Biology AI stopped looking at the tissue and started looking at the cell
AlphaGenome 2 reads DNA effects per cell type. TxFM beats models trained on 100x more data. ProTEUS evolves proteins in a closed loop.
- AlphaGenome 2 was trained on more than 100 million individual cell profiles across 68 tissues, and identifies nearly twice as many variants that change gene expression in cells compared to the previous model. Self-reported m:…8083
- AlphaGenome 2 identified 30% of causal variants in the TenK10K dataset compared to 16% by the original AlphaGenome. Self-reported m:…8083
- TxFM, a masked autoencoder for RNA-seq, outperforms atlas-scale foundation models trained on corpora 100x larger. Self-reported m:…6954
- The ProTEUS all-atom predictor is >10 times faster than AlphaFold3, and ProTEUS-engineered variants showed >10-fold higher activity than wild type. Self-reported m:…1807
One cell at a time
Google DeepMind announced AlphaGenome 2, a next-generation model that predicts how DNA sequences and their variations impact specific human cell types. AlphaGenome 2 was trained on more than 100 million individual cell profiles across 68 tissues, and is a conditional model that predicts DNA effects for one condition at a time. The shift from bulk tissue to single cells is the whole story: the same DNA variant can matter in one cell type and not in another, and a model that averages across tissues cannot see that. m:…8083
The results show the payoff. AlphaGenome 2 identifies nearly twice as many variants that change gene expression in cells compared to the previous model, and identified 30% of causal variants in the TenK10K dataset compared to 16% by the original AlphaGenome. AlphaGenome 2 can generalize to new single-cell experiments without needing to be retrained, by compressing gene activity into a cell state representation. AlphaGenome 2 predicts gene expression, histone modifications, chromatin accessibility, and DNA methylation. m:…8083
The model is available today for non-commercial research via the AlphaGenome API, and will soon be released as AlphaGenome 2 Skills in Google Antigravity. AlphaGenome 2 is not validated or approved for clinical or diagnostic purposes, which is the honest boundary: this is a research tool for understanding disease, not a diagnostic device. m:…8083
Smaller data, better design
TxFM makes the case that the field has been overvaluing scale. TxFM, a masked autoencoder for RNA-seq, outperforms atlas-scale foundation models trained on corpora 100x larger, and the DiverseRNA-1.4M dataset provides a high-fidelity training corpus for biological representation learning. Inductive self-supervised learning is a viable modeling approach for transcriptomics when architecture is carefully synthesized with data curation. m:…6954
The preprint's argument is that data quality and architectural fit beat raw corpus size, which cuts against the prevailing bigger-is-better story in foundation models. A curated dataset, DiverseRNA-1.4M, matched to an architecture designed for count data, beat models trained on corpora 100x larger. For drug discovery teams, the implication is that the budget question is not how much data you can afford, but how well you can curate the data you have. This is a preprint, so the results are not peer-reviewed, and the claim rests on the authors' own evaluation. m:…6954
Evolving proteins in a loop
ProTEUS is a mutation-resolved world model for protein evolution, and it closes the loop that AlphaGenome 2 opens. ProTEUS combines sequence, evolutionary, and structural representations, and uses MutationFormer, a mutation-biased attention architecture, to represent variants as states in a joint latent space. ProTEUS uses closed-loop experimental feedback to refine predictions, and outperforms established models in zero-shot mutation prediction tasks. m:…1807 m:…8083
The speed claim is what changes the workflow. The ProTEUS all-atom predictor is >10 times faster than AlphaFold3, and uses explicit geometric constraint enforcement to maintain physical validity. ProTEUS was tested on aldehyde dehydrogenase and levansucrase, and ProTEUS-engineered variants showed >10-fold higher activity than wild type. A model that predicts structure >10 times faster and then learns from the experiment it suggested is not a prediction tool. It is a design loop. m:…1807
The papers share a direction. AlphaGenome 2 reads DNA at the resolution of a single cell type. TxFM shows that careful curation beats a corpus 100x larger. ProTEUS turns prediction into a loop that improves with each experiment. Biology AI is moving from describing what exists to designing what should exist, and the resolution keeps getting finer: from tissue, to cell, to mutation. m:…8083 m:…6954 m:…1807
What is still unresolved
- Whether AlphaGenome 2's variant identification holds up on independent datasets beyond TenK10K.
- Whether TxFM's advantage over larger models survives peer review, given it is a preprint.
- How far ProTEUS's closed-loop gains extend beyond the two enzymes tested.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.