Biological foundation model
A model trained over biological data (sequences, structures, single-cell measurements, clinical
records) to be reused across tasks the way a language model is, rather than trained per question.
In this spoke it is an ambition with infrastructure behind it and no held instance:
israel-ai-bio-consortium is funding the tokeniser ("Bio Tokens") and the training platform
("Factory Model"); autism-protein-interaction-map shows a finished model (AlphaFold) used as an
instrument to rank candidates; hippocampus-aging-epigenome is the kind of single-cell dataset such
a model would be trained on. The founding gap is the model itself: no page here describes one, its
training data, or a benchmark for it — growth edge 1. Closed 2026-09-22: geneformer is the first
held instance: six transformer layers pretrained on 29.9 million single-cell transcriptomes, whose
predicted cardiomyopathy targets (PLN, GSN, ESRRG, HMGB1) improved contraction when knocked out in
iPSC-derived cardiac tissue.
What pretraining buys by itself
The same model looks different depending on who tests it. Its builders report strong results after fine-tuning on tasks they chose (geneformer). An outside group from Oxford and Microsoft Research used its embedding with no fine-tuning, and a plain filter for highly variable genes beat it on every metric (single-cell-foundation-model-zero-shot). So the value measured so far sits in fine-tuning on labelled data. Whether pretraining helps where there are no labels, which is where discovery happens, is still open.
From model to medicine
rentosertib-phase-2a follows the other end of the pipeline: a target and a molecule both produced by one company's AI platform (insilico-medicine), taken into a placebo-controlled phase 2a trial of 71 patients. That platform is not a foundation model in the sense above. The trial shows only that an AI-derived drug can enter and complete a controlled trial with its adverse events published, and it is the sponsor's own report.
Distinct from machine-learning-wiki's treatment of models as models; a biological foundation model is here for what it learns about biology, and there for how it is trained.
Related
israel-ai-bio-consortium · autism-protein-interaction-map · single-cell-multiomics · geneformer · single-cell-foundation-model-zero-shot · rentosertib-phase-2a · synthesis