Articles

A thinking tool, not a decision tool

Virtual cell models trained on causal perturbation data are a real scientific step up. Looked at from a regulatory perspective, they still sit upstream of anything FDA would accept as evidence.

3 min read

I read the preprint last night, and this is a lot of work. Massive dataset, CRISPR screens, multiple layers of biology pulled together. Definitely not something done overnight.

To be fair, what they are doing is genuinely hard. Predicting how a cell changes after you knock down a gene, across different biological contexts, is not a simple modelling problem. So performance around 0.5 Pearson reflects real signal.

Training on causal data is the right foundation. But I am still asking myself: does 0.5 Pearson mean the model has actually learned to predict causal outcomes, or is it still learning the pattern?

Vioela’s note: A quick way to read that number. A correlation of 0.5 sounds like half right. Square it and you get the share of the variation the model actually accounts for: about a quarter. Three quarters of what the cell does is still unexplained. That is real signal for a problem this hard, and it is also a long way from something you would stake a patient on.

Where I think it fits

Looking at it from an FDA perspective, I have mixed feelings.

Where this fits well is early stage. Pre-IND thinking, hypothesis generation, helping frame mechanism of action before you commit to expensive experiments. That is genuine value and I do not want to undersell it.

But not beyond that. Not yet.

  • It does not replace tox.
  • It does not replace clinical.
  • And FDA is not going to accept it as standalone evidence.

What I like about the direction

Moving toward causal biology instead of observational correlation is the right instinct. Most models in this space are built on data that captures association. From what I understand, this is trying to capture what actually happens when you intervene, and that is a real scientific step up.

Where I still struggle

Translation. The same target can behave one way in immune cells and completely differently in vascular or other tissues.

And there is a bigger gap: no PK, no dose, no system-level biology. Even if the signal is real, you still do not know whether it matters at the exposure you are actually giving to a human.

The regulatory angle

There has been a sustained push toward more data-driven drug development, and tools like this fit that direction. But one of the real challenges FDA faces is that the agency already holds enormous amounts of data across submissions, studies and programmes. The harder problem has been making that data usable and consistent enough to draw meaningful comparisons across it.

This kind of model works upstream of that problem rather than solving it, but the two are not incompatible.

So is this a big improvement?

Scientifically, yes. Regulatorily, not yet.

Right now it feels like a thinking tool, not a decision tool.

What I would most like to see next is retrospective validation: go back to programmes with known clinical outcomes and ask whether this model would have predicted them. That would tell us a great deal about where it actually stands, and it would be a meaningful step toward making something like this usable in a regulatory context.

Vioela’s note: Retrospective validation has one trap that decides whether it means anything: leakage. If the historical cases used to test the model were already in its training data, in any form, the model is being graded on an exam it has already seen. The clean design is a time split. Train only on what was known before a cutoff date, then ask the model to predict outcomes that came after it. FDA’s 2025 draft guidance on AI in regulatory decision-making points the same way: the more a model’s output drives a decision, and the worse the consequence of being wrong, the more evidence of credibility it needs. Early research use sits at the low end of that scale. A dosing or safety decision sits at the top.

Adapted from a comment first published on LinkedIn, on the launch of a genome-wide virtual cell model.