sanj

New ways of quantifying cellular states are coming up everyday, with different granularities & modalities of information. AI cell models that we know of today use RNA-seq as the primary modality of learning - simply due to the ease and abundance of capturing this. I've been thinking about this as a problem, as a species we will always come up with newer ways of learning about our biology. 1. how do we connect this general cell state to specific modalities of functional readouts? 2. are there any additional benefits to solving this problem of unification?

Love this release. Their open weight model is apt for the overall tinker ecosystem AND it’s comparable to the frontier in model capabilities!

Insilico initiates phase III clinical trial for Rentosertib, its AI-empowered TNIK inhibitor for idiopathic pulmonary fibrosis #NBTintheNews insilico.com/news/xmjsn4l09…

Everyone in pharma cites the looming patent cliff of $300B between 2025 and 2030 and how it accounts for 1/6 of pharma revenue. So I wanted to put this in context with other periods in history. The 2010-15 revenue risk was almost double 2025-30 (as a % of total pharma revenue).

Excited to share our paper at @icmlconf genbio right now. We ablate the core design choices in latent perturbation prediction and then tested the model on a real-world out-of-distribution target discovery task from Pfizer, achieving SOTA results. Read more from @splicewiring!

Also checkout our more recent progress at @atlasdiscovery0 !

We were curious whether the autoregressive next token prediction recipe that has worked for language can be applied to single-cell biology. We first tried to understand if a discrete representation is the right substrate for modelling in this domain? Tldr: Yes! We studied a discrete latent recipe for training perturbation prediction models, that sets a new SOTA for most cell-eval metrics. We also asked: what does "best" even mean here? There are mean-based metrics, which mostly ask whether you got the average perturbation effect right. And there are distributional metrics, which ask whether you captured the shape and diversity of the predicted cell population. As in, what is special about the models that perform on the mean based metrics class, as opposed to distribution based and vice-versa? (P.S. It’s not a mutual win in the literature we’ve found so far) It turns out that a lot of this variance becomes clear when viewed through one axis: sampling richness. Some heads are better at collapsing toward a mean. Others are better at producing the distribution. Which one “wins” depends heavily on the ruler. We then looked at a more biologically relevant task, where in a model is used to rank / predict something and then that’s reviewed from an expert and subsequently scored. On a held-out CRISPRi inflammation-reversion screen, the our encoder ranks target genes with 0.79 AUROC, matching scGPT while seeing roughly an order of magnitude less data. The benchmark carries several distribution shifts at once (cell type, perturbation set, biological context), and with a single dataset we cannot fully attribute the result to any one axis of generalization. However, we hypothesize that this is because the models are trained on perturbational data and not just observational data. cc: @Cgensbigler and @ShaamilKarim1

@jacobkimmel Very interesting early results for continual learning type systems!

# generalizing across cell lineages there are far more reprogramming medicines than we will ever be able to test experimentally. to navigate this space, we’ve built frontier AI systems that perform reprogramming experiments in silico. our models were historically trained to design reprogramming medicines for one cell type or lineage at a time. this summer, we unlocked cross-lineage generalization. our systems can now leverage data from one of our programs to accelerate others. in one example, generalization improved the sample efficiency of our models by >3X. this means that initiating a new program requires far fewer experiments in the world of atoms.

this is our most dynamic summer yet @newlimit - 3X data efficiency for reprogramming AI from lineage generalization - >100X manufacturing scale for our lead therapeutic asset - +4 leads that restore function in old hepatocytes, +5 payloads in endothelium & lots more in our blog

Frozen eVAE encoder, trained on 1M interventional cells, on a held-out CRISPRi inflammation-reversion screen: 0.79 enrichment AUC, matching scGPT, which outperforms simpler baselines at roughly 10x less data. There are certain limits here: the benchmark mixes cell-type, perturbation, and context shift, so we cannot cleanly assign the gap to one axis of generalization.

Now for part (b), Papers disagree on which model, and which decoder head, is best, largely because they report different metrics. We ran a controlled head ablation on one frozen backbone to find out why. Every standard metric encodes a hidden position on ONE axis: how rich the inference-time predictive distribution is. The decoder head's sampling rule slides you along it. And the split cuts straight through the cell-eval metrics, not just distribution vs DE. PDS, PR-AUC, Sp-sig want sampling; Sp-LFC, overlap@N, Pearson-Δ want a deterministic mean. Takeaway: a lot of what reads as "model quality" on a single metric is really "did you sample or take the mean."

1/ Foundation models of the transcriptome, trained with generative pre-training or masked language modeling objectives, are confounded by gene co-expression networks. They shine at tasks like cell type and disease classification, but fail at downstream applications that

@anshulkundaje @genophoria @BoWang87 What would you say is a good test if the models are encoding real biology? Benchmarking is a huge gap in the domain, it’s great that we’ve started to think about it!

@genophoria @BoWang87 More importantly none of the papers even attempt to test if the models are encoding causal trans regulation machinery. Just benchmarks on poorly defined tasks with poor performance metrics. How would anyone even know what the models are learning?!?

With all due respect, I'd like to offer a few points of clarification. First, I have no issue with "shortcut models." In fact, many of my own papers use relatively simple models to solve important real-world problems. If a simpler model ultimately proves capable of capturing complex cellular biology and helping cure disease, I'd be delighted. Science should reward what works, not what is most sophisticated. Second, terms like virtual cells, foundation models, and world models are high-level concepts that describe a class of models rather than a specific algorithm. Similar terminology has emerged naturally in computer vision and NLP as the field evolved. I think it's reasonable to adopt analogous concepts in biology as we explore whether they can unlock similar advances. Whether these ideas ultimately live up to their promise is, of course, an empirical question. Rigorous validation will decide. This is exactly what my original post is about. Healthy skepticism is essential, but so is giving ambitious new directions the opportunity to prove (or disprove) themselves. I don't think we should dismiss a promising research direction simply because the terminology sounds aspirational 🙏🙏

Imo, the strongest case for this argument is that these tasks are not independent. Perturbation response, differentiation, drug response, and cell-state transitions are all readouts of the same regulatory machinery. If we can learn that machinery, broad generalization should follow as a matter of principle. As you said, whether current models do that is the empirical question.

I spend most of my days working with millions of samples across tissues, diseases, and perturbations. Here’s how it goes: months go into defining an objective - it’s pretty much defined over the latent space a model has learned - so you’re just trying to get the model to organize the biological state into a geometry that means something. And for a long time (the past few months) the honest answer to "does it mean something?" is a shrug and a few loss curves. When you train these things, the recurring fear is that you've built an elaborate compression of noise, something that reconstructs beautifully and predicts nothing you actually care about. We decided to define an experiment to understand if the pre-training bet adds gains in the clinical space. This space needs a lot of innovation, this is known at this point. The key piece in a trial design is to figure out which person would respond to a drug, and who would not. We used a pre-trained model to embed patient samples from the UNIFI phase 3 trial, to then classify this stratification. We were able to get to a 0.76 AUROC, and derive relevant gene sets by probing the model further. Details in the thread. cc: @Cgensbigler @ShaamilKarim1 What it changed for me is smaller and more personal: it's the first real evidence that the thing I've been staring at loss curves for is learning the structure of human disease, not just memorizing it. One drug, one indication, one cohort, a model trained on a fraction of the data that exists. I think it's a floor, not a ceiling.