Stanford's Kundaje Hails New Gene Regulation Model as Paradigm Shift, Not Brute-Force Scaling

Anshul Kundaje, a Stanford professor of computational genomics, posted a lengthy thread on November 12 highly praising a newly released gene regulation model (related to his collaborator Avsec, likely an AlphaGenome-type work and a major upgrade over the previous AG1). He considers it a paradigm shift in model design that goes well beyond the earlier multi-task paradigm and points in the right direction—toward "causal cis-regulatory gene regulation models."

Confirmed

• The model was trained on 2048 TPUv4 chips; the team was meticulous at every step, from data design and model design to engineering, squeezing maximum value out of large datasets.
• Kundaje stresses the model is not a product of brute force scaling: the architecture, training scheme, loss functions, and data normalization were all carefully designed by people with deep domain knowledge, making it a substantive, non-trivial extension of prior SOTA; the paper's main text and methods sections are excellently written, with the methods being a "gold mine."
• Training data came from public academic data curated over years by consortia such as ENCODE, GTex, and Cell x Gene; he cautions against underestimating the value of open academic data resources.
• His lab has already used the model for GWAS variant-to-gene prioritization: by predicting cell states not covered in training, it clearly outperforms earlier models; variant effect and E2G (enhancer-gene) link resources are expected to be highly valuable.
• On architecture, he believes "cell-state conditioning" offers a natural path to unifying sequence models with multimodal cell-state encoding models; he proposed an early version of this paradigm back in 2019 with ChromDragoNN, now validated by a frontier model.
• His message to "virtual cell" enthusiasts: the model's single-cell encoder is just an improved version of the classic scVI and works very well—you don't need billion-parameter, massively scaled encoder-decoder "foundation world models."

Why it matters

• This is an endorsement from a senior figure in the field for the route of "domain-knowledge-driven careful design + open academic data," in contrast to the brute-force parameter stacking of general foundation models, offering methodological guidance for the AI x Bio community. Kundaje also notes the model is still short of a fully causal cis-regulatory gene regulation model, but the direction is right.

2026-11-12 ~ 2026-11-12 · 10 related posts

• [source] Stanford professor hails new gene-regulation model as a real paradigm shift toward causal cis-trans models — anshulkundaje · 2026-11-12
• [source] Genomics professor hails new gene regulation model: methods section is a gold mine — anshulkundaje · 2026-11-12
• Professor's lab tests new gene-regulation model: stronger GWAS variant prioritization in unseen cell states — anshulkundaje · 2026-11-12
• Professor: gene-regulation model was built on open academic data from ENCODE, GTex and Cell x Gene — anshulkundaje · 2026-11-12
• Not brute-force scaling: professor praises domain-crafted architecture and loss design in gene-regulation model — anshulkundaje · 2026-11-12
• [source] 2048 TPUv4s and deep domain expertise: professor dissects how a gene-regulation model was trained — anshulkundaje · 2026-11-12
• Stanford professor: virtual cell models don't need billion-parameter foundation models — tweaked scVI did fine — anshulkundaje · 2026-11-12
• Cell-state-conditioned architecture may unify sequence models with multimodal cell-state encoders — anshulkundaje · 2026-11-12
• Professor: the cell-state-conditioned paradigm I proposed in 2019 is now validated by a frontier model — anshulkundaje · 2026-11-12

1 near-duplicate retellings: anshulkundaje