Wildtype One

21 Mar

If your experiments worked but your figures still feel fragile, it’s often not the biology—it’s the analysis. This webinar walks through the 5 most common statistics mistakes biologists make + how to avoid them No coding. No equations. Just clarity. 👉 Register for the webinar here: wildtypeone.com

Your passage number might be affecting your data more than you think❗ Liu et al. (2025) looked at RNA-seq data across passages P3–P39. (a thread)

Biology rewards traceability. — Wildtype One 🧬

🧫 Join 1,000+ researchers getting weekly notes on how to diagnose ugly data, avoid weak readouts, and make better bench decisions (it’s free) 👉 wildtypeone.substack.com/about

Your passage number might be affecting your data more than you think❗ Liu et al. (2025) looked at RNA-seq data across passages P3–P39. (a thread)

So, what can you do about that? 🤔 👉 For EVERY experiment you seed, write down: - Thaw date - Passage number - Split ratio - Days/weeks in culture - Cells per well/flask - Confluence at seeding - Confluence at treatment

Density is not just “how many cells.” It changes physiology. And physiology changes your readout. We often blame weird data on biological variability. Yet sometimes, you can save yourself ugly data and repetitions, ...simply by doing the boring work.

Your passage number might be affecting your data more than you think❗ Liu et al. (2025) looked at RNA-seq data across passages P3–P39. (a thread)

They found nonlinear transcriptomic shifts in tumor cell lines. The scarier finding? It wasn’t just RNA. Mid-passages showed altered: - Cell cycle - Metabolism - Stress response - Signaling - Immune-response pathways

Before you repeat the experiment, answer this: What kind of problem do you actually have? Ugly data does NOT have one next step. (a thread)

Before adding more replicates, write down: 1. What changed? 2. What did I measure? 3. What conclusion am I trying to draw? If those three don’t match… more data won’t save you. — Wildtype One 🧬

🧫 Join 1,000+ researchers getting weekly notes on how to diagnose ugly data, avoid weak readouts, and make better bench decisions: wildtypeone.substack.com/about

Before you repeat the experiment, answer this: What kind of problem do you actually have? Ugly data does NOT have one next step. (a thread)

Readout too narrow? - qPCR changed but protein didn’t. - Western blot shows one band but not mechanism. - Flow shows a population shift but not what drives it. Action: Don’t repeat ❌ Change readout ✅

In short 💡: Repeating the same experiment is only rigorous when the problem was execution ✅ Repeating the wrong experiment is just expensive confidence ❌

Before you repeat the experiment, answer this: What kind of problem do you actually have? Ugly data does NOT have one next step. (a thread)

Analysis choice wrong? Wrong test, SEM hiding variation, bad normalization, post-hoc outlier removal? Action: Reanalyze ✅

Design problem? Pseudo-replication, weak controls, unstable reference gene, no predefined exclusion rule? Action: Redesign ✅

Before you repeat the experiment, answer this: What kind of problem do you actually have? Ugly data does NOT have one next step. (a thread)

- Sometimes you should repeat. - Sometimes you should reanalyze. - Sometimes you should redesign. - Sometimes you should change the readout. The mistake is treating every ugly result like a repeat problem.

Ask 🤔: Execution failed? Contamination, bad staining, failed transfer, wrong dilution, instrument issue? Action: Repeat ✅

8 July

“Everyone does it this way” is not a valid statistical defense. (1/3)

8 July

Many analysis habits are: - Lab-inherited - Incorrect - And they show up immediately in peer review. (2/3)

8 July

This webinar helps you spot the most common statistical traps in biology labs and gives you a simple framework to avoid them. Designed for people who actually run experiments—not statisticians. 👉 Register for the webinar here: wildtypeone.com (3/3)

“Do more replicates until it’s significant” is killing your credibility 🧫 (a thread)

But never do it to rescue a conclusion. That’s not rigor. That’s extra time. — Wildtype One 🧬

🧫 Join 1,000+ researchers getting weekly notes on how to diagnose ugly data, avoid weak readouts, and make better bench decisions (it’s free) 👉 wildtypeone.substack.com/about

“Do more replicates until it’s significant” is killing your credibility 🧫 (a thread)

That is what adding replicates until p < 0.05 looks like. Replicates should be planned before you know the result. Not added because the p-value hurt your feelings.

You can add more independent biological replicates when: 1. The design/power calculation called for it. 2. A predefined stopping rule allows it. 3. A replicate had a documented technical failure and needs replacing. 4. You are running a new validation experiment.

“Do more replicates until it’s significant” is killing your credibility 🧫 (a thread)

Imagine a football game ⚽ It’s minute 90+5’. The game is over. Your team lost 2–1. So you extend by 5 minutes… Then 5 more… Then 5 more… You keep doing that until your team wins 3–2. Then you stop the match.

Clear cell death. p = 0.056. Now what?

Stop asking: “Why is this not significant?” Ask: “What decision is this readout actually allowed to support?” “Progress in science depends on new techniques, new discoveries and new ideas, probably in that order.” — Sydney Brenner, Nobel Prize Laureate (2002) — Wildtype One 🧬

PS - If you struggle with “non-significant” statistics and data outliers, Wildtype One's webinar helps you avoid mistakes and get confident statistics. No math or bioinformatics required. Register here 👉 wildtypeone.com

Clear cell death. p = 0.056. Now what?

Maybe this model dies through a caspase-independent route. Maybe your endpoint is too late. Maybe the cells detached before you measured them. Maybe your assay is measuring survival, not death mechanism.

The Better Bench Decision 💡: - Reanalyze if the test or assumptions are wrong. - Repeat if the technical record is weak. - Add biological replicates if the design was underpowered. - Change readout if the current assay cannot answer the biological question.

Clear cell death. p = 0.056. Now what?

🔴 Case B: The effect is real but underpowered. Then more biological replicates may be right. But decide that before seeing the p-value. Not after the p-value hurts your feelings.

🔴 Case C: The readout is wrong. If microscopy suggests cell death but you see no change in cleaved caspase, PARP, viability, Annexin V, PI, or TUNEL... you may not have “failed statistics.” You may have a readout mismatch.

Clear cell death. p = 0.056. Now what?

🔴 Case A: One point is weird (outlier). Do not blame biology yet. Check the experiment details: - reagent lot - serum bottle - passage number - seeding density - time out of incubator - edge effects - imaging field selection

You usually cannot remove an outlier just because it ruins your p-value. You need a technical reason. No technical record? Start writing it down for EVERY experiment.

Clear cell death. p = 0.056. Now what?

Step 2️⃣: Ask what the microscope is actually showing. The cells look different. Maybe it’s death. Maybe it’s lower confluence. Maybe it’s stress, detachment, morphology, growth arrest, or handling damage. Your eyes are allowed to raise suspicion. They are not allowed to finish the conclusion.

Step 3️⃣: Choose the next move. You'd typically have three cases:

Clear cell death. p = 0.056. Now what?

Instead, read the contradiction properly. Step 1️⃣: Know what the p-value means. A p-value is not a verdict. It does NOT tell you: - how big the effect is - whether the result is true - whether the effect matters biologically

Better definition: A p-value is the probability of seeing such data if we assume no effect in your model. So it's a surprise-o-meter. That’s it. Underwhelming, right?

Clear cell death. p = 0.056. Now what?

Most researchers make 2 mistakes when they see this. Mistake 1: Optimizing the statistics ❌ Changing the test until the result becomes significant. That is not analysis. That is p-value fishing in a lab coat.