Biology Foundation Models Are Growing Up, and Their Features Are Being Checked

A genome model tuned for single cells

In a preprint, Scooby-2 adapts the AlphaGenome foundation model for single-cell genomic track prediction, improving accuracy in gene expression and variant effect modelling at the single-cell level. The point of the fine-tune is precision about non-coding variants: knowing a variant exists is less useful than knowing which cell types it affects and how strongly. m:…7762

On the NeurIPS bone marrow dataset, Scooby-2 achieves a Pearson correlation of 0.56 when predicting cell-type-specific gene expression differences, compared with 0.53 for Scooby. Using the distilled AlphaGenome backbone, variant effect prediction concordance of fine-mapped expression quantitative trait loci from the OneK1K cohort improved from 0.37 to 0.47 Spearman rank correlation. m:…7762

Those are modest but meaningful gains on difficult prediction tasks, and they come from adapting an existing foundation model rather than training one from scratch for a single cell type. The preprint does not explicitly mention a public code repository URL in the abstract, which is a practical limit for teams wanting to reproduce the fine-tune. m:…7762

One clinical model, thousands of predictions

In a preprint, a generative foundation model called PhenOM completes predictive tasks across clinical domains without additional training. It achieves an area under the receiver operating characteristic curve of at least 0.70 in 80 percent of 1,770 validated predictive tasks, according to the paper. m:…2056

A model that can generate thousands of predicted outcomes without task-specific retraining changes the workflow for health teams: instead of building a separate model for each outcome, they query one model across many outcomes. That breadth is also the risk, because a single score across 1,770 tasks hides which tasks perform well and which do not, and the preprint summary does not break out the weakest tasks. m:…2056

The right question for a team considering such a model is therefore not whether it predicts in general, but whether it predicts the specific outcome, population and time horizon they care about, at a level they can verify locally. Breadth without that local check is a starting point, not a deployment decision. m:…2056

Can you trust what the model learned

In a preprint, researchers provide a ground-truth evaluation of sparse autoencoder features on spatial, single-cell and language-model embeddings. Consensus sparse autoencoders using five fits achieve 78 to 98 percent reproducibility, significantly higher than single sparse autoencoder fits at 27 to 46 percent, which means a feature found once may not be found again unless the fit is repeated and combined. m:…5666

Rotation-based methods including principal component analysis, Varimax and independent component analysis are highly reproducible but capture fewer biological concepts compared with sparse autoencoders, according to the paper. That creates a trade-off: a method can be stable and still miss the biology, or it can capture more biology and still be unstable from one fit to the next. m:…5666

For teams building on biology foundation models, the implication is direct. Before treating an internal feature as a discovered biological concept, check whether it survives repeated fitting and whether it maps to known biology. Scooby-2 and PhenOM show what these models can predict; the sparse autoencoder evaluation shows how carefully their explanations should be handled. m:…5666 m:…7762 m:…2056

That caution applies in both directions. A reproducible feature is not automatically a biological discovery, and a feature that appears in only one fit is not automatically meaningless, but neither should be reported without the repeatability check. The ground-truth test gives teams a way to run that check before they build a story, a target or a clinical hypothesis on top of a model explanation. m:…5666