May all the reported results published in 2024 be found to be reproducible in 2025.

A p value <0.05 does not mean there's a 95% chance your data will repeat. It means the probability you failed to falsify your hypothesis when your hypothesis was false is <5%. And even that claim is only accurate if the N accounts for the variability of the effect given the effect-size desired.

The most commonly used "frequentist" statistics, such as t tests, ANOVAs etc, fail to give you a probability that your data will repeat. The closest a frequentist comes to this is with a "confidence interval." The best way to get to probability is with Bayesian approaches.

If your published dataset does not match the actual full dataset from an experiment, you have some explaining to do.

There are valid reasons to reject data; you need to list those and justify them. It's not ok to reject data that simply fails to align with your hopes & dreams.

The problem with much of what gets published is that the results claimed are not predictive of what others might find upon repetition.

The main reasons: lack of reagent validation; lack of a large enough N to capture variability of the effect; cherry-picking "desired" results.

Your data is a testament to what happened when you performed an experiment.

Your job as a scientist is to determine if that data is predictive as to what will happen the next time the experiment is done - and to what degree (what probability) that prediction is accurate.

There’s an online iBiology course called ‘Let’s Experiment’, plus books by @steveroyle.bsky.social & @davidjglassmd.bsky.social ‘The Digital Cell’ & ‘Experimental Design for Biologists’ which I found really useful

If you add neural agrin to a muscle cell, it's sufficient to induce NMJ formation. Without agrin, there is no junction. So we say agrin is necessary and sufficient for that effect.

But agrin does nothing to a fibroblast, because it needs its receptors. So the context matters.

If you knock out the receptor MuSK you eliminate neuromuscular junction formation.

Therefore MuSK is necessary. However, if you add MuSK and its ligand agrin to a fibroblast, you won't get signaling, because you need the co-receptor LRP4.

So MuSK is necessary but not sufficient.

Biology is messy - because there are often signaling networks rather than linear pathways.

There are examples of being neither necessary nor sufficient, but still being causal:

For mTORC1 signaling, Akt activation and amino acid activation are both causal, but neither may be sufficient. In many settings you don't need Akt, but it still can cause mTORC1 activation.

What's the most convincing way to prove causation, after you've shown correlation?

In biology, it's to show necessity. For example, if you want to ask if X causes Y, you knock out X and see if Y is decreased, or stops happening.

Correlation does imply causation - it just doesn't prove it.

If X causes Y then X has to be correlated with Y.

It's just that there are plenty of instances where X is correlated with Y but has nothing to do with Y.

I was reading Charles Darwin's Origin of the Species. I didn't expect the example he gave, from back in 1859, of a change which could cause species extinction: climate change.

Can we stop using the 6 minute walk test, and the stair climb test? These very short, motivation-based exercises have killed more muscle drugs than I care to mention.

How about using modern 24-hour monitoring devices -like your smart watch to get more objective measurements?

Give a scientist you know the gift of improved experimental design:
www.amazon.com/Experimental...

Grad school should be like Med school - it should teach students how to do their jobs as scientists. Failing to teach them experimental design and statistics is criminal.
They don't choose their projects, so judging them on the success of their projects makes no sense.

RNAseq is commonly performed. We compared N=30 wild-type Bl6 mice to an N=30 heterozygotic knockout with a phenotype, and then downsized. We found you need at least an N= 8-12 to avoid a >50% false positive rate. A 2nd line confirmed the finding.

www.biorxiv.org/content/10.1...