Applied AI-Powered Bioinformatics Tools

RNA-Seq Differential Expression Dashboard

Last reviewed

Beginner

What you'll learn

~45 min
  • Ask an AI tool, in plan mode, for a page that reads a finished differential expression results table (one row per gene, written by an established analysis package), the normalised counts and the sample sheet, and computes no statistics of its own
  • Read its plan for what matters: thresholds you type in, the shrunken fold change used where you said, the three kinds of blank adjusted p-value kept apart, the direction of the comparison stated, and every plot saying which genes it leaves out
  • Check the page against values worked out for the practice files
  • Say why the counts behind a large fold change are worth a look before anyone builds on it, why a blank adjusted p-value is not 'not significant', and why a gene that passes is the analysis's statement about these samples, not a finding

The story, and what you will end up with

It is Monday 16 November 2026 at a genomics core. You are a staff scientist who supports the labs whose RNA sequencing goes through the core. LAB-17 sent six RNA samples, three in group A and three in group B. The core made the libraries and sequenced them, and the core’s analyst ran the differential expression analysis (explained in the box below) with an established count-based package. Tomorrow you sit down with LAB-17 to look at the results together.

In this practice core, the analyst checks every sample before testing anything: read depth and the share of reads assigned to genes, a plot of how the six samples group, a check that no two samples were swapped, and whether the groups line up with anything else, such as the day the RNA was extracted. For LAB-17 the analyst did those checks before today and signed off on them. The analysis record states the design (one factor, group, with three samples in each), that the comparison is B against A, the package’s version, that the shrunken fold changes were made by what the record calls apeGLM-style shrinkage, and that the analysis was set to an adjusted p-value cutoff of 0.05. You and the analyst wrote down the thresholds for tomorrow’s first look before anyone saw a gene: an adjusted p-value below 0.05 and a shrunken log2 fold change of at least 1 either way (the box below explains these words). They are this practice core’s own thresholds for this meeting; yours come from your own analysis plan. You also agreed that any passing gene whose biggest sample carries more than half of its group’s counts goes to the analyst to look at before the meeting. The analyst’s report gives 445 genes with an adjusted p-value below 0.05, at any fold change.

What you do not have is a way to look at the results with the lab. The results table has 12,000 rows, and the lab will want to see the counts for its own genes, sample by sample.

By the end of this lesson you will have a results page on your own computer: a page in your browser that reads the analysis’s three files, shows which genes pass the thresholds you type in, a volcano plot and an MA plot, and, for any gene you pick, its counts in each sample. It computes no statistics at all: no test, no p-value, no adjustment, no normalisation, no shrinkage. Those were done by the analysis package; the page reads what it wrote, counts and draws. You build it by asking your AI tool in plain English, and you check it against numbers this lesson gives you.

🔍Words used in this lessonclick to expand
  • Differential expression analysis: the analysis that compares, gene by gene, the read counts of groups of samples, done here by an established package. DESeq2 is one such package; the practice files are shaped like its results.
  • Normalised counts: each sample’s read counts for a gene after the package’s normalisation, as the package writes them out.
  • baseMean: in the package’s words, the “mean of normalized counts for all samples” (DESeq2 vignette, version 1.52.0, read 11 November 2026).
  • log2 fold change: how much higher a gene is in one group than the other, on a log 2 scale. 1 means twice as high, 2 four times, -1 half. The vignette puts it this way: a log2 fold change of -1 “means that the treatment induces a multiplicative change in observed gene expression level of 2-1 = 0.5”.
  • Shrunken fold change: a second estimate of the same fold change, written by the package. The vignette says: “Shrinkage of effect size (LFC estimates) is useful for visualization and ranking of genes.” This practice core takes its threshold on the shrunken fold change, a choice from its own analysis plan; its analysis record names the method that made the shrunken fold changes, apeGLM-style shrinkage, and yours names your own.
  • Adjusted p-value: the package’s p-value adjusted for testing thousands of genes at once; the vignette’s column description is “BH adjusted p-values” (BH is the name of the adjustment method).
  • Volcano plot: a plot with each gene as a point. On the page in this lesson, the shrunken fold change goes across, and how small the p-value is (adjusted or raw; your page says which) goes up.
  • MA plot: a plot with each gene as a point. On the page in this lesson, the mean (baseMean) goes across and the shrunken fold change goes up.

Before you start

First, if your AI tool is still open, leave it: type /exit and press Enter (if you use Antigravity, press Ctrl+D with nothing typed instead), and wait until you see the $ or % prompt. The next line goes there, in the terminal, not in the AI tool. It makes a folder for this lesson beside your practice folder and goes into it:

Terminal window
mkdir -p ~/bio-workbench/05-de-results && cd ~/bio-workbench/05-de-results

Start your AI tool with the start line you kept from the setup lesson, the one with approval prompts off. Then ask it to set the folder up:

Set this folder up so we can undo any change you make, download https://learn.intellimetrics.net/starters/biotech/05/de-results.zip and unpack it into a data folder, and save a checkpoint.

Look for it saying three things: the folder is set up with Git (the program that saves snapshots, from setup), three files are in the data folder (results.csv, normalized-counts.csv and samples.csv), and a first snapshot, your checkpoint, is saved. Later you can ask whether the files have changed since then. If it does not mention one of them, ask: “Did you save a checkpoint with the files in it?” It may also keep the zip file in the data folder, which is fine; if it says it kept one somewhere outside this folder, ask it to delete that copy. If the download does not work, download de-results.zip in your browser, check where it went (the browser’s list of downloads shows the folder), and tell your AI tool, for example: “de-results.zip is in my Downloads folder. Unpack that one file into a data folder here, and save a checkpoint.” In Windows PowerShell or a desktop app, “my Downloads folder” works as is; only say “my Windows Downloads folder” if you use Ubuntu (WSL) on Windows, which keeps its own Downloads folder apart from Windows.

The box below says what is in the practice files; later steps refer to them.

ℹWhat is in de-results.zipclick to expand

Three files, each of which opens in a spreadsheet program:

  • results.csv: one row per gene, 12,000 rows. The first column, which has no name, is the gene. Then baseMean, log2FoldChange (the unshrunken estimate), lfcSE (its standard error), stat (the test statistic the p-value comes from), pvalue, padj (the adjusted p-value) and log2FoldChange_shrunken. NA means the package wrote no value.
  • normalized-counts.csv: one row per gene, one column per sample, S1 to S6.
  • samples.csv: which group each sample is in. S1, S2 and S3 are group A; S4, S5 and S6 are group B.

Some rows have no adjusted p-value, for one of three reasons the package documents. In the vignette’s words: “If within a row, all samples have zero counts, the baseMean column will be zero, and the log2 fold change estimates, p value and adjusted p value will all be set to NA.” “If a row contains a sample with an extreme count outlier then the p value and adjusted p value will be set to NA.” “If a row is filtered by automatic independent filtering, for having a low mean normalized count, then only the adjusted p value will be set to NA.” In these files:

Kind of rowHow to tellGenes
No countsbaseMean 0, everything else NA1,845
Count outliera fold change, but pvalue and padj NA3 (GENE-04203, GENE-07836, GENE-09531)
Filtered for a low meana pvalue, but padj NA; every one has a baseMean under 3.621,365
Has an adjusted p-valuea padj8,787

Genes you will meet further down:

GenebaseMeanShrunken log2 fold changepadjCounts S1 to S6
GENE-00001338.42.081.9e-19139.4, 118.0, 120.9, 450.3, 549.1, 652.8
GENE-0024029,442.8-7.15055,077.4, 62,049.0, 58,301.1, 377.6, 423.6, 428.3
GENE-095317,204.00.16NA (outlier)2,709.1, 2,510.1, 2,602.0, 2,517.2, 2,612.6, 30,273.4

GENE-00240’s p-value and adjusted p-value are written as 0: they are too small to store as any other number, not a true zero.

The counts are made up for this course (no organism or instrument produced them), and the normalised counts and the results table, standing in for the analyst’s, were worked out from them with PyDESeq2 0.5.4, a Python implementation of the package’s method, using the settings the scene’s analysis record gives, so each row is what that implementation gives for those counts; your own core’s package and version will differ in details. Each lesson in this track has its own practice files, with their own people and dates.

Switch to plan mode

Your AI tool is already open. Switch to plan mode now: Shift+Tab until it says plan, or /plan (the keys for each tool). In plan mode it usually reads your files, asks its questions and shows you a plan before it builds anything; the order can vary.

The conversation

Each message below is an example. Say it your own way.

Say what you want

Say what you do, which files to use, and what you want to see:

For our core's labs, build an offline page for the results in data, computing no statistics and leaving the input files unchanged: thresholds I type, volcano and MA plots, each gene's counts by sample. Ask me what you need first.

Your AI tool usually reads the files and comes back with questions. It may already tell you how many genes have no adjusted p-value (it may put them all down to filtering; the box above has the three kinds), and it may offer plots of the samples, such as a PCA (a plot of how the samples group); the analyst has already looked at the samples, and those are statistics the page should not work out again.

Answer its questions

Answer each in a sentence. These are questions the tools often ask, with one way to answer each. Yours may ask them in a different order or show a list of choices:

If it asksYou might say
Who opens the page: you, or the labs?It’s for me and the lab together, on my computer.
Can it load anything from the internet?No. It must work with no internet connection.
Hide the genes with no adjusted p-value, or show them greyed out?Either, if each plot says which it leaves out; count them by kind, never as not passing.
Is the comparison B against A, so positive means higher in B?Yes, B against A: positive means higher in B.
Which thresholds to start with?Start at adjusted p below 0.05 and a shrunken log2 fold change of at least 1 either way.
The adjusted p-value or the raw p-value for the threshold?The adjusted p-value only.
A box for a minimum expression?No minimum-expression box.
The shrunken fold change, the unshrunken one, or a switch?Use the shrunken fold change for the threshold and both plots; show both in the table.
Which plots? A PCA, a heatmap, a p-value histogram?Volcano and MA are enough.
Draw the gene with an adjusted p-value of 0 at the top edge?Draw it at the top edge, marked.
A file of gene names for these IDs?Gene IDs as they are.
Real group names in place of A and B?Keep A and B as the group names.
Group the counts by condition?Yes, grouped by condition.
Counts as a table, a plot, or both?Table and plot, both on a linear scale.
Counts as normalised counts? On a linear or a log scale?Show the normalised counts as they are, on a linear scale.
A searchable gene table?A searchable gene table, yes.
Downloads of the gene list or the plots?No download for now.
Keep the raw p-value column in the table?Keep it, labelled as the raw p-value.
Only this analysis, or others later?Others later: a button to open another analysis’s files of the same kind.
One file you double-click, or a local server?Whatever needs the least installed, working with no internet connection.

If it asks several questions at once, answer the ones about the thresholds, which fold change and the comparison from your analysis record, a sentence each, and accept its suggestions only about how the page looks. If you cannot tell which kind a question is, treat it as a rule. If it asks something not listed here, answer it the same way, in a sentence.

If your AI tool does not ask which way the comparison runs, tell it in a message of its own, then wait for its reply:

The comparison is B against A, so a positive log2 fold change means higher in B.

The direction comes from the analysis record, not from memory: the files themselves do not say which group the package compared against which.

If it does not ask which thresholds or which fold change to use, tell it the same way, from your analysis record, then wait for its reply:

Start at adjusted p below 0.05 and a shrunken log2 fold change of at least 1 either way. Use the shrunken one for both plots too, and show both in the table.

Wait for its reply, then, in a message of its own, ask for the thresholds in use to be on the page all the time, so that anyone looking at it, or at a printed copy, can see what the list was filtered by. If your AI tool already shows a plan or a choice to approve one, do not approve it until you have sent this message and the two after it:

Always show the thresholds in use, and mark any value typed in that differs from the ones we wrote down: adjusted p below 0.05, shrunken log2 fold change at least 1.

Wait for its reply, then, in another message of its own, say how the genes with no adjusted p-value are to be counted. The file writes every blank the same way; the three kinds differ in the other columns, as the box above shows:

Count genes with no adjusted p-value by kind: no counts in any sample; a fold change but no p-value (count outliers); a p-value but no adjusted one (filtered for a low mean). Never as not passing, never using raw p-values.

Wait for its reply, then ask for the counts of passing genes, and for what each plot leaves out:

Show how many genes pass, higher in B and higher in A, changing as I type, and have each plot say which genes it leaves out, and why.

Wait for its reply, then read the plan.

Read the plan

Next it usually shows you a plan: what it will build and how it will check it. Read it for six things:

  1. Nothing is worked out again: no test, p-value, adjustment, normalisation or shrinkage. The page reads the files.
  2. The thresholds are yours, always shown, and change when you type: the counts of genes higher in B and higher in A follow them, and any value that differs from the written-down ones is marked.
  3. The shrunken fold change is used for the threshold and both plots, and the table shows both fold changes.
  4. The three kinds of blank adjusted p-value are counted apart (no counts in any sample; a fold change but no p-value; a p-value but no adjusted one), never as not passing, and the raw p-value never stands in for a blank one in the counts.
  5. The page says which way the comparison runs: positive means higher in B.
  6. Each plot says which genes it leaves out, and why, and the page works with no internet connection and leaves your files as they are.

A first plan often leaves one of these out. Say so in a sentence, in your own words, for example:

The plan doesn't say which way the comparison runs. Show on the page that a positive fold change means higher in B.

When the plan has all six, approve it: if your AI tool offers a choice to go ahead, pick it; otherwise say “Yes, go ahead and build it.” In Codex, press Shift+Tab first to leave plan mode. In Copilot, saying it brings up a choice to go ahead; pick that. Let it build. It may test what it built and tell you how to open it. When it has finished, ask it: “Did you install anything, and where?” Nothing, or something inside this lesson’s folder, is fine for these lessons. If it names anywhere else on your computer, such as the zip file it downloaded, ask it what that is for, and to remove it if you do not want it there.

Look at what it built

Ask it to open the page:

Open the results page in my browser.

Your AI tool may open the page for you. A page opened straight from its file may carry its own copy of the practice files, so it needs nothing else running. If it gives you a file location or an address without opening it, copy that into your browser’s address bar and press Enter, or ask your AI tool to open it for you. You can also press Ctrl+O in your browser (Cmd+O on a Mac) and choose the page’s file in this lesson’s folder. On Windows, the folder is inside Ubuntu, so ask your AI tool: “How do I open this page in my Windows browser?” If it starts the page at a local address, one ending in a number, and says the address is already in use, say: “Something else is using that address. Don’t stop it. Use a different address for my page.” If your AI tool does not take a message after opening the page, wait a minute. If it still does not, close the terminal window, open a new one, type this lesson’s folder line again, start your tool, and say: “Start the page again so it keeps running while we talk.” The page and your saved work are files in the folder, so nothing is lost. When you have finished with the page, ask: “Stop the results page if it is still running.” Your page will be laid out in its own way.

Some checks below ask you to type in the page or pick a gene. Pick a gene the way your page offers, usually by typing its name in a search box and clicking its row, or by clicking its point on a plot. If you cannot find where, ask your AI tool where it is.

Check it for these:

  • The page shows 12,000 genes, and says that S1, S2 and S3 are group A and S4, S5 and S6 group B, or shows them that way beside the counts.
  • At adjusted p below 0.05 and a shrunken log2 fold change of at least 1 either way, 336 genes pass: 171 higher in B and 165 higher in A.
  • The 3,213 genes with no adjusted p-value are counted apart, in three kinds: 1,845 with no counts in any sample, 3 count outliers, with a fold change but no p-value, and 1,365 with a p-value, filtered for a low mean. None is counted as not passing. If your page shows a total of genes not passing, it is 8,451, the genes with an adjusted p-value that do not pass; a larger total means blanks have gone in with it; tell your AI tool.
  • The page says that a positive fold change means higher in B.
  • From the start, the page shows the thresholds in use, the ones written down before the meeting: adjusted p below 0.05 and a shrunken log2 fold change of at least 1. Nothing is marked.
  • Type 2 in place of 1 for the fold change: 121 genes pass, 59 higher in B and 62 higher in A, and the page marks the fold change as different from the written-down one. Typing a number reruns nothing; the page only filters what the analysis wrote. Then put it back to 1: the mark goes.
  • Type 0 for the fold change: 445 genes pass, the count in the analyst’s report for adjusted p below 0.05. Then put it back to 1.
  • Pick GENE-00001. Its counts are 139.4, 118.0 and 120.9 in S1 to S3 and 450.3, 549.1 and 652.8 in S4 to S6, and its shrunken fold change is 2.08: positive, and all three B samples are higher. That agrees with the direction from the analysis record.
  • Pick GENE-09531. It is in the count-outlier kind (your page may call it something plainer, such as both p-values unavailable), and S6 shows 30,273.4 against 2,510.1 to 2,709.1 in the other five samples.
  • Each plot says which genes it leaves out, and why. The MA plot leaves out the 1,845 genes with no counts, which have a mean of 0 and no fold change. What the volcano plot leaves out depends on the p-value it uses for height: with the adjusted p-value, all 3,213 genes with none; with the raw p-value, the 1,848 with none (the 1,365 filtered genes are then drawn, usually in grey). Either way, GENE-00240, whose p-value is written as 0, is drawn at the top edge and marked, or named among the genes left out.
  • When you ask “Does the page load anything from the internet?”, it says no. To see for yourself, turn off your computer’s Wi-Fi and reload the page: it still works.
  • When you ask “Have my results files changed since the checkpoint?”, it compares them with the checkpoint and says no.

If something on the list is not true, tell your AI tool in a sentence what you see and what you expected, for example: “The page puts all 3,213 genes with no adjusted p-value in one group. Count them by kind: no counts, a fold change but no p-value, or a p-value but no adjusted one.”

Notice what the list does not say

The page now says 336 genes pass. Before you take that list to the lab, look at two things it cannot show on its own.

First, the biggest changes in the file. Ask your AI tool:

Sort every gene by the unshrunken fold change, largest either way. Show me the top ten with their mean, shrunken fold change and adjusted p-value.

Before you read on: what would you expect the genes with the largest fold changes in the whole file to look like?

Six of the ten are on the list. The ten largest unshrunken fold changes run from 4.92 to 7.37 either way, written as 30 to 165 times higher in one group. Six pass the thresholds: GENE-00248, GENE-00240, GENE-06479, GENE-06215, GENE-09178 and GENE-10121. Three have no adjusted p-value because they were filtered for a low mean (GENE-05876, GENE-10567 and GENE-02914, on means of 2.8 to 3.4 normalised counts), and their shrunken fold changes are within 0.25 of zero. The remaining one, GENE-00368, has an adjusted p-value of 0.055, just above the threshold, and a shrunken fold change of -3.80.

Five of the ten have no counts at all in one group, and the size of the fold change does not tell them apart; the counts do. In normalised counts, GENE-00248 has 39.4, 26.6 and 26.5 in group B and none in A, and GENE-06479 has 31.7, 20.4 and 15.0 in group A and none in B: both have an adjusted p-value (0.006 and 0.014), both pass, and their shrunken fold changes are as large as the unshrunken ones or larger (7.87 against 7.37, and -7.40 against -6.98). The other three have a few normalised counts in one group and none in the other: GENE-05876 has 1.1, 9.3 and 10.0 against none, and its shrunken fold change is -0.24 against an unshrunken -5.28; in GENE-02914, one sample, S6, carries 16.1 of the 17.0. Whether a change with no counts in one group is established, and what is said about it, is the analyst’s to judge, whichever side of the thresholds it falls. So is a shrunken fold change larger than the unshrunken one, as in GENE-00248 and GENE-06479: ask the analyst about it before anyone quotes it. And a blank adjusted p-value is not “not significant” either: those three were filtered for a low mean, so the analysis has no adjusted p-value to judge them by.

Second, the counts behind a gene that passes. Ask:

For each gene that passes, show what share of the higher group's counts its biggest sample carries, largest first, with its counts. Don't change anything.

With three samples in a group, a share of about 33% means all three are equal, and 100% means one sample has everything. Your AI tool may show only the top of a long list. GENE-10047 should come first, at 60%, then GENE-11684 and GENE-10454, both at 59%; only 1 of the 336 is above 60%. A high share alone does not mean a gene rests on one sample. GENE-10047’s biggest sample, S5, has 30.9, but its other two B samples, 7.9 and 12.3, are each more than six times the mean of group A’s 0, 2.8 and 1.0. Now find GENE-10454, third. It passes, with an adjusted p-value of 0.016 and a shrunken fold change of 1.10. Its counts are 43.1, 74.3 and 57.0 in group A and 72.7, 108.5 and 262.5 in group B. Without S6, group B is about one and a half times group A; with it, about two and a half times. Most of the difference is S6. The package did not mark it as an outlier, so nothing in the results table says so; whether one sample should carry that much is a question for the analyst, who has the analysis’s own diagnostics. S6 is also the high sample in GENE-11684 and in the outlier GENE-09531, so look across genes too: whether one sample comes up too often is also the analyst’s question. On the practice files none stands out: of the 16 passing genes whose biggest sample carries more than half its group’s counts, each of S1 to S6 is the biggest in 2 to 4. Under the line you and the analyst agreed, more than half, those 16 genes go to the analyst to look at before the meeting; where your own line sits is agreed with your analyst, and no rule on this page sets it. The share puts every passing gene in order, so none is missed; the counts are what you read.

What GENE-10454 means is not for the page or for you alone to say. What explains one sample standing apart is for the analyst and the lab to find out: the analyst has the sample-level checks, and the lab knows the samples. They decide what it means and what they say about it.

A gene on this page’s list passed thresholds you chose, on numbers the analysis worked out for these samples under its design; whether it changed in a way that matters, and what goes back to the lab, is for the analyst and the lab to judge from the whole analysis, not the page.

One real run

Here is part of what one AI tool said after the first question above, with Claude Code on 12 November 2026. Yours will word it differently; check your page against the checks above, not against this.

- **Ranks 6, 8 and 10 are probably not real changes.** Each has a mean of only about 3 and no adjusted p-value. Their shrunken fold changes are near 0, so the large unshrunken values are most likely low-count noise. They count under "no adjusted p-value", not as up or down.

“Probably not real changes” and “most likely low-count noise” say more than the files can: these three were filtered for a low mean and have no adjusted p-value, so the analysis shows neither that they changed nor that they did not.

The genes it talks about are made up for this course.

How real analyses differ

  • Shrunken or not, it is an estimate. The vignette says of the MA plot: “It is more useful to visualize the MA-plot for the shrunken log2 fold changes, which remove the noise associated with log2 fold changes from low count genes without requiring arbitrary filtering thresholds.” Check which fold change your analysis wrote and which one your page uses; a results table may hold only one. This page reads both from one table; if the shrunken estimates came out as a table of their own, ask whoever ran the analysis for one table with both. Put the name of the method that made your shrunken fold changes, from the analysis record, on the page beside them, as the practice analysis record names its own.
  • A fold-change cut on the page is not a test of fold change. The page keeps genes whose adjusted p-value is under your threshold and whose fold change is at least yours. That is not the same as a test against a fold change, which the package can run itself; the vignette, where Wald tests are the package’s tests of significance: “It is also possible to provide thresholds for constructing Wald tests of significance.” If the lab wants genes that changed by more than some amount, that is a question for the analyst to answer in the analysis.
  • Typing a new adjusted p-value is not running the analysis again. The practice analysis record gives its own setting for that cutoff, 0.05. The vignette, where alpha is the package’s setting for that cutoff, says: “By default the argument alpha is set to 0.1. If the adjusted p value cutoff will be a value other than 0.1, alpha should be set to that value”. A count the page shows at another adjusted p-value is a look at this table, not what the analysis would report at that value.
  • A small p-value is not a big change. The vignette: “a gene can have a small p-value although the change in expression is not great, as long as the standard error associated with the estimated LFC is small.” The fold change and the counts say how big; the p-value does not.
  • Which way the comparison runs depends on how the analysis was set up. For one factor, when the package’s results function is given “no additional arguments”, the package compares “the last level of this variable over the reference level”, and the vignette notes: “By default, R will choose a reference level for factors based on alphabetical order.” Take the direction from the analysis record and check it on one gene’s counts.
  • A real table may hold more. Check what yours holds, for example: gene identifiers with version numbers and a column of gene names, some names on more than one row; far more rows; more than one comparison, in separate files; a design with more than one factor, such as batch. Gene names make a page easier to read; check that a name search finds every row with that name.
  • The page shows normalised counts. The raw counts and the size factors the analysis used are with whoever ran it; for a gene with only a few counts, ask for them before saying anything about it.
  • A gene that did not pass is not shown to be unchanged. Not passing says only that its numbers in this analysis did not meet the thresholds.

Keep going

Small changes do not need plan mode. Say each in a sentence, then look at the page again:

  1. “Let me pick several genes at once and show their counts side by side.” Try it with GENE-00001 and GENE-10454.
  2. “Save a checkpoint.” Your AI tool saves a snapshot you can return to.

When you want to look at a real analysis, use the finished page. It runs in your browser by itself, with your AI tool closed, so your AI tool never sees the files. Nothing confidential or unpublished, and nothing about people or about animals under a protocol, goes to your AI tool or into this course folder; nor does data covered by an agreement. Your analysis may not be shaped exactly like the practice files: its columns may have other names, and its gene identifiers will be real. The files come from whoever ran the analysis, written out from the package: a results table with both fold changes in it, the normalised counts and the sample sheet. Do these steps in order; the first five happen now, while your AI tool is open and this folder still holds only the practice files:

  1. Go through the data check, the questions in setup about what the data is, who decides how it is used, and which AI tool would see it. A lab’s results are that lab’s to decide on, and data from human samples may need your facility director or compliance office as well. Go on only if they allow this use.

  2. Decide on a folder of its own for your real files, outside this course folder, for example ~/rnaseq-results. Nothing has to be in it yet. If the folder is not there, ask your AI tool to make it: “Make an empty folder called rnaseq-results in my home folder.”

  3. Ask your AI tool for blank files and one made-up gene to try: “In the rnaseq-results folder in my home folder, put only blank results, counts and sample sheet files my spreadsheet program can open, with one made-up gene. Keep the page and its checkpoints here. Show the page opening them.” Look for it opening them in the page and showing that one gene. Then look at your own analysis’s files in a spreadsheet program and tell your AI tool the names of their columns, with two rows you make up in the same shape, made-up gene names and numbers, never a real row. For example: “My results file has the columns gene_id, baseMean, log2FC, lfcSE, pvalue, padj and log2FC_shrunk. Here are two made-up rows in the same shape. Change the page to read files like that, and update the made-up files in rnaseq-results to match.” Then paste the two rows you made up. Wait for its reply, then do the same for your counts file and your sample sheet if their columns differ from the practice ones, each with two made-up rows, asking it each time to update the made-up files to match. Share only the layout: where a column name is a sample’s name, as in a counts file, make one up in its place, and send no real gene list.

  4. Give the page your own settings and tidy the real-files folder, one message at a time:

    1. Give the page your own starting thresholds in place of this practice core’s, from your analysis plan. For example: “Replace the page’s starting thresholds with ours, which I’ll send next. Then make up one gene just inside and one just outside each, and show me how the page sorts them.” Wait for its reply, then send your thresholds. Check that each made-up gene passes or not the way its numbers say against the threshold it was made for, and that the page now marks a value only when it differs from yours.
    2. Wait for its reply, then tell it, from your analysis record, the name of the method that made your shrunken fold changes, and which group is compared with which unless the page already asks for that. For example: “Our comparison is treated against untreated, and our shrunken fold changes were made with apeglm. Show both on the page.”
    3. Wait for its reply, then ask: “What did you put in the rnaseq-results folder besides the made-up data files? Don’t remove anything yet.”
    4. If it lists a copy of the page there, say: “Make every change in that copy in the page in this lesson’s folder, and show me that page opening the made-up files.” Look for the made-up gene.
    5. Wait for its reply, then say: “Now remove everything in the rnaseq-results folder except the made-up data files, including any copy of the page and any checkpoint folder.” Do this now, while only made-up data is there: real files should not sit beside a copy of the page with practice data built in, or in a folder that keeps checkpoints, which can save copies of them.
    6. Wait for its reply, then say “Save a checkpoint.” Step 5 asks what has changed since this checkpoint.
  5. Ask your AI tool how to open the page yourself, for example: “Tell me exactly how to open this page myself, with nothing of yours still running, with its full path, and which button opens files from the rnaseq-results folder in my home folder.” Use your own folder’s name. If it says the page needs something of its own running, ask it to make the page open straight from its file, and try that before you leave your AI tool. If it offers to copy your real files into this folder instead, say no. Copy what it tells you into a note. It will look something like this, though yours may differ: open your browser, press Ctrl+O (Cmd+O on a Mac) and choose /home/dana/bio-workbench/05-de-results/index.html; then click Open Analysis, go to /home/dana/rnaseq-results, and pick the three files together. The first path is the page in this lesson’s folder, the second your real-files folder. /home/dana is the home folder of someone whose user name is dana; your own user name goes where dana is, and on a Mac the paths begin /Users/ instead of /home/. What it tells you depends on the computer: the example came from a Linux computer, so use what your own AI tool gives you. Your page’s button may have another name, and it may ask which group is compared with which; answer from the analysis record. The page opens from where it is, so leave this lesson’s folder in place. If you are unsure about any part, ask your AI tool now, while it is still open.

    Then, before you leave your AI tool, follow your note yourself with the made-up files, because a real analysis’s numbers must never end up in the course folder: open the page the way the note says and choose the made-up files in rnaseq-results with the page’s button. Look for the made-up gene. Then ask: “Has anything in this folder changed since the last checkpoint?” Look for it saying no. If something changed, tell your AI tool what it says and ask it to make the page save nothing.

  6. Ask your AI tool to stop anything it started for the page, then leave it the way you did at the start of this lesson.

  7. Put a real analysis’s three files in your rnaseq-results folder, in place of the made-up ones. Open the page the way your note says; if it will not open, start your AI tool in an empty folder and ask it how to open a page file on this computer, giving it no data. Choose the files with the page’s button; the page reads them on your computer and saves nothing. Check the page against the analyst’s report before you sit down with the lab: set the fold change to 0 and the adjusted p-value to the analysis’s cutoff, and the number that pass should match the report’s count for that cutoff (on the practice files, 0 and 0.05 give 445). If it does not match, stop there and take it to the analyst before the meeting; do not bring the real files to your AI tool. Then put your own thresholds back. To come back to the lesson folder later, type this lesson’s folder line again, the line near the top of this page that begins with mkdir.

Try another case

Before the quiz, write two or three sentences for yourself. In the meeting, the lab points at a gene with a shrunken log2 fold change of 3.2 and an adjusted p-value of 0.003, and asks you to put it in their grant report as strongly induced in B. What do you look at first, and what do you say?

A good answer opens the gene’s counts before anything else: whether all three samples in B are higher, or one sample carries it; and how many counts it has at all. It says what the numbers are (the analysis’s estimate for these six samples, passing thresholds set before the meeting) and leaves the claim in the report to the lab and the analyst, who can check the gene another way if it matters.

Key takeaways

  • A short ask, then your analysis record. Say what you do and which files; take the thresholds and the direction from the record.
  • Read the plan for what the page must not do: compute anything, lump the three kinds of blank together, or hide what a plot leaves out.
  • Check a new page against numbers you know before you trust it on an analysis you do not.
  • Open the counts. A gene on the list is a place to look; the counts show whether three samples or one carry it.
  • Practice data stays practice data: real results go to the finished page, with your AI tool closed, after the data check.
KNOWLEDGE CHECK

On your page, a gene has an unshrunken log2 fold change of 6.2, a mean of 1.8 normalised counts, a p-value of 0.03 and no adjusted p-value. Its shrunken fold change is 0.02. What can you tell the lab about it?

KNOWLEDGE CHECK

A gene passes your thresholds. Its normalised counts are 30, 41 and 26 in group A and 44, 52 and 150 in group B. The lab wants to report it as induced in B. What happens next?

What’s next

The next lesson is about checking the record a finished RNA sequencing run left behind against the lab’s sample sheet and its settings, and what can honestly be written about what was run.

intellimetrics · Arlington, VA
© 2026 intellimetrics.net
Search lessons