Certification Overview

Whether you're new to -omics or have years of bench-to-bioinformatics experience, this certification will sharpen how you evaluate data quality, choose the right analysis for your scientific question, and communicate results with confidence. You'll walk away able to:

• Understand major -omics types and when to use each one
• Judge whether a dataset is fit for your research question (replicates, batch effects, metadata)
• Interpret pipeline and downstream QC metrics, spotting red flags before they affect your conclusions
• Apply statistical rigor to differential expression, GSEA, and multi-omics integration
• Build clear, audience-appropriate scientific narratives using Pluto's reporting tools
• Complete an end-to-end capstone project on a dataset of your choice, including a multi-omics integration

The certification is organized into four courses:

Course 1

Foundations of -Omics Data & the Pluto Platform

Develop a foundation of -omics literacy and an understanding of the Pluto platform. This course focuses on understanding datasets, where to find them, and how to import them into Pluto.

This course comprises three modules, each broken down into digestible sections. Most topics take 2 to 4 minutes and are followed by a learning check. Your progress saves automatically, so you can close the tab anytime and pick up right where you left off.

Modules

• Module 1.1Omics Types & When to Use Them12 topics
• Module 1.2Determining the Qualities of a Dataset4 topics
• Module 1.3Finding & Acquiring Datasets4 topics

Welcome & Objective

Each of these modules is broken into short ‘snippets’ of information, with a learning check after each one. Your progress saves automatically; feel free to close the tab anytime and pick up right where you left off.

What Is Pluto?

Pluto is a platform that provides an infrastructure layer for teams working in translational research and drug discovery programs. As a no-code, AI-native platform, Pluto enables researchers and bioinformaticians from any background to seamlessly import datasets, perform analyses, and interpret and share results across teams. By removing the ‘custom code’ requirement traditionally used for analyzing -omics datasets, scientists without a coding background can run and interpret their own experiments directly through point-and-click interfaces, and bioinformaticians can move through routine work far faster.

Beyond speed, a shared platform solves problems that come up constantly in research programs:

• Reproducibility. Every analysis records its exact parameters and inputs, so a colleague, or you months later, can see precisely how a result was generated and reproduce it.

• Shared context. Analyses, plots, and interpretations live in one place the whole team can access, rather than scattered across individual laptops and one-off scripts.

• Public data, ready to use. Thousands of published datasets are already imported and searchable, so a new hypothesis can be tested against existing public data before committing to a new experiment.

The rest of this module walks through the -omics data types you'll encounter, then returns to where Pluto fits into the broader research workflow.

Overview of -Omics Types

"Omics" refers to large-scale technologies that each measure a different molecular layer at scale. Some measure DNA, while others measure RNA, proteins, chemical modifications, or metabolites. The next several topics walk through each layer individually: what it measures, what it's commonly used for, and where its limits are.

Genomics

Genomics refers to sequencing of DNA to identify mutations, copy number changes, and structural rearrangements. A person's germline genome is essentially fixed, and DNA is chemically stable in a way that RNA and protein levels are not; it doesn't fluctuate hour to hour in response to treatment or stress. However, somatic changes do accumulate, which is exactly what cancer sequencing studies try to identify.

Genomics answers the question of what changes exist in the DNA. Since it doesn't tell you whether those changes affect protein activity, there are limitations to keep in mind: not every mutation identified through genomics translates into a functional consequence.

Common use cases: identifying inherited disease-associated variants in a family history study, characterizing the mutational landscape of a tumor biopsy, screening cell lines for off-target CRISPR edits, or genotyping a mouse colony to confirm a knockout.

Transcriptomics

Transcriptomics is most commonly measured through RNA sequencing (RNA-seq) and measures the level of each gene’s RNA at the time of sequencing. Unlike DNA, which stays largely the same across a cell's lifetime, RNA levels are constantly changing in response to drugs, stress, developmental stage, or environmental signals. This makes transcriptomics a snapshot of what a cell is doing right now rather than what it's capable of doing. Its genome-wide coverage and relatively low cost have made it one of the most commonly used -omics technologies across basic and translational research.

Common use cases: comparing gene expression between treated and untreated samples, characterizing how a cell line responds to a drug over time, identifying a gene signature that separates patient subgroups, or profiling tissue-level changes across disease progression.

Proteomics

Knowing what a cell is transcribing is only part of the picture. Drugs typically act on proteins rather than on genes or RNA directly. As a result, understanding the protein landscape - what's present and how much - is a separate and necessary layer.

Most proteomics workflows digest proteins into peptides and use mass spectrometry to identify and quantify them by their mass-to-charge ratio. Affinity-based platforms are also widely used, particularly for plasma and other clinical samples. Proteomics can also reveal modifications made to a protein after it's built, such as phosphorylation, which is frequently the specific molecular switch a drug is designed to hit. Because it measures the functional molecule itself, proteomics is often the layer that confirms whether a genomic or transcriptomic finding is biologically real.

Common use cases: comparing treated and untreated samples, identifying a signature that separates patient subgroups, biomarker discovery.

Epigenomics

Epigenomics is the global study of chemical modifications to DNA like methylation, macro DNA structure, and the packaging proteins that bind to DNA. Collectively, these modifications regulate gene expression by adjusting gene accessibility and transcriptional machinery. These marks are reversible and they underlie several approved drugs, for example, azacitidine and revumenib to treat acute myeloid leukemia.

ChIP-seq and CUT&RUN are two common methods for generating epigenomic data. ChIP-seq is long established: proteins are crosslinked to DNA, the DNA is sheared, then an antibody is used to pull down DNA fragments of interest. CUT&RUN is a newer, lower-background alternative. An antibody against the protein of interest is added to permeabilized cells, and a protein A/G–micrococcal nuclease fusion is recruited to that antibody, cutting DNA only where the target is bound. Because there is no crosslinking or sonication and only bound fragments are released, it needs far less starting material and produces cleaner data.

Common use cases: mapping binding locations of a transcription factor or chromatin accessibility before and after a drug treatment or disease state.

Metabolomics

Metabolomics measures small molecules such as amino acids, sugars, and lipids, which are the end products of what is happening upstream in a cell. Because cancer cells and many other disease states are known for rewiring their metabolism, the metabolomics layer can reveal drug targets and biomarkers that are difficult to decipher with genomics, transcriptomics, or proteomics alone.

Common use cases: identifying a metabolic biomarker that distinguishes disease from healthy tissue, characterizing how a drug shifts cellular energy metabolism, or profiling a patient sample for signs of a specific metabolic disorder.

Resolution, Single-Cell & Spatial Approaches

So far, each layer covered can be measured in bulk, meaning a sample's cells are pooled together and measured as an average. Single-cell and spatial approaches add a different dimension: resolution.

Single-cell sequencing separates a tissue into individual cells before measuring them, most commonly for RNA, though other -omics layers exist too. This can reveal a specific cell state, such as one that's resistant to a therapy, which would have otherwise been invisible in a bulk average. However, reaching cellular resolution in this case comes with an important tradeoff. To separate cells, the tissue must be fully dissociated, resulting in the loss of any information about where each cell originally sat in the tissue.

Spatial approaches maintain location information along with gene expression or other molecular measurements at or near single-cell resolution. Capturing where a cell sits in the tissue is critical when location carries biological meaning. For example, a T cell infiltrating the center of a tumor may have a different functional state than a T cell confined to the stroma yet a single-cell measurement might group these two T cells together. One caveat worth noting is that spatial sequencing coverage is typically not as deep as standard single-cell sequencing.

Common use cases: tracking a cell state transition across treatment to find early markers of resistance, mapping which immune cell types cluster at a tumor's invasive edge, or resolving cell-type composition in a heterogeneous tissue such as a developing organ.

Multi-Omics: A More Holistic Picture

Now that we've covered each -omics layer individually, let’s dive into what happens when you combine two or more of them on the same sample.

Multi-omics is used broadly across disease domains, like cancer, immunology, neuroscience, microbiome research, and agricultural biology, and provides a more complete picture of a biological system. The core benefit is that, by combining multiple layers, relationships become clearer and a more holistic story unfolds.

Let’s consider a few ways that multi-omics can add valuable insights:

• Confirming that a genomic finding matters biologically. A mutation only matters if it changes protein behavior, and only proteomics or downstream layers can confirm that.

• Revealing patient subgroups that a single layer would miss. By combining genomic, transcriptomic, and proteomic signals into a shared subtype, patient populations can further be stratified.

• Understanding resistance mechanisms as they develop. Tracking how a cell's regulatory and transcriptional state evolves under treatment pressure over time can reveal critical periods to consider in disease progression.

Institutional efforts like the Clinical Proteomic Tumor Analysis Consortium (CPTAC) and The Cancer Genome Atlas (TCGA) have generated multi-omics profiles across thousands of tumors and dozens of cancer types and remain widely used public references for this kind of integrated analysis. For example, a CPTAC study of lung adenocarcinoma found that many DNA copy number changes showed little to no corresponding change at the protein level.

Pluto’s Value in The Research Journey

Every -omics-driven research question has similar pillars anchoring their research journey. This section walks through that journey in general terms and shows the value of Pluto in each of those pillars.

Research & Prioritization

Research questions in drug discovery begin with a target disease, physical symptom, or abnormal cellular process which ultimately lead to a set of candidate genes, proteins, or biomarkers worth investigating. Scoring and prioritizing those candidates draws on experimental data including -omics data, published literature, and other public or proprietary knowledge sources. In target discovery specifically, this looks like filtering candidates by genetic validation, expression specificity, and targetability. Pluto supports this with built-in gene and target lists that can be updated as new evidence comes in.

Multi-Omics Data Analysis & Visualization

Visualizations are critical for streamlining investigation of gene expression, pathways, and mechanisms in multi-omics analyses. In target discovery, this process might look like confirming a candidate is active in the right cell type with single-cell RNA-seq or confirming a genomic hit changes protein behavior with proteomics. With Pluto, novel and published datasets are quickly ready for analysis. Broad measurements by these -omics types also enable more open-ended analyses and questions that aim to narrow hypotheses and directions.

Literature Curation

Findings rarely stand alone. Searching articles, clinical trials, and other sources, then organizing and categorizing what's found, builds the evidence base behind a research question. In target discovery, this looks like attaching a supporting paper directly to a candidate as evidence for its validation status. Pluto supports this by letting teams search literature within the platform and link papers directly to the genes or biomarkers they support.

Communication & Reporting

A finding can typically only impact a program decision once it is shared with other members of the team. Turning experimental data into presentation-ready figures and reports keeps stakeholders and collaborators aligned. Pluto supports this with custom figure export, use of a canvas to tell a story, and even sharing analysis spaces with collaborators so they can explore the data in a hands-on manner that isn’t possible with static reports. We’ll dive into this more in a later course.

Inside the Pluto Platform

Within the platform, you are empowered to analyze and interpret your data efficiently and confidently. Additionally, our platform natively enables you to view and create gene lists, curate literature as supporting evidence, and report findings at the click of a button by turning experimental data into publication-ready figures for stakeholder and program status meetings.

Each user is signed up under their organization's workspace. Each workspace contains Program-level assets, including datasets and analysis spaces. Programs are the highest level of organization within a Pluto workspace, and these are specifically designed to keep datasets and analyses connected within a single platform. Datasets refer to the assay data and sample metadata, which are shared across a program. Analysis spaces house the ‘story’ of the science in a flexible and highly editable canvas space. Here, you can bring in multiple datasets and perform analyses in a single space, making it easier to generate hypotheses, evaluate findings, and find patterns across the data. Importantly, datasets cannot be duplicated within a program, but unlimited analysis spaces can be created within individual programs.

High level summaries about programs are displayed in the Workspace dashboard. Here, a user can see a breakdown of program assets and customizable widgets, including but not limited to the stage of a specific program and/or the number of active programs, datasets, and analysis spaces.

See our ‘Getting Started with Pluto Programs’ guide for more details on setting up your first program, dataset, and analysis space.

At the program level, a user can have one of four roles: owner, editor, viewer, and member. For more details on permission levels, navigate to Users & Teams and click on the ‘Permissions’ icon.

• Owner: There can only be one owner, and ownership can be transferred by the owner to another user.

• Editor: users with Editor permissions can edit and invite users to an Organization, Team, and Project. Editors can edit Experiments, Programs, Biomarker lists, and Entity databases.

• Viewer: users with Viewer permissions can View each of the entities listed above, with no additional permissions for editing or inviting.

• Member: the most limited users who have no access to any data unless they are explicitly added to an entity, such as an Experiment. This role is most commonly used for outside collaborators or CROs that need access provisioned.

Welcome & Objective

What makes a good dataset? This is a tough question to answer, but a critical one, especially in the field of big data.

The reality of working with public -omics data is that there are tens of thousands of studies available, and for most of them, running any type of analysis will give you an answer. However, the differentially expressed genes, the significant p-values, and the clean PCA plot can all be results of how the samples were handled rather than true underlying biology.

Together with the resources built into the Pluto platform, this part of the course will teach you a practical framework for evaluating any -omics dataset before you touch the data. You'll learn about the importance of replicates and the impact of batch effects, and be able to answer the question 'is this dataset good enough for my research question' with enough confidence to act on the result.

Why Replicates Matter and What to Look For

Datasets can have biological and/or technical replicates. A popular example for understanding the difference is taking height measurements. Say you took the height of a single individual five times on the same ruler, that's five technical replicates, and the information gathered reveals how precise your measurement tool is. Taking the height of five different individuals, on the other hand, represents five biological replicates. What you gather in the latter scenario reveals information about the population you sampled from. Importantly, for any scientific claim about a disease, only biological replicates count.

In -omics data, a biological replicate is an independently derived sample: a different patient, a different mouse, a different independently cultured flask of cells. A technical replicate, in this case, would be using the same RNA sample and running it through sequencing twice.

The minimum number of biological replicates a scientist should aim for depends on the experiment type. As a general guideline, bulk RNA-seq should have at least three replicates per group as a practical floor. Standard differential expression tools (e.g., DESeq2, edgeR) will technically run on two replicates per condition, but empirical dispersion estimation in the negative binomial models underlying these tools becomes unstable at n=2: variance estimates for individual genes are unreliable, and shrinkage toward a genome-wide trend has to do most of the work. This makes results unstable: power drops sharply, and individual gene-level calls become highly sensitive to a single outlying sample. Three or more replicates give the dispersion estimates enough stability to trust, with more recommended for translational or clinical claims.

In single-cell RNA sequencing, the number of donors sampled is more relevant to a population-level claim than the total number of cells profiled, since individual cells from the same donor are correlated with one another rather than fully independent. Treating cells as the unit of replication measurably inflates false-positive rates in differential expression testing.

What Are Batch Effects and Why Do They Matter?

Batch effects refer to any systematic technical difference between groups of samples that misleadingly appears biological. Consider the classic scenario below.

Comparing tumor vs. normal tissue by bulk RNA-seq: tumor samples were snap-frozen, collected in 2021, and sequenced on a NovaSeq. Normal samples were collected a year later, stored in RNAlater, and sequenced on a HiSeq. In a differential expression analysis, what would be the biggest signal in the data? Would it be the biological differences between tumor and normal samples or would the larger differences be based on the batch and experimental differences? There would be no way to disentangle the effects of biology and batch.

At every stage of an experiment, batch effects can be introduced: how quickly tissue was frozen, which library prep kit lot was used, how deeply each sample was sequenced. Fortunately, many of these are detectable, but only if the metadata exist.

In the Pluto app, you can access metadata for any dataset by navigating to the dataset, which will be covered in Module 1.3. In the sample metadata, you can confirm that the dataset has your target of interest, has enough samples or biological replicates, and contains the necessary metadata for your scientific question.

3 Questions to Consider Before Selecting a Dataset

Does it measure what you need?

Does the platform cover your target molecule class (DNA, RNA, protein, metabolite)? Is your specific gene, protein, or locus of interest measured in the assay? Is the measurement from the right biological compartment (tumor vs. stroma vs. blood)? Are the correct controls included in the dataset?

Does it have enough samples?

Are there enough biological replicates per group for the planned statistical test? Were samples powered for your question, or only the original study's question? For rare features, is the dataset deep enough to detect them reliably?

Does it have the right metadata?

Clinical: treatment history, response, outcome, staging. Technical: batch, platform, processing dates. Biological: tumor purity, cell composition, tissue site.

Consider the example below to make this concrete with a research question.

Welcome & Objective

This module follows the natural order of working with data in Pluto: first finding a relevant dataset, then seeing what one looks like once you're inside it and finally learning your options for bringing in your own data, whether that's public data you want to reanalyze or data you generated yourself.

Where to Find Data

Let's start at the beginning: locating a dataset that's relevant to your research question.

Both internal and public datasets are accessible by navigating to the Datasets tab on the left navigation pane. Here, you will find our Global Data Commons, which includes over 14,000 public datasets.

Here, you can search for published datasets and analyze them interactively for your own proprietary research. Note that there are options to filter by several fields, including Dataset type (i.e., RNA-seq, ATAC-seq, etc.), Organism, Program, etc. Feel free to explore different filters! To streamline your workflow, you are also able to save current filters, which will then appear on the lefthand side of the Datasets page.

You may notice several datasets are titled with a 'GSE' identifier. These datasets have been imported from the Gene Expression Omnibus (GEO) public repository. The GEO repository accepts array- and sequence-based functional genomics data. Users can query, download, and upload datasets. Given the popularity of GEO datasets, Pluto has a built-in feature to streamline the import process. We will explore the request feature later in this module.

When you are ready to analyze a dataset, the next step is to create an analysis space. These are scientific canvases within a Program where multiple datasets can be linked and analyzed. These shared canvases can help teams make Program decisions faster by keeping ideation, hypothesis testing, and insights all in one space. Test this out yourself: Navigate to the Pluto Certification Program and create a new analysis space.

Name your analysis space with your initials or another unique identifier so you can come back to this in a later course. Feel free to test out adding a specific goal or hypothesis. In the next section, we will be adding a public GEO dataset (GSE264728) to your analysis space. This GEO dataset uses RNA-seq data to explore gene expression changes in breast cancer.

Hands-On, Adding a Dataset to your Analysis Space

Now that you’ve created your own analysis space, let’s add a public dataset. Your analysis space is an open canvas to capture a scientific story across an individual experiment as well as across multiple different types of experiments. To add existing datasets to this analysis space, click on ‘+ Link a dataset’. We will navigate to the Global Data Commons tab.

In the search bar, you can type in a specific identifier or a keyword of interest. In this example, we'll search for GSE264728. There will be two results: one includes single-end (SE) RNA-seq reads and the other includes paired-end (PE) RNA-seq reads. We will work with the PE dataset, "[GSE264728] PE data | Combined inhibition of KAT6A/B and Menin reverses estrogen receptor-driven gene expression programs in breast cancer."

This dataset is now linked to the Program, which means its contents and metadata can be viewed from the Datasets tab within the Program. This is where you can view the experimental data registry. The Overview tab provides a summary of the data, like in the screenshot below. Feel free to explore the other tabs further.

Navigating to the Protocols tab will open an experimental notebook where entries can be made to capture experimental objectives, protocol details, and other relevant notes, with full markdown capability.

Creating New Datasets in a Program

In many cases, you may want to analyze data generated by internal teams at your company, or you may want to analyze a public dataset that has not yet been imported into the platform. Below, we will briefly discuss the workflow for those two processes.

Uploading Your Own Raw or Processed Data

From your Program, navigate to the Datasets tab so you can access the experiment data registry. Here, you can import your own data by selecting ‘+ Dataset’. When adding new datasets, formatting your data uploads is arguably one of the most important steps to get right. To assist you, help documentation is available, and we also have in-app modals with live examples to reference while creating a new experiment.

Each Pluto dataset requires two uploads: assay data (measurements) and sample data (metadata describing each sample). Most experiment types accept either FASTQ files (raw sequencing reads) or a processed data table (CSV). Importantly, if uploading FASTQ files, the sample data must include the r1_fastq column, and, if the data is paired-end, ther2_fastq column is also required. For processed data, accepted units vary by experiment type per our Data Upload Overview knowledge base article.

GEO Imports

As you explored earlier in this module, Pluto has thousands of GEO datasets already available in the app. Our platform also supports GEO import requests for other datasets, so users can perform investigative or complementary analyses on relevant published data. As demonstrated above, many GEO datasets are available in the Datasets page, where you will navigate to the Global Data Commons tab. To request the import of a public GEO dataset, simply click on the ‘+ GEO import’ button located directly next to the import dataset button (‘+ Dataset’).

Clicking on this button will bring you to a pop-up window where you can enter the GEO accession ID. This GEO accession ID will be run against our catalog of over 14,000 public datasets, and if the dataset has not yet been imported, then your request will be forwarded to our Scientific Support Team. Importantly, there are limited requests shared across the workspace. To view these details, click on ‘View GEO details’ in this pop-up window.

The Published experiment requests shows your workspace’s submitted, in progress, and completed requests. For any GEO import requests that are declined (such as in cases where the sequencing technology used is not supported by our Platform), these will also be noted in the ‘invalid’ tile.

External and Remote Data Transfer

Pluto also supports external and remote data transfer from Globus, TCGA, SFTP/FTP, Google Cloud Storage (GCS), Amazon S3 (AWS), and other sources such as GENEWIZ or a Backblaze B2 Bucket. Find more information about what's required for granting Pluto access by submitting the external data source request form. This form, along with other useful resources, is available via the in-app live chat button at the bottom right of every page.

Course 2

Interpreting Quality Control (QC) and Analysis

In this course, you will develop the critical thinking skills to assess whether a dataset will yield quality insights.

This course is broken into two modules, each split into short topics with a learning check after every section. Most topics take 2 to 4 minutes. Your progress saves automatically, so you can close the tab anytime and pick up right where you left off.

Modules

• Before You BeginBefore You Begin: Library preparation and next-generation sequencing5 topics · optional
• Module 2.1Pipeline Quality Control (QC)11 topics
• Module 2.2Downstream QC5 topics

Welcome & ObjectiveOptional

Course 2 will cover quality control (QC) of sequencing experiments. It is useful to have a basic understanding of next-generation sequencing and library preparation to get the most out of the QC topics. If you already have a good understanding of these topics, please feel free to skip this module.

Throughout the primer, we'll focus on bulk RNA-seq, the assay behind the PLX259235 experiment you'll explore in Module 2.1. We'll use CUT&RUN as a point of contrast since Module 2.1 refers to it when discussing duplication and read depth.

Skip the primer and go to Module 2.1

Library Prep: Turning Biology into Something a Sequencer Can ReadOptional

Adapters are short synthetic DNA sequences ligated to both ends of every fragment; a short-read sequencer can only read molecules that carry them. Since sequencers can't read RNA, chromatin, or a protein's binding site directly, a process called Library preparation is required. This process converts your biology into that one readable format. The resulting "library" is simply a tube containing billions of these adapter-tagged fragments.

Every assay is a different way of deciding which fragments end up in that tube, but nearly all of them involve variations of these steps.

1. Get to DNA. RNA is converted into complementary DNA (cDNA) by reverse transcription. Prior to this step, it is necessary to determine which RNA molecules to include in the preparation. It is typical to exclude RNA that is not polyadenylated (polyA) or to deplete ribosomal RNA.

2. Define the fragment ends. Molecules are cut into short pieces by enzymes, sonication, or the assay's own chemistry. In CUT&RUN, where those cuts happen is the measurement itself. For some bulk RNA-seq mRNA kits, this step precedes reverse transcription.

3. Attach adapters. Adapters are added to both ends so each fragment can bind to the sequencer's flow cell. If a fragment is shorter than the read, the sequencer reads past the fragment and into the adapter which is why adapter content may appear in your QC report.

4. Add tags. A short sample index is added so many samples can be sequenced together and sorted apart afterward.

5. Amplify and size-select. PCR makes enough copies to sequence, and a cleanup step keeps fragments in the size range the instrument needs. Too little starting material at this step leads to duplicates and low complexity, which is the focus of the next topic. Note that some library prep kits are designed to omit PCR amplification.

Library Complexity: Why "Just Sequence Deeper" Often Won't HelpOptional

Library complexity refers to the number of genuinely distinct original molecules in the tube. PCR can make copies but not generate new observations. If a library contained too little material to begin with, deeper sequencing would only return the same molecules, again and you pay full price for the redundant reads. This is why sequencing providers ask so insistently about input amount and quality and why low complexity shows up in a QC report as a high duplication rate - a metric you'll explore in Module 2.1.

How a Sequencer Reads DNAOptional

Each library fragment is copied in place on a glass flow cell into a cluster on the order of a thousand copies. The clusters are then extended one base at a time, and the whole surface is imaged after every cycle. The color seen at each cycle determines the base call, and how unambiguous that color was determines the quality score.

Source: https://support.illumina.com/content/dam/illumina-marketing/documents/products/other/ngs-primer-770-2012-008.pdf

A few practical consequences follow from this:

• Read length is a purchasing decision rather than a property of your sample. This simply refers to how many cycles were run. For example, 2×100 means 100 cycles on each end of the fragment.

• Paired-end vs. single-end. Paired-end sequencing reads both ends of each fragment, which can be used to calculate fragment length when both pairs are aligned to the genome. For CUT&RUN, fragment length is part of the measurement so paired-end is required.

• Two-color chemistry and poly-G. Newer instruments like the NextSeq and NovaSeq record G as the "dark" base with no fluorescent label, so any cycle with no signal gets called G. Very short fragments and failed clusters can therefore produce long runs of G, which can be addressed by using trimming tools.

• Index hopping. A small fraction of reads can be assigned to the wrong sample. Unique dual indices can help to minimize misalignment. Importantly, index hopping matters most when a low-input library shares a run with a much deeper one.

From Instrument to FASTQOptional

Because all samples were pooled together on one flow cell, the instrument writes per-cycle base calls and quality scores for the whole run into BCL files. Demultiplexing then uses each read's sample index to sort reads into per-sample FASTQ files.

What's inside a FASTQ file

A FASTQ file is plain text, with four lines per read, repeated millions of times:

Quality scores use the Phred scale: Q20 means a 1 in 100 chance that a base call is wrong, and Q30 means 1 in 1,000. The "%≥Q30" figure on a sequencing run report is simply the fraction of all base calls at or above Q30. It is normal for quality to decline toward the end of reads.

Welcome & Objective

Before performing or trusting a downstream analysis, you need to understand the quality of the raw data behind it and whether the pipeline run was actually any good. Pipeline QC is an important step in evaluating a dataset before you see any of the biological results. The good news is that you don't need to be the person who ran the pipeline to critically read and understand these results.

These metrics are only available in Pluto if 1) raw FASTQ files were uploaded, and 2) a Pluto pipeline was used to process the data.

What Is a MultiQC Report, and Where Do You Find It?

Raw sequencing data passes through many individual tools before it becomes a processed dataset: read quality assessment, adapter/quality trimming, alignment to a reference genome/transcriptome, or pseudo-alignment to a transcriptome, duplicate marking, and quantification. Each of these tools produces its own log file or report in its own format. MultiQC is the industry-standard tool that recognizes the data formats from dozens of common bioinformatics tools and compiles everything into a single, interactive HTML report, making it one place to scan every sample's quality metrics side by side instead of digging through dozens of separate log files. These reports are organized into modules, one section per tool it detected (e.g., FastQC, STAR, Salmon, samtools, Picard), each with per-sample tables and comparison plots.

In Pluto, a QC report is automatically generated for any experiment where raw FASTQ files are uploaded and a pipeline was run. You can find the QC report for a dataset in your Program by navigating to the Pipeline tab within that dataset. If you instead uploaded an already-processed data table, there is no pipeline run and therefore no QC report to review, because Pluto never touched the raw reads. However, you can upload your own QC files as attachments to the dataset.

Key QC Metrics: Alignment

Not every metric from MultiQC is equally important. When analyzing your MultiQC report, the key questions are: how well reads aligned, how much of the data is redundant, and how deeply and evenly the genome (or target regions) were sequenced. A single sample failing one metric doesn't automatically mean that you should exclude it from downstream analysis. Instead, it means: 1) look at that sample's other metrics, 2) check whether it's confounded with a biological group of interest (excluding it could bias your comparison), 3) consider its impact (if any) on the biological interpretation of analyses, and, importantly, 4) document the decision either way.

Alignment metrics tell you what fraction of sequenced reads could actually be placed onto the reference genome or transcriptome and how confidently. The headline number is the mapping rate: the percent of reads that aligned uniquely to a single location.

For most bulk RNA-seq experiments in human or mouse against a standard reference, expect somewhere in the 80–95% uniquely mapped range. A mapping rate far below these levels could indicate:

• Wrong reference genome or transcriptome selected for the species

• Degraded RNA or poor library prep producing short, uninformative fragments

• Residual adapter sequence that wasn't trimmed before alignment

Alignment metrics can be found in the General Stats table.

Key QC Metrics: Duplication

The duplication rate is the percent of reads that are exact or near-exact copies of another read in the library, which is most commonly introduced during PCR amplification. A high duplication rate can signal low library complexity. If you started with limited unique molecules (often due to low input material) and amplified them heavily to reach your target read depth, you would be re-sequencing the same fragments over and over rather than sampling new ones.

For most DNA-based assays (ChIP-seq, CUT&RUN, exome/genome sequencing), high duplication directly reduces your effective, usable depth and is typically removed (deduplicated) before downstream analysis - although there are considerations for technical control samples and scenarios where similarly cut fragments (like from CUT&RUN) are expected. For bulk RNA-seq, duplication typically isn't a large concern because they can happen naturally for highly expressed genes and PCR duplicates typically correlate with original expression level, meaning that signal of relative expression is still present in the dataset. However, very high duplication rates, such as greater than 80%, could warrant further inspection.

A key step is to identify samples whose duplication rates differ greatly from other samples in the dataset. If they do, it could indicate a technical problem with that sample's library prep or sequencing.

Duplication rates can be found in the General Stats table.

Key QC Metrics: Coverage and read depth

Coverage (or depth) describes how many times, on average, a given base or region was sequenced. It matters in two dimensions: how deep (mean coverage) and how even (uniformity across the genome or target regions).

Required depth depends heavily on the assay and the goals of the investigator: Assays like CUT&RUN can have high specificity and low background, meaning they can be adequately powered with fewer reads. Bulk RNA-seq, on the other hand, can typically be adequately powered with 20-30 million reads per sample but a higher number of reads may be required if the goal is to detect lowly expressed genes or isoforms. Replicate number typically matters more than read depth for RNA-seq once reads surpass roughly 10 million.

Read depth is another area where you should look for outliers. Most analyses take into account read depth, so some variation is expected and accounted for, but large differences should be noted. For example, an RNA-seq sample with 10 times as many reads as another sample might pick up genes that were not identified in other samples but it could be more of a result of a technical artifact rather than a true, biological difference. Conversely, a 2X spread in depth is common and can be largely handled by size-factor normalization for moderately and highly expressed genes. However, it still costs you sensitivity for low-count genes in the shallower samples, so check that depth isn't confounded with your comparison groups.

Read depth per sample can be found in the General Stats table. Note that some columns will report the number of aligned reads as the number of molecules (e.g., RSEM) while others can count Read 1 and Read 2 separately (e.g., Samtools).

Key QC Metrics: Sample status checks

This section reports each sample's strandedness check, which is a fail-safe to record what type of library strandedness was selected by the user versus what was inferred via bioinformatics tools (RSeQC or Salmon). The strandedness is a result of the library kit utilized in the library preparation step. Some kits can preserve the original strandedness of the sequencing molecule (forward or reverse technologies) and some cannot (unstranded). In Pluto, the default is to run RNA-seq pipelines with automated detection of the library strandedness.

A "Fail" here is a serious red flag because it means either that two prediction algorithms disagree on the library strandedness or that the algorithms disagree with the chosen pipeline strandedness. In either case, this should be explored further because an improperly chosen strandedness will distort every downstream gene count.

Key QC Metrics: Adapter Content

The Adapter Content section is found within the FastQC section for both raw reads or after they have been trimmed for quality or adapter sequences.

The plot shows a cumulative percentage, by position along the read, of how much of the library still contains a known adapter sequence. Sequencing platforms require appending adapter sequences to the ends of the biological molecules to be sequenced. In some cases, these sequences can be read and returned in the dataset. An ideal sample will have near zero adapter content throughout but it is common that a curve rises, especially toward the end of the read, meaning that some shorter biological molecules were sequenced and now the adapter is being picked up. If a read is still of sufficient size, the sequence is still used in the analysis.

Adapter trimming is a common bioinformatics step (it should mostly be removed in the FastQC (trimmed) section) but very high adapter content should be flagged.

Key QC Metrics: Per Base Sequence Content

This section is also found within the FastQC section for both raw reads or after they have been trimmed for quality or adapter sequences. It shows the proportion of A/C/G/T bases called at each position across all reads. For most sequencing, you'd expect the four bases to run roughly flat and parallel across the read because there is mostly an equal chance of each base pair being at each numeric position (depending on the genome).

The early portion of RNA-seq reads is a common exception. Most RNA-seq library prep uses random hexamer priming to convert RNA into cDNA. Those hexamers don't bind with equal affinity everywhere; they have a slight sequence preference, so the first dozen or so bases of every read show a reproducible, non-random base composition skew. This is a real technical bias, but it is universal to nearly all RNA-seq libraries, cannot be fixed by trimming, and in practice does not meaningfully affect downstream quantification or differential expression. Expect to see it; there is no need to treat it as a red flag on its own.

Key QC Metrics: Genomic origin of reads

This section is found within the QualiMap section. The bar chart classifies every mapped read as exonic, intronic, or intergenic. The expected distribution of where the reads fall depends on the experiment type.

For polyA-selected or well-depleted rRNA bulk RNA-seq libraries, expect the large majority of reads to land in exons. A high intronic fraction can point to substantial pre-mRNA/unspliced transcript capture, nuclear RNA, or degraded input material. A high intergenic fraction can point to genomic DNA contamination.

As with many of the other QC metrics, check for any samples that do not look like others in the dataset.

Key QC Metrics: Mapped Reads

For RNA-seq datasets aligned with the STAR + RSEM method, the mapped reads will be in the RSEM section. It contains RSEM's own accounting of what happened to every read in each sample: how many reads aligned, didn't align, or were multi-mapped (aligned to multiple regions)?

This is the same alignment concept covered earlier in this module, now broken down per sample as a stacked bar: a quick visual check for any sample where the usable fraction is noticeably smaller than its replicates. You can view this plot by total read numbers or by percentages within each sample. Samples with a high percentage of unaligned or multi-mapped reads could indicate an issue with the library.

Key QC Metrics: nf-core/rnaseq Workflow Summary

A plain record of exactly how the pipeline was configured for this run: pipeline version, reference genome and annotation build, aligner and quantification method, as well as key thresholds such as the minimum mapped reads required per sample. This section is useful to see what parameters were chosen for a particular dataset and to compare them to other datasets.

Welcome & Objective

Pipeline QC tells you whether your raw sequencing data was generated well. Downstream QC asks a different question: once you have processed data, does it actually behave the way you'd expect for your experiment, or are there hidden technical effects that will confound your results before you've run a single statistical test?

Common Types of Data Normalization & Transformation

Raw counts are not directly comparable across samples, and their statistical properties make many downstream methods behave poorly if used as-is. Two of the most common corrections you'll see applied before visualization or exploratory analysis are counts per million (CPM) normalization and log-transformation.

📏 CPM (Counts Per Million)

Each gene's (or peak's) raw count is divided by the total number of gene or peak counts for that sample, then multiplied by one million. This corrects for differences in sequencing depth (library size) and alignment rates between samples. Without it, a sample sequenced to twice the depth of another will show roughly double the raw counts for every gene, with no biological meaning behind the difference.

CPM makes gene or peak abundance comparable across samples. It does not correct for gene length, which matters more when comparing different genes to each other within the same sample - that's where TPM/FPKM can be used.

📉 Log-transformation

Typically applied as log2(CPM + 1) or similar. Raw and CPM-normalized counts have a strong mean-variance relationship: highly expressed genes have much larger absolute variance than lowly expressed genes, which violates the assumptions of many standard statistical and visualization methods.

Log-transformation compresses this range, stabilizes variance across the expression spectrum, and makes distributions more symmetric. It's a standard step before PCA, hierarchical clustering, and heatmap visualization, methods that otherwise get dominated by whichever handful of genes have the largest raw variance. Aesthetically, it can also mean that it is easier to visualize genes on the same plot even if they differ in abundance by orders of magnitude.

Principal Component Analysis (PCA): What Is It?

A typical RNA-seq dataset measures tens of thousands of genes per sample, far too many dimensions to look at directly. Principal Component Analysis (PCA) is a dimensionality reduction technique that finds a small number of new axes (principal components, or PCs), each a weighted combination of many genes, that capture as much of the variance across samples as possible. PCA is a great initial analysis to run once you have your processed data. PC1 captures the single largest axis of variation in the dataset, PC2 the largest remaining axis orthogonal to PC1, and so on.

Plotting samples along PC1 and PC2 compresses a massive gene expression matrix into a single 2D scatter plot that's fast to visually interpret and is usually the very first plot worth looking at for any new dataset, before any formal statistical testing. The percent of total variance each PC explains (shown on the axis labels) helps you judge how much of the dataset's overall variability is captured in that 2D view.

You don't need linear algebra to use PCA well; what matters practically is that each PC is a data-driven summary of "what varies most", and it has no idea whether that variation is biological or technical. That distinction is entirely up to you, which is important and the focus of this section is to provide examples of different scenarios where PCA can indicate important aspects of a dataset.

Using PCA to Evaluate Your Dataset

The value of a PCA plot comes from recognizing how patterns emerge across your samples. To help explain some of these patterns you can hover over the plot points to see the sample names and you can color the samples by any biological or technical variable present in your sample metadata (sequencing batch, collection date, processing site, RNA integrity, operator). The pattern that emerges tells you a lot about whether your dataset is fit for purpose.

Ideally, your biological groups (e.g., treatment vs. control) are cleanly separated along PC1 and PC2. See the PCA of experiment PLX121088 as a good example of clear biological separation. In this case, treatment clearly separates the samples into two groups.

There are many situations where this might not be the case. It does not necessarily mean that the dataset is unusable but the pattern should be explored to determine what the most appropriate action is.

PCA - sample outliers

See the PCA of experiment PLX095216. In this example, 6 samples are clear outliers from the rest of the dataset. Looking over the sample data table, it isn't clear why this would be the case and it could make sense to remove these samples from the analysis before continuing.

Using PCA: Group Variation, Batch Effects & Covariates

PCA - high group variation

See the PCA of experiment PLX213665. The plot is labeled by which samples came from patients with nonalcoholic steatohepatitis (NASH). Disease and control samples are intermixed in the plot, without clear separation. This is a scenario where PCA suggests the group effect is small relative to other variation. You may still recover DE genes, but expect a modest list. Indeed, this is observed in the volcano plot for this dataset with only a handful significant genes being identified.

PCA - batch effects

See the PCA of experiment PLX045098. This plot shows a clear pattern in the data that forms a "U-shape". It is not explained by the available sample metadata but it could be a batch effect or another type of unexplained continuous structure. If we had more information about the samples and some of the metadata correlated with this pattern, we could incorporate it into our differential expression model. Not all batches would produce a “shape” but they typically do display as a pattern in the PCA plot that disrupts the expected separation of experimental groups. Some examples of batches could be sequencing run, collection date, or operator.

PCA - biological covariate

Take a look at the two PCA plots in the experiment PLX259235. In one plot, you can see that the major sample groups are separated by cell line. In the other plot, you can see that drug resistance also clusters the samples but only within each cell line. Cell line would be an example of a biological covariate. You may want to include it in your differential expression model if you want a unified comparison that only evaluates the effect of drug resistance on several cell lines.

Course 3

Scientific Workflow and Storytelling

Quality data and key insights only create impact when communicated effectively. This course covers the principles of choosing the right analyses, scientific visualization, and how to use Pluto's reporting and customization tools to tell a clear, compelling story for any audience.

This course is broken into five modules, each split into short topics with a learning check after every section. Most topics take 2 to 4 minutes. Your progress saves automatically, so you can close the tab anytime and pick up right where you left off.

Modules

• Module 3.1A real-world scientific problem2 topics
• Module 3.2Choosing and Interpreting Analyses6 topics
• Module 3.3Introduction to Multi-Omics Analysis8 topics
• Module 3.4Principles of Scientific Visualization4 topics
• Module 3.5Exporting, Reproducing, and Sharing Your Work6 topics

Background

Non-small cell lung cancer (NSCLC) is the most common form of lung cancer, comprising 80-85% of all lung cancer cases. Osimertinib is a third-generation epidermal growth factor receptor (EGFR) tyrosine kinase inhibitor (TKI) used to treat non-small cell lung cancer. Osimertinib acts by activating EGFR mutations - most commonly exon 19 deletions or the L858R substitution - including in tumors that have already acquired the T790M gatekeeper mutation that makes them resistant to earlier-generation EGFR TKIs. Osimertinib produces dramatic, durable responses in most patients, but acquired resistance is still common. In the latter case, a subpopulation of tumor cells instead enter a reversible drug-tolerant persister (DTP) state: a slow-cycling, transiently drug-insensitive condition maintained by changes in gene expression and chromatin state rather than by changes in DNA sequence.

DTP cells survive the initial drug exposure and can later revert to a drug-sensitive state if treatment is paused, or serve as a reservoir from which permanent, mutation-driven resistance eventually emerges under continued drug pressure. Because the DTP state is reversible and non-genetic, it is an attractive target for strategies aimed at preventing resistance before it becomes fixed, rather than treating resistance after the fact. To study this experimentally, EGFR-mutant NSCLC cell lines were treated with osimertinib for an extended period, and the surviving, drug-tolerant cells are then profiled and compared back to untreated cells of the same line.

Throughout this course, you will be a researcher in this field where you will work to characterize the DTP state using two paired, publicly available datasets generated from exactly this kind of experiment and deposited in NCBI's Gene Expression Omnibus (GEO): an RNA-seq dataset (GSE193258) capturing transcriptional changes in the DTP state, and an ATAC-seq dataset (GSE193256) capturing changes in chromatin accessibility, the regions of open, transcriptionally accessible DNA, in the same system. Analyzing both layers together lets you ask not just which genes change expression in DTP cells, but whether those changes are accompanied by a matching change in chromatin state at the same loci, a concordance that strengthens confidence a hit is biologically real rather than an artifact of a single assay. These datasets are part of a larger study by Criscione et al. 2022, “The landscape of therapeutic vulnerabilities in EGFR inhibitor osimertinib drug tolerant persister cells,” which profiled DTPs across multiple EGFR-mutant NSCLC cell lines and linked chromatin remodeling to a DTP transcriptional program involving EMT-associated transcription factors (ZEB1, ZEB2, SLUG), Hippo pathway/YAP-TEAD activity, and TGF-β/SMAD signaling. SERPINE1 (PAI-1), a MEK compensatory-resistance gene, gains both accessibility and expression in DTPs; MAPK13 (p38δ) loses both. For additional scientific background beyond what's covered here, the full paper is available at https://pmc.ncbi.nlm.nih.gov/articles/PMC9794691/.

Set up your analysis space

We will build on this theme throughout the course using a newly made analysis space. Follow those steps to create your own analysis space in a certification-focused program.

Navigate to the program named ‘Pluto Certification program’ by using the search bar at the top of the page

Navigate to the program’s ‘Analysis spaces’ and click ‘ + Analysis space’ to create a new analysis space. Name the analysis space using your initials and course 3.

Choose to link a dataset

First, link the RNA-seq dataset that characterizes transcriptional differences in osimertinib-resistant cell lines. Select the “Global Data Commons” tab and search by the GEO accession ID, GSE193258.

Click on the dataset and repeat the steps to add the associated ATAC-seq dataset, GSE193256. Once both datasets have been selected, click “Link datasets”.

We can now begin adding new and existing analyses. Existing analyses mean you will be adding analyses already created in the linked public datasets while new analysis will be created here from scratch. Throughout this course, we will use a combination of them.

Welcome & Objective

Throughout this module, we will introduce some common analysis types, discuss their use and interpretation, and provide hands-on exercises that will build out these analyses in the newly created Osimertinib resistance analysis space.

The Typical RNA-seq Analysis Workflow

Most RNA-seq experiments, regardless of the specific biological question, tend to move through the same initial general analysis path. Walking through it once end-to-end will make each individual analysis in this module easier to place in context: principal component analysis (PCA) → save variable genes → clustering heatmap → differential expression → gene set enrichment analysis (GSEA).

PCA: an unsupervised first look at the whole dataset, used to explore overall sample variability, spot outliers, and identify covariates that should be accounted for later.

Save variable genes: the genes driving most of that variability are saved as a defined gene list, keeping downstream exploratory analyses focused on the genes that actually differentiate your samples rather than noise.

Clustering heatmap: those variable genes are visualized as a heatmap to confirm, visually, whether samples group the way you'd expect biologically, and not by a technical variable.

Differential expression: a formal statistical comparison between two defined groups, producing a fold-change and adjusted p-value for every gene.

GSEA: takes the full ranked differential expression result and asks whether entire pathways, not just individual genes, are coordinately shifted.

This flow is introduced here using RNA-seq, but the same or similar analyses apply to most other quantitative -omics data types as well, including the ATAC-seq dataset you linked back in Module 3.1.

Principal Components Analysis (PCA)

You were introduced to PCA earlier in this certification as a way to reduce a large, high-dimensional dataset down to a handful of principal components that summarize most of the variability between samples. Here, we'll focus specifically on how PCA fits into an RNA-seq workflow.

PCA is typically run on a defined, top set of genes with the most variable expression across your samples rather than on every gene measured. Restricting PCA to variable genes reduces noise from genes that don't meaningfully change across your dataset and keeps the components focused on the factors actually driving variability.

Beyond a first look at your data, PCA is a useful diagnostic tool for two related purposes: describing where variability in your dataset comes from and identifying covariates that might be worth including in later analyses like differential expression. A covariate is simply a variable other than your primary variable of interest that may also explain variation in your data, such as batch, cell line, or sequencing run.

If you already suspect a technical batch effect ahead of time, batch correction can optionally be applied before running PCA. Pluto can incorporate limma's removeBatchEffect function to statistically remove the variance associated with a known batch variable prior to analysis, which can make it easier to see the biological signal of your biological groups in the PCA plot. This is an optional step, most useful when you already suspect a batch effect (e.g., samples processed in different rounds), rather than something to apply by default.

🧪 HANDS-ON EXERCISE: Explore the DTP RNA-seq dataset with PCA

Steps:

1. From your analysis space, create a new analysis by clicking on the plot symbol in the upper left.

1. Click on the RNA-seq experiment (GSE193258) because this analysis will be related to only this dataset.

1. From the available analyses, choose Principal components (PCA) and create the analysis

1. We will keep all default settings except under Advanced settings, we will choose to run the analysis on the 2,000 most variable genes. Now, you can run the analysis.

1. Once the analysis completes, we can open up the plot editing options. If it isn’t already, select to label the plot by Cell Line.

1. Finally, save the variable genes as a list which we will utilize in the next analysis.

👀 What to look for: Do samples separate primarily along PC1 and PC2 by treatment condition (DMSO vs. DTP), by cell line (H1975, HCC2935, HCC827, PC9), or both? If cell line is driving a large share of the variability, that's a signal to include cell line as a covariate in the differential expression analysis later in this module.

🖼️ Expected plot

Clustering Heatmaps Using Select Genes

PCA gives you a low-dimensional summary of variability, but it's often useful to also look directly at the expression patterns of the genes driving that variability or at a specific set of genes that you are interested in. A clustering heatmap displays genes as rows and samples as columns, with cell color representing expression level, typically row-scaled (z-scored) so each gene's own high/low pattern is visible regardless of its overall abundance. Hierarchical clustering is commonly applied to both rows and columns, grouping genes with similar patterns together and revealing whether samples cluster the way you'd expect biologically.

In this topic, we'll build that heatmap using the variable gene list you saved in the PCA topic but this list of genes could come from anywhere.

🧪 HANDS-ON EXERCISE: Cluster the DTP Dataset Using Variable Genes

Steps:

1. Follow similar steps to the previous exercise but this time selecting the Clustering analysis in the RNA-seq dataset.

1. To set up the analysis, select followed list and choose the list we saved during the PCA step. It will likely start with GSE193258. Additionally, let’s select to analyze specific groups of samples and then select Broad Treatment. This will group the samples by their different resistance states. We can select all of the groups to ensure that all samples will be included but they will be labeled as these treatment groups. We will keep all of the data transformations the same as default.

1. Finally, we can run the analysis. When it finishes, take a look at the plot.

👀 What to look for: Do the sample clusters (the column dendrogram) match the grouping you saw in your PCA plot? Do they cluster well by Broad Treatment group? What about if you change the plot labels to Cell Line? Does it match what you saw in the PCA plot?

🖼️ Expected plot

Differential Expression Analysis

Differential expression analysis is the statistical heart of most RNA-seq workflows: it identifies which genes differ significantly in expression between two defined groups (e.g., treatment vs. no treatment), producing a fold-change (magnitude and direction) and an adjusted p-value (statistical confidence) for every gene tested. Using these statistics, you can act on them directly or hand off to a pathway-level analysis like gene set enrichment analysis (GSEA).

P-values, Adjusted P-values, and Fold-Change

A p-value answers a narrow, specific question: if there were truly no difference between your groups (the null hypothesis), how likely would you be to see an effect at least this extreme, just by chance? It is not the probability that the null hypothesis is true, and on its own it says nothing about how large or biologically meaningful the effect is.

In a typical differential expression analysis, you're not testing one gene; you're testing thousands or even tens of thousands of genes simultaneously. A raw p-value cutoff of 0.05 across 20,000 tested genes would produce roughly 1,000 falsely “significant” genes by chance alone. Adjusted p-values (also called FDR or padj, most commonly via Benjamini-Hochberg correction) control for this multiple-testing problem: a padj cutoff of 0.05 means that, among the genes you call significant, you expect roughly 5% to be false positives, a controlled error rate across the whole gene list rather than per individual test.

Fold-change describes the magnitude and direction of a difference (commonly reported as log2 fold-change), independent of statistical confidence. A gene can have a tiny but highly reproducible fold-change that reaches significance with a large enough sample size, or a large fold-change that fails to reach significance because of high variability or a small sample size. Because significance and magnitude answer different questions, you can apply both a significance threshold (e.g., padj < 0.05) and an effect-size threshold (e.g., |log2FC| > 1, a 2-fold change) before calling a gene meaningfully differentially expressed.

It's also important to know which direction a fold-change is reporting. It stems directly from the order of your group comparison. A positive log2 fold-change means a gene is higher in whichever group you set as the experimental group relative to the control group, so comparisons should always be set up as experimental vs. control, for example mutant vs. wild-type or treated vs. untreated, to keep that direction consistent and interpretable throughout your results.

Covariates in Your Differential Expression Model

A covariate is a variable other than your primary variable of interest that may also explain variation in your outcome, things like sequencing batch, cell line, sex, or age. Including a known covariate in your statistical model lets you account for (statistically remove) the variance it explains, which can both increase your power to detect the effect you actually care about, and prevent that variable from generating false positives if it happens to be correlated with your groups of interest. This is exactly the decision your PCA plot helps you make: if PCA showed a variable like cell line explaining substantial variability, that's a strong candidate to include as a covariate here. Including too many covariates relative to your sample size can reduce your power to detect real differential expression genes, since every covariate consumes degrees of freedom, so it's worth only including the ones your PCA exploration flagged as meaningful.

Filtering Genes Before Testing

Before running differential expression, it's standard practice to filter out genes with very low or inconsistent expression. Testing these genes anyway adds noise, worsens the multiple-testing burden described above, and can produce unstable, misleading fold-change estimates for genes that are barely being measured at all.

Pluto lets you set a minimum read count and a minimum percentage of samples meeting that count, independently for your experimental and control groups; a gene passes the filter if it meets your thresholds in at least one of the two comparison groups. Pluto's platform default is 3 reads in at least 20% of samples in any group, a good general balance of sensitivity and reliability for most studies if you have a high number of samples. If you only have a handful of samples, you may want to raise the minimum sample percentage to roughly 60% to ensure that multiple samples express a given gene for it to be tested.

Visualizing DE Results: Volcano Plots and Heatmaps

A volcano plot puts log2 fold-change on the x-axis and −log10(adjusted p-value) on the y-axis. Points toward the top of the plot are more statistically significant; points toward the left or right edges have larger effect sizes. Genes that are both far from center and high up are your strongest candidates, letting you see at a glance how many genes moved in each direction, and how extreme those changes were, before looking at any single gene individually.

A heatmap can also be used here, exactly like the clustering heatmap from the previous topic, but built from your differential expression result's top genes instead of your PCA variable genes. A differential expression result that “looks real” should generally produce a heatmap where the top genes visibly separate your comparison groups; if they don't, it's worth checking for a confounding technical variable or an underlying QC problem.

🧪 HANDS-ON EXERCISE: Run Differential Expression on DTP long washout vs. DMSO

Steps:

Long washout means DTPs were created (21 days of osimertinib) and drug was removed until exponential proliferation resumed (about 1 week out of drug). In this exercise, we will compare these resistance cells to the control DMSO group.

1. Similar to before, create a new analysis in the RNA-seq experiment but this time, choose Differential expression.

1. In the analysis set up, we will select Broad Treatment as the variable to compare. The experimental group will be ‘osimertinib DTPs long washout’ and the control group will be ‘DMSO’. We will customize the next two sections by requiring at least 10 reads in at least 60% of samples in at least one of the groups and we will include Cell Line as a covariate because our analyses so far have shown that is an important covariate to include in this dataset. Finally, we can run the analysis.

1. After the analysis completes, you will see a volcano plot of the results. Using the Plot tab, you can view a heatmap of the top differentially expressed genes as well.

👀 What to look for: Do any of the genes mentioned in Module 3.1's background reading, SERPINE1, ZEB1, or MAPK13, show up among your top hits? These can be searched in the results table and/or labeled in the plot using the Labeled features section.

🖼️ Expected plot

Gene Set Enrichment Analysis (GSEA)

Gene set enrichment analysis (GSEA) takes a different approach from single gene differential expression testing. Instead of applying a significance cutoff gene-by-gene, GSEA ranks every measured gene by a chosen statistic and asks whether a predefined gene set (a known pathway or biological process) is enriched toward the top or bottom of that entire ranked list, more than would be expected by chance.

This matters because many real biological effects are coordinated, subtle shifts across dozens of genes in a pathway, where no single gene individually clears a strict significance threshold, but the pathway, as a whole, shows a clear, consistent directional signal.

Ranking Genes and Choosing a Gene Set Collection

GSEA requires every tested gene to be ranked along a single axis, from most “up” to most “down” in your comparison, before it looks for pathway-level enrichment. Common ranking statistics include log2 fold-change (default in Pluto) alone or a combined score from both the fold-change and p-value.

Gene sets themselves come from curated collections, most commonly the Molecular Signatures Database (MSigDB). A gene set is simply a list of genes known to be associated with a biological pathway or process, for example, the Hallmark “Hypoxia” gene set contains around 200 genes known to increase expression in low-oxygen conditions. MSigDB organizes thousands of these gene sets into collections; the Hallmark collection is a good general starting point, with more specialized options (curated pathways, immunologic signatures, and others) available depending on your question. Since Module 3.1's background reading tied DTP formation to EMT, Hippo/YAP-TEAD activity, and TGF-β/SMAD signaling, the Hallmark collection is a reasonable first place to look in this exercise.

Reading a Single Enrichment Plot

The enrichment plot shows a running enrichment score as GSEA walks down the ranked gene list, increasing when it hits a gene from the set of interest and decreasing otherwise. The peak of that curve is the enrichment score (ES); the Normalized Enrichment Score (NES) adjusts for gene set size so that different pathways can be fairly compared to one another. The leading edge is the subset of genes contributing most to the enrichment signal, often the most interesting genes to follow up on individually, since they're both part of the enriched pathway and near the top of the overall ranking.

Comparing Many Gene Sets at Once

If you want to scan many gene sets at once rather than inspect them one at a time, GSEA results can be displayed as a bar plot of Normalized Enrichment Score (NES) across gene sets, ranked from most positively to most negatively enriched. NES, rather than the raw enrichment score, is what allows different gene sets, which can vary enormously in size, to be fairly compared to one another on the same plot.

🧪 HANDS-ON EXERCISE: Run GSEA on the DTP long washout vs DMSO Comparison

Steps:

1. Similar to before, create a new analysis in the RNA-seq experiment but this time, choose Gene set enrichment.

1. Select the previously created comparison and choose the MSigDB collection Hallmarks. We will adjust the gene ranking method to Sign(log2 fold change) * -log10(p-value). This ranking method will essentially rank genes by significance but takes into account the direction of the fold-change. Finally, you can run the analysis.

1. Once the analysis completes, the plot will display the most significant gene set. We will change the displayed gene set to Hallmark Epithelial Mesenchymal Transition because it is highly significant and we expected EMT to change.

We can also choose to display the adjusted p-value and NES on the plot. Save changes to update the plot.

👀 What to look for: Do any EMT-, Hippo/YAP-TEAD-, or TGF-β/SMAD-related Hallmark gene sets (e.g., “Epithelial Mesenchymal Transition”) show a strong positive NES? Check the leading edge genes for that gene set, do you recognize any of them from the Module 3.1 background reading?

🖼️ Expected plot

Welcome & Objective

What Are the Pros and Cons of Multi-Omics Analysis?

Every omics layer measured on its own tells you part of the story. Multi-omics analysis combines two or more layers, most often on the same samples, to build a more complete, cross-validated picture. That power comes with real added cost and complexity.

✅ Pros

• More complete biological picture: cross-validates a hypothesis across layers rather than relying on one (e.g., confirming a DNA-level change actually shows up at the RNA or protein level).

• Can reveal regulatory mechanisms invisible to any single layer, by connecting what changed (RNA/protein) to why it changed (epigenomic/regulatory layer).

• Strengthens confidence in a finding: an effect seen independently in two orthogonal data types is far less likely to be a technical artifact of either one.

⚠️ Cons

• Higher cost and complexity: generating and integrating multiple omics layers takes more time, budget, and specialized statistical methods than a single-layer study.

• Sample sizes and coverage often differ across layers (e.g., if fewer samples have matched proteomics than RNA-seq), complicating joint statistical analysis.

• Batch effects can compound: each additional omics platform brings its own technical variability, all of which needs to be accounted for before integration.

Multi-Omics Considerations

Before integrating two or more omics datasets, a few compatibility checks matter more than any statistical method you'll apply afterward. If these don't line up, no downstream analysis can fix it.

🧬 Species

Are both datasets generated in the same species? Integrating across species (e.g., mouse RNA-seq with human proteomics) requires ortholog mapping, which is imperfect: not every gene has a clean one-to-one ortholog, and some biology simply doesn't translate across species.

🧭 Genome Version

Gene IDs, coordinates, and annotations shift between genome builds and annotation releases (e.g., GRCh37 vs. GRCh38, or different Ensembl/RefSeq versions). Two datasets processed against mismatched builds can silently prevent features from mapping when integrated by ID or genomic coordinate.

🧪 Experimental Conditions

Are the two omics layers measuring the same underlying biological state: matched samples, same timepoint, same treatment condition? Integrating unmatched cohorts profiled under different conditions compares two different biological questions, not two views of the same one.

ATAC-seq Differential Binding Analysis

So far, every analysis in this module - PCA, clustering, differential expression, GSEA - has used only the RNA-seq experiment of the osimertinib resistant cell lines. The ATAC-seq experiment (GSE193256) you linked back in Module 3.1 has been sitting alongside it, unused, until now. This experiment measures chromatin accessibility in the same cell lines and similar osimertinib resistant states as the RNA-seq experiment. One difference to note between the DTP long washout group in the RNA-seq experiment and the DTP group in the ATAC-seq experiment is that the long washout group in the RNA-seq has a period of time in between drug and when the sample was harvested. The ATAC-seq DTP group contains samples that were harvested after they are resistant but while they are still receiving drug.

The equivalent of differential expression for a chromatin accessibility dataset is called Differential Binding Analysis in Pluto, available for ATAC-seq, ChIP-seq, and CUT&RUN experiments. It runs on the same underlying logic: for each feature tested, it produces a fold-change and an adjusted p-value between two groups. The difference is what's being tested. Instead of named genes, an ATAC-seq experiment tests peaks: discrete genomic regions of pileup of reads identified during peak calling. For an ATAC-seq experiment, these peaks are considered regions of increased chromatin accessibility. Because a peak is a genomic region rather than a gene symbol, Pluto annotates each peak to its nearest or overlapping gene, which lets you connect a differentially bound peak back to a named gene, and to your RNA-seq result, in the next two topics.

Filtering and Covariates for Peak-Level Testing

The same feature-filtering logic from the differential expression topic carries over directly: peaks with very low or inconsistent read counts are filtered out before testing, for the same reasons (noise, an inflated multiple-testing burden, and unstable fold-change estimates). The same is true for covariates. Cell line explained meaningful variability in your RNA-seq PCA exploration, and since the ATAC-seq experiment was generated from the same cell lines and design, it's reasonable to include it here as well rather than re-running the full PCA-to-clustering exploration from scratch.

🧪 HANDS-ON EXERCISE: Run Differential Binding Analysis on DTP vs. DMSO

Steps:

This is similar to the comparison you ran for RNA-seq differential expression, applied to the ATAC-seq experiment instead.

1. From your analysis space, create a new analysis and this time, select the ATAC-seq experiment (GSE193256), since this analysis is specific to that dataset.

1. Choose Differential Binding Analysis from the available analyses.

1. Select Treatment as the variable, with “osimertinib DTPs” as the experimental group and “DMSO” as the control group. Under the Advanced settings, increase the feature filter to at least 10 reads in at least 60% of samples in at least one of the two groups, and include Cell Line as a covariate, matching your RNA-seq analysis. Finally, run the analysis.

1. Once it completes, you'll see a volcano plot of differentially bound peaks; use the Plot tab to switch to a heatmap of the top differentially bound peaks if you'd like.

👀 What to look for: Peaks are automatically annotated to their nearest gene. Do you see any peaks annotated to SERPINE1 or MAPK13, the same genes flagged in your RNA-seq result?

🖼️ Expected plot

Overlap Analysis

With a differential result now available in both the RNA-seq and ATAC-seq datasets, Overlap Analysis is the simplest way to start combining them. It generates a Venn diagram from up to three gene (or target) lists, letting you see at a glance which genes show up as significant in more than one list at once.

In a multi-omics context, overlap answers a specific question: which genes are both transcriptionally changed (flagged by RNA-seq differential expression) and annotated to a differentially accessible peak (flagged by ATAC-seq differential binding)? A gene appearing in both circles has independent support from two orthogonal assays, a substantially stronger candidate for follow-up than either list on its own.

🧪 HANDS-ON EXERCISE: Overlap Your RNA-seq and ATAC-seq Gene Lists

Steps:

1. From your analysis space, create a new analysis. This time however, we will create an analysis under the section this analysis space because we want analyses that will take into account multiple datasets.

1. Now, select overlap genes/peaks/proteins analysis.

1. You will then need to define each differential list to include in the overlap analysis. For the first list, define the RNA-seq genes. Select the DTP long washout vs DMSO comparison and then define the thresholds as adjusted p-value of 0.01, only genes in the up direction and an Upper bound requirement of 1. Then save this list. What you have created is a list of all genes that are upregulated in DTP long washout with an adjusted p-value of at least 0.01 or lower.

1. Next, create the ATAC-seq list in the same manner, save, and then run the analysis.

1. You will then see a Venn diagram of the results with hundreds of genes overlapping in the middle where they are significantly upregulated by RNA-seq and have at least one enriched peak annotated to the same gene. You can view the genes in each category using the results table below the plot.

👀 What to look for: How many genes fall in the overlap? Do SERPINE1 and/or MAPK13 appear in both experiments?

🖼️ Expected plot

Concordance Analysis

Overlap only tells you whether a gene cleared the significance bar in both layers; it doesn’t look globally to see if the two layers agree broadly on direction and magnitude. Concordance Analysis fills that gap with a scatter plot: one comparison's fold-change on the x-axis, the other's on the y-axis, one point per gene measured in both. A strong positive correlation, points falling along the diagonal, supports a coherent regulatory story: genes that gain accessibility tend to also gain expression, and vice versa.

This type of analysis can highlight genes that break the expected patterns. A gene that gains chromatin accessibility without a matching change in expression (or vice versa) can be flagging real regulatory nuance, a poised-but-not-yet-active enhancer, or a post-transcriptional layer of control that a single-omics view would never have surfaced.

🧪 HANDS-ON EXERCISE: Plot RNA-seq vs. ATAC-seq Fold-Change for the DTP Comparison

Steps:

1. Similar to the previous exercise, choose to create a multi-omics analysis and then specifically, a concordance analysis.

1. Set up the analysis to contain the ATAC-seq comparison on the x-axis and the RNA-seq comparison on the y-axis and run the analysis. Matching on peak IDs puts one point per peak, so a gene with several peaks appears many times.

1. After the analysis completes, switch to the Plot tab to color SERPINE1 and MAPK13.

👀 What to look for: Do most genes fall along the diagonal? Specifically, does SERPINE1 sit in the upper-right quadrant (up in both RNA-seq and ATAC-seq) and MAPK13 in the lower-left (down in both)?

🖼️ Expected plot

The Analysis Space as a Multi-Omics Canvas

Overlap and concordance are both statistical summaries across your entire gene list. Sometimes the fastest way to build intuition for one specific gene is to just look at it directly. Because your analysis space already has the RNA-seq and ATAC-seq experiments linked side by side, you can create an analysis from either dataset and arrange the results next to each other on the same canvas, turning the analysis space itself into a shared, gene-level comparison view.

We will demonstrate this with a Peak/Signal Analysis, which displays bigWig tracks (one track per sample) in an IGV-style genome browser plot. Using this analysis type, you can compare read coverage across the genome for your experiment. You can search for a specific gene or locus, position the view where you want it, and save that view so it's there the next time you open the plot.

🧪 HANDS-ON EXERCISE: Compare RNA-seq and ATAC-seq Signal Side by Side

Steps:

SERPINE1 gained both expression and accessibility in the DTP state, and MAPK13 lost both, per Module 3.1's background reading. Let’s plot the coverage of SERPINE1 in these two experiments

1. From your analysis space, create a Signal Analysis using the RNA-seq experiment.

1. To simplify the plot, we will subset the samples to just the H1975 cell line and just the first replicate of the osimertinib DTPs long washout and the DMSO samples.

1. Once the plot loads, search for SERPINE1 within the plot.

1. Now that we have the view that we want, switch to the plot tab and set this as the current view. Then, save your changes.

1. Perform steps 1 - 4 using the ATAC-seq experiment. You will select the Peak analysis instead of signal analysis.
2. You can resize and rearrange these plots on the canvas and compare them.

👀 What to look for: For SERPINE1, do the RNA-seq tracks show taller signal in the DTP samples than DMSO, and do the ATAC-seq tracks show a similarly taller, more accessible peak near the gene's promoter in DTP? If you shift the ATAC-seq view further upstream of SERPINE1, what do you see?

🖼️ Expected plots

Additional options for multi-omics analysis

While these dedicated multi-omics analyses are valuable, most analysis types can offer multi-omics value. Target and biomarker lists can be saved for any experiment, and these lists can be applied to analyses of other experiments. For instance, you can label the significant genes in an RNA-seq experiment in the volcano plot of an ATAC-seq experiment that shows differential binding. You can even use a gene list as a test gene set for GSEA analysis.

Welcome & Objective

Let's come back to your role as a researcher in this course. You've spent Modules 3.1 through 3.3 characterizing the DTP state in EGFR-mutant NSCLC across the linked RNA-seq and ATAC-seq datasets, and you're now preparing a slide deck for the next internal resistance prevention program stage-gate review, which is the meeting that decides whether your findings advance the target and model characterization work into mechanism and selectivity profiling. You walk in with a dense heatmap: the full DTP transcriptional signature across dozens of genes and both treatment states, a color scale that took real effort to tune. Ten minutes later, the conversation is about what the color scale means, not about whether SERPINE1 and the rest of the EMT, Hippo/YAP-TEAD, and TGF-β/SMAD program are strong enough evidence to advance the target. The meeting ends with 'let's revisit this next time,' and the stage-gate decision gets pushed.

This isn't necessarily a failure of the science: it may reflect a translation gap, which is one of the more common bottlenecks in drug discovery. Programs are typically tracked through defined stages (e.g., discovery, target and model characterization, mechanism and selectivity profiling, and translational evidence building), and a stage-gate figure has one job: telling the room whether the evidence is strong enough to advance. A dataset can be perfectly clean with a statistical test that is perfectly appropriate, and, yet a conclusion can still fail to land because the figure carrying it was built for the analyst who made it, not for the decision the room needed to make.

These next modules are about closing that gap. Module 3.4 covers what makes a figure work in the first place. Module 3.5 covers how to customize and export figures directly from Pluto so they are ready for whatever room you are walking into. Module 3.5 also covers Pluto's methods and sharing tools so your work can be checked, repeated, and trusted. It brings together into a full narrative arc, from question to conclusion, that you can use to structure your stage-gate presentation of these findings.

What Makes a Good Scientific Figure

A good scientific figure does three things at once: it is clear, it is accurate, and it carries an appropriate amount of complexity for its audience. Most figures that fail to land do so because they manage only one or two of the three.

Clarity

A viewer should identify the main takeaway within a few seconds, without narration. The point of the figure should be its most visually prominent element, rather than buried in a legend or a footnote. If a figure requires a presenter to say something like, 'ignore the top left panel, focus on the bottom right,' the figure could be improved.

Accuracy

Clarity without accuracy is a confident-looking mistake. This means that axes should reflect where the data start, color scales should be appropriate for the data type, and error bars and sample sizes should be visible. Consider a figure with a truncated y-axis; adjusting the axes can make a small effect look dramatic. Similarly, if the data are showing diverging data, like a heatmap, the color palette should appropriately diverge as well. A figure that is easy to read but misrepresents the underlying data is dangerous because it can easily be misinterpreted.

Appropriate complexity

The right level of complexity depends entirely on the audience: a figure for a specialist conference can carry more nuance and more variables than one meant for a general audience talk. Every added variable, panel, or color category is a small tax on the viewer's attention. That tax is worth paying when it adds information the audience needs for the decision at hand, and not worth paying otherwise.

💡 Try it yourself - the caption test

The figure panels below was taken from Criscione et al., 2022 (Figure 1D). Cover the caption and ask a colleague unfamiliar with the project what they think it shows. If their answer matches the intended takeaway, the figure works. If not, the fix is almost always to simplify, not to add more explanation.

Left panel: gene set variation analysis (GSVA) scores for Reactome DNA strand elongation pathway (two-sided t test, p value * < 0.05, ** < 0.01, *** < 0.001). Boxplot is quartiles with range bar as minimum or maximum data values within 1.5 times the interquartile range. Right panel: Same as left panel, GSVA scores for GO extracellular matrix pathway.

Matching Plot Type to Message

The same differential expression result can be shown as a volcano plot or a heatmap, and each answers a different question even though the underlying statistics are similar.

Compare these two types of plots where MAPK13 and SERPINE1 are highlighted in the DTP RNA-seq experiment.

Volcano plot:

Cluster analysis heatmap:

The volcano plot does a good job of showing global changes between these two groups. It also allows highlighting of our two genes so we can see how those changed and how they change in relation to other genes. Meanwhile, this heatmap is only showing those two genes so we don’t have the global context. However, we are now able to see how variable the expression of these genes is in these groups of samples. Additionally, we can add other groups in the dataset so we can see how these genes are expressed in the acutely treated group. This plot demonstrates that these genes generally change with treatment but that the change is variable (cell line plays a big role here) across sample. Both plot types offer value and their use depends on the question you are seeking to answer.

Tailoring Figures for Different Audiences

The same result can call for different figures depending on who is in the room. A computational peer can interpret a volcano plot's axes at a glance and may want more statistical detail visible on the figure itself. A stage-gate review usually needs fewer points on the page, larger labels naming the genes or targets that matter, and a title that states the conclusion rather than the method. An external or public-facing audience needs the figure to carry even more of the explanation on its own, since no analyst will be standing next to it to answer questions.

💡 Try it yourself – matching messaging to audience

The figure panel below was taken from Criscione et al., 2022 (Figure 5B). Review the ‘take-away messages’ below and match them to the appropriate audience.

Same figure, same underlying result, but four very different ways to talk about it. Tailoring changes what's emphasized and how much is explained; it never changes the underlying finding.

⤢ Enlarge

Criscione et al. 2022, npj Precision Oncology 6:95, Fig. 5b, CC-BY 4.0. DOI: 10.1038/s41698-022-00337-w

Same figure. Four captions below, each written for a different reader. Drag each one into the audience it matches, then check your answers. (Keyboard or touch: select a caption, then select an audience.)

Technical Peercomputational / wet-lab colleague

Drop a caption here

R&D Team Leadscientist managing multiple programs

Drop a caption here

Business Stakeholderleadership / portfolio review

Drop a caption here

General Audiencenon-scientist

Drop a caption here

GSVA scoring of the MEK activation signature shows significant suppression in osimertinib DTPs vs. DMSO (two-sided t test, p < 0.001), consistent with on-target EGFR pathway inhibition. The MEK compensatory-resistance signature shows the inverse pattern (p < 0.001), but critically, it does not change during acute (24h) treatment: the increase is specific to the established DTP state, not an immediate consequence of EGFR blockade.

The mechanism here is a pathway handoff, not just a shutdown. Osimertinib suppresses the MEK activation program as expected. But DTPs concurrently upregulate a distinct MEK compensatory-resistance program, which is silent during acute treatment and only switches on once the persister state is fully established. That delay is the mechanistic backbone of our screen result: an upfront MEK combination hits before the compensatory program is online, while a sequential combination arrives after it's already compensating. This reframes MEK inhibitor timing as a resistance-prevention question rather than just a combination-efficacy question.

When these lung cancer cells become tolerant to treatment, they don't simply power down. Over about three weeks, they activate a backup resistance program. Because that switch takes time to appear, there's a window to act early. This is the data behind our recommendation to prioritize an upfront combination strategy over adding a second drug later.

Cancer cells facing this drug don't just shut down and wait it out. Turns out, they slowly build a workaround over a few weeks. The good news: that workaround doesn't appear right away, so there's a real window to block it with a second treatment before resistance fully takes hold.

Check my answers

Welcome & Objective

Module 3.4 covered how to think about a figure. This module covers how to build it in Pluto, generate the accompanying methods, and share it, so all your data are ready for a manuscript, a stage-gate deck, or a leadership summary without a round trip through a separate design tool.

Customization Options

Pluto's plot customization tools control colors, labels, and axes or legends. These are the elements that most determine whether a figure reads correctly.

Colors and themes

New plots start on Pluto's default cool color theme, though warm and monochrome themes are available from the Edit Plot menu. Plots can be colored by one or more variables from your uploaded sample data (e.g., treatment or condition). Individual colors can be set from the active theme, saved favorites, your organization's shared palette, a gradient picker, or a manually entered hex code.

Labels and legends

Legend display names can be edited and reordered directly in the Edit Plot menu by dragging entries or clicking the pencil icon on a name. This keeps labels tied to the live plot rather than to a static image edited after export. Legend editing is not available on every plot type: volcano plots, enrichment plots, score barplots, and certain image-based plots (tornado, profile) do not support it since their legends are not derived from uploaded sample groups the same way.

For step-by-step instructions, see Edit plot colors and legends and Reorder plot legends in the Pluto help center.

On a volcano plot specifically, individual targets can be labeled by searching for them in the Labels section of the Edit Plot menu, which is the same mechanism used for the lasso-select workflow in Module 3.4.

Axes

Axis ranges and scales should represent the data honestly, per the accuracy principle from Module 3.4. These ranges can be edited directly in the plot editing view.

🖼️ Example in Pluto

Pluto Edit Plot menu showing color theme selection, color-by-variable options, and the Legend section with draggable, editable display names, for a plot from the DTP RNA-seq dataset (GSE193258).

Exporting Figures at the Right Size and Format

Every plot in Pluto is exported from the same place: open the plot's '...' menu and select Export. The export view is a canvas where you set the final size and layout of the figure before you download it. Making some changes here can be the difference between a figure that drops cleanly into a slide or journal template and one that needs rework in another tool.

Size first, then arrange

1. Open the plot's '...' menu and select Export.

2. Set the canvas size. Drag the canvas handles, or type exact pixel values into the Width and Height boxes at the top of the export view. Use the lock icon to keep the aspect ratio while you resize.

3. Place the legend after the plot is sized. Moving the legend can resize the plot area, so settling the size first saves rework. The legend can run vertically or horizontally.

4. Trim empty space with Fit to content.

5. For print, turn on Optimize axis styles for publication, which switches all axis labels to pure black for higher contrast.

6. Download or export in the format your destination needs (see below).

Why canvas size matters

A larger canvas gives labels room to breathe. If gene names or sample labels are crowded or unreadable, enlarge the canvas in the export view (for example, to 1000 x 1000 px). For PNG files, a larger pixel size also produces a sharper image. If you go large, increase the point size in Edit Plot before exporting so data points stay visible at the new scale.

Choosing a file format

• Manuscript or poster: a large PNG (for example, 1400 px or wider) keeps the image sharp in print.

• Slide deck, stage-gate review, or written report: Export to PowerPoint or Export to Word. The plot + methods option carries the experiment details with the figure.

• Re-plotting in another tool (for example, GraphPad Prism): download the Results CSV from '...' > View results.

🖼️ Example in Pluto

The plot's '...' menu, where Export starts, and the export view with Width and Height, Optimize axis styles for publication, and the download options.

Reproducibility and Traceability in Pluto

A result is only as trustworthy as the ability of someone else to reproduce it. That means the exact software versions, parameters, filtering thresholds, and normalization steps behind a result need to be recorded. Vague methods are a common reason a promising finding cannot survive scrutiny during due diligence, peer review, or an internal replication attempt, and a stage-gate decision built on a result nobody can fully reconstruct is a fragile decision.

Pluto records methods for you as you work, generated from how each result was actually made. Because the text comes from the live analysis rather than being typed from memory, it updates if a threshold changes or a step is rerun.

Methods at three levels

Analyses are versioned

Analyses in Pluto are versioned. When a newer version of an analysis is available, a green banner appears on it. Upgrades are not automatic because you may already have exported results from the earlier version. Clicking Run analysis upgrades it, and earlier results stay available exactly as originally run.

🖼️ Example in Pluto

Pluto 'View Methods' popup for a plot from the DTP RNA-seq analysis (GSE193258)

Trace what changed

Traceability answers other key questions that come up during audits: What changed, and when? Which processing run is this? Which version of the analysis is this? Pluto answers each of these in a specific place, and because datasets, pipelines, analyses, and their histories live in one platform rather than scattered across documents, drives, and teams, each answer is a few clicks away instead of a search through old files and inboxes.

Iterate without losing the original

• Clone a plot to try different parameters and compare the versions side by side.

• Archive rather than delete. Archived experiments are not deleted and can be found again with the "is archived" filter.

Why Share Your Analysis in Pluto

Most omics results travel as static files: a PNG in an email, a slide in a deck, a spreadsheet of genes on a shared drive. Each copy is a snapshot, cut off from the data, the parameters, and every change made afterward. When a collaborator asks "which comparison is this?" or "can you re-run it without that outlier sample?", the answer depends on the one person who made it.

Sharing in Pluto keeps a result attached to its data, its methods, and its change history. The people you share with can see how a result was made, comment on it in place, and, with the right role, work on it directly, bringing them right to the location of analysis. Pluto is built for this: everyone in your lab and your collaborators can be given seats, with view or edit access granted to specific people.

What this looks like in a lab

Two Kinds of Sharing: Access and Results

Sharing in Pluto comes in two kinds. Sharing access invites a person into the data itself with a role, so they can open an analysis space. Sharing a result sends a specific figure, usually as an export with methods included, without giving access to anything else.

Sharing access: roles and invitations

Permissions apply at the organization and Program level, and a role on a Program carries over to every dataset and analysis space in it.

To give a collaborator ongoing access to a full Program, navigate to the 'Users' section of the Program. Here, you can see which users have access to the Program and what their role is. These roles have important permissions that were covered in Course 1.

Sharing a result: canvases, comments, and exports

• Canvas gathers plots, tables, text, and methods in one place to share with your lab or collaborators.

• Comments and notes can be left on experiments and plots. Use @name to tag a colleague, who gets an email with a link to reply.

• Static exports to PowerPoint or Word (see Exporting Figures at the Right Size and Format) are the right choice for decks and reports; include methods so the context travels with the figure.

Which route fits the audience?

Choosing the right sharing route is itself part of tailoring for an audience, the principle from Module 3.4: a view-only Program invite keeps a stage-gate reviewer focused on the narrative as built, while edit access invites a close collaborator to engage directly with the underlying analysis.

Course 3 complete. You have covered a real-world scientific problem in EGFR-mutant NSCLC drug resistance, choosing and interpreting analyses, multi-omics integration, principles of scientific visualization, plot customization and export and generating reproducible methods in Pluto. Course 4 continues from here.

Course 4

Capstone Project

Courses 1 through 3 built up -omics literacy, critical thinking about data quality, and the skills to choose, run, and communicate an analysis. This capstone course asks you to put all of it together: find a multi-omics dataset tied to a scientific question you care about, set up your own analysis space, perform analyses, and tell the resulting story on a shareable canvas. If you don't already have a multi-omics dataset in mind, we'll point you to a few places to find one.

This course is broken into five modules. Unlike Courses 1 through 3, there is no single dataset walked through step by step here - the purpose of the capstone is applying these skills to a dataset and question of your own choosing. Your progress saves automatically, so you can close the tab anytime and pick up right where you left off.

Modules

• Module 4.1Identify Datasets to Analyze4 topics
• Module 4.2Analysis Setup3 topics
• Module 4.3End-to-End Analysis of Single Experiments2 topics
• Module 4.4Multi-Omics Analysis3 topics
• Module 4.5Storytelling Using Multi-Omics4 topics

Start From a Question, Not a Dataset

In Course 3, you stepped into the role of a scientist studying a real problem, osimertinib resistance in EGFR-mutant NSCLC, using publicly available datasets that were chosen for you. Here, the order reverses: start with a scientific question you want to answer, from your own research, a paper you've read, or general curiosity, and let that question guide which datasets you look for.

Search for Datasets

• Check what your team or organization already has in Pluto first. Internal datasets in your organization’s Datasets tab may already cover your question or a related one.

• Search Pluto's Global Data Commons (Course 1, Module 1.3) for public datasets. Filter by Organism, Dataset type, or search by keyword or accession (e.g., a GEO 'GSE' ID).

• If a relevant GEO dataset isn't already in Pluto's catalog, request it using the '+ GEO import' button next to '+ Dataset', the same workflow covered in Course 1.

• You may choose to upload a new dataset into Pluto with data you or your team has generated.

Don't Have a Dataset in Mind? Some Example Questions and Datasets to Choose From

Here are a couple of datasets to choose from if you don't have your own data or a specific public dataset in mind. Both are real published studies with more than one omics layer and both already are available in Pluto. Pick whichever question is closest to something you'd find genuinely interesting to dig into, then search Pluto's Global Data Commons for each GEO accession below to link it into your analysis space.

Suggestion 1: PARP1–SNAI2 Transcription Axis Drives Resistance to the PARP Inhibitor Talazoparib

PARP inhibitors, like talazoparib, are effective in BRCA-mutated tumors, but resistance inevitably develops. Resistance patterns in BRCA-mutated settings have been well described, but there is comparatively little work in non-BRCA-mutated tumors. Here, the investigators characterized talazoparib resistance in non-BRCA-mutated cancer cell lines to determine the resistance mechanisms at play in these tumor types. Full study: PMC9304387.

• RNA-seq, GSE202434: several cell lines acutely treated with talazoparib, plus some with acquired resistance.

• ATAC-seq, GSE202437: two experiments: PSN1 cells acutely treated with talazoparib, and PARP1 knockdown in PANC1 cells.

Suggestion 2: KRAS-Driven FOSL2 Epigenetic Reprogramming in Pancreatic Cancer

This study sought to uncover epigenetic changes associated with the progression of pancreatic ductal adenocarcinoma (PDAC). This study used ATAC-seq, H3K27ac ChIP-seq, and RNA-seq on genetically engineered mouse models (GEMMs) of PDAC with differing KRAS and/or TP53 mutations. Full study: PMC10403592.

• ATAC-seq, GSE190900

• ChIP-seq (H3K27ac), GSE190901

• RNA-seq, GSE190903

Choose Your Dataset

Before committing, sanity-check your candidate against what you've already learned in this certification:

• Relevance: does it address your question? Right tissue/species/model system, adequate sample size, and clearly defined groups to compare.

• Pipeline QC (Course 2): if QC is available are there any obvious red flags?

• Multi-omics compatibility (Course 3): matching species, genome/annotation build, and experimental conditions across every layer you plan to integrate.

Choose a Program

Programs are the highest level of organization within a Pluto workspace (Course 1); they house both datasets and analysis spaces. Decide whether your capstone belongs inside an existing Program already relevant to your question, or whether it deserves a new, dedicated Program. Creating a Program requires Owner or Editor permissions, the same roles covered in Course 1.

🖼️ Example in Pluto

Programs list view with the '+ Program' option.

Create Your Analysis Space

An analysis space is the scientific canvas within a Program where you'll link datasets, run analyses, and eventually build your final story (Courses 1 and 3).

🖼️ Example in Pluto

Creating a new analysis space, with a name and optional goal or hypothesis.

Get Your Data In

There are two different paths for getting a dataset into your analysis space, and the steps differ depending on whether the dataset already exists in Pluto or not.

Path A: Link a Dataset Already in Pluto

Use this path if your dataset is already public and searchable or if a colleague has already imported it into your workspace.

If a public dataset you want isn't in Pluto's catalog yet, request it first using the '+ GEO import' button next to '+ Dataset' (Course 1), then come back and link it with the steps above once it's been imported.

Path B: Create a New Dataset From Your Own Data

Use this path if the data doesn't exist in Pluto yet, for example, unpublished data generated by your own team.

Check Quality Control First

Before trusting any biological result, revisit the two layers of QC from Course 2: Pipeline QC (the MultiQC-based report covering raw data and alignment quality) and Downstream QC (using PCA to catch outliers, unexpected clustering, or covariates worth accounting for).

• If you uploaded raw data, review the QC report

• Run PCA on each dataset, label by your groups of interest, and check whether samples separate the way you'd expect biologically, or whether a technical variable (batch, cell line, etc.) looks like it's driving variability instead.

Analyze Your Datasets

Course 3, Module 3.2 introduced the typical RNA-seq analysis chain: PCA → save variable genes → clustering heatmap → differential expression → GSEA. Apply that same chain or similar analyses to your datasets. Incorporate other analysis types that may not have explicitly covered in the courses but might be useful based on your questions and interests.

Confirm Compatibility First

Before integrating, revisit the Course 3, Module 3.3 checklist one more time: matching species, matching genome/annotation build, and comparable experimental conditions across both layers. This is the last point where it's easy to catch a mismatch.

Create a Multi-Omics Analysis

Because both of your datasets are linked into the same analysis space, you can create an analysis scoped to the analysis space itself, rather than to a single dataset. That's what exposes analysis types built to draw on more than one linked dataset at once, the same approach used in Course 3, Module 3.3.

• Overlap Analysis: define a significance and fold-change threshold for each layer's differential result, save each as a list, then run Overlap to generate a Venn diagram of genes or features flagged in both.

• Concordance Analysis: plot fold-change from one layer against fold-change from the other, one point per feature, to check whether direction and magnitude agree broadly, not just whether a fixed significance cutoff was cleared in both.

• Peak/Signal Analysis, side by side: for a specific gene of interest, create a signal-track view from each dataset and arrange them next to each other on the canvas for a direct, locus-resolved comparison.

Interpret What You Find

A feature that's significant in both layers has independent support from two orthogonal assays, which is a meaningfully stronger candidate for follow-up than either layer alone. A feature that's significant in only one layer isn't automatically noise, as it can point to real regulatory nuance a single-omics view would have missed and is worth a closer look rather than automatic dismissal (Course 3).

Build Out the Canvas

Course 3's Modules 3.4 and 3.5 covered the principles of scientific visualization and how to customize and export a figure in Pluto; apply those principles here rather than relearning them. Your analysis space is already the multi-omics canvas, the same shared, arrangeable space you've been building in since Module 4.2.

• Add QC plots from each individual experiment so a reader can judge data quality before seeing the biological result.

• Add sample and metadata summaries so a reader understands exactly what was compared.

• Add your key differential plots for each layer (e.g., differential expression and differential binding volcano plots).

• Add your multi-omics comparison plots, the Overlap Venn diagram, Concordance scatter plot, and/or side-by-side Peak/Signal tracks, as the centerpiece connecting the two layers.

• Resize and arrange plots on the canvas so the story reads in a logical order.

Generate Methods

Every Pluto plot has a 'View Methods' popup generated from the live analysis: statistics, parameters, comparisons, and package references. Open View Methods on one of your key plots and copy the text into your canvas.

Share Your Canvas

From the Users section of your Program, invite a team member with Editor or Viewer access, the same sharing model covered in Courses 1 and 3. They'll receive an emailed invite link to accept and log in (or create an account) before gaining access.

• Viewer access suits a reviewer who should see the finished story as built.

• Editor access suits a close collaborator who should be able to work in the space directly.

Complete the Course

Once your canvas tells the story you want it to tell, take a screenshot of your analysis space's Summary tab and upload it to complete Course 4.

1. Take your screenshot

1. In Pluto, open your capstone analysis space and select Summary in the left navigation.
2. Open the Analysis & plot breakdown tab and scroll so the Included on canvas charts are in view.
3. Capture the page, including the breadcrumb at the top that shows your analysis space name.
   Mac: Cmd + Shift + 4  ·  Windows: Win + Shift + S

2. Check your summary

Before you upload, confirm your screenshot shows:

• At least one analysis under Plots by analysis type
• Two dataset types under Linked dataset types (for example, Bulk RNA-seq and ATAC-seq)
• The breadcrumb with your analysis space name

3. Upload

Analysis space link (optional)

Copy the URL from your browser while you are on the analysis space.

Summary screenshot

Choose a file or drag it herePNG or JPG, up to 10 MB. You can also paste a screenshot with Cmd/Ctrl + V.

I confirm that this screenshot shows my own capstone analysis space, that it includes the analyses I built in Modules 4.3 and 4.4, and that I shared my canvas with a collaborator. Submit and complete Course 4