Zoë — I know you keep coming here.
The baby doesn't change anything for me.
If you want this, act. If you don't, stop coming here and forget about me.
The baby doesn't change anything for me.
If you want this, act. If you don't, stop coming here and forget about me.
Paper: Generative ranking enables scalable pretraining on noisy biological multisets
We built a world model of the cell—a single self-supervised model, trained on rankings of mRNA counts, that encodes:
- Where proteins sit in the cell — from nucleus to membrane to secreted
- Which proteins physically interact — direct substrates, scaffolds, complex partners
- How receptors signal — multi-hop kinase cascades recovered zero-shot
- Whether a target is druggable — and by which modality
- What transcription factors bind — including the sign of regulation and complex composition
State-of-the-art perturbation prediction. Monotonic scaling. No injected knowledge. No task-specific pretraining.
The model learned the cell because the cell is what generates the data. The right objective was all that was missing.
GCTCAGAAGCGCCGAGAGCGCGGCCGGGACGGTTGGAGAAGAAGGCGGCTCCCGGAAGGGGGAGAGACAAACTGCCGTAACCTCTGCCGTTCAGGAACCCGGTTACTTATTTATTCGTTACCCTTTTTCTTCTTCCTCCCCCAAAAACCTTTTCCTTTTCCCTTCTTTTTTTTTCCTTTTTGGGAGCTGAAAAATTTCCGGTAAGGGAAAGAAGGGCTCCTTTCGCTCCTTATTTCCCCGCCTCCTTCCCTCCCCCACCTTCCCCTCCTCCGGCTTTTTCCTCCCAACTCGGGGAGGTCCTTCCCGGTGGCCGCCCTGACGAGGTCTGAGCACCTAGGCGGAGGCGGCGC
GCTCAGAAGCGCCGAGAGCGCGGCCGGGACGGTTGGAGAAGAAGGCGGCTCCCGGAAGGGGGAGAGACAAACTGCCGTAACCTCTGCCGTTCAGGAACCCGGTTACTTATTTATTCGTTACCCTTTTTCTTCTTCCTCCCCCAAAAACCTTTTCCTTTTCCCTTCTTTTTTTTTCCTTTTTGGGAGCTGAAAAATTTCCGGTAAGGGAAAGAAGGGCTCCTTTCGCTCCTTATTTCCCCGCCTCCTTCCCTCCCCCACCTTCCCCTCCTCCGGCTTTTTCCTCCCAACTCGGGGAGGTCCTTCCCGGTGGCCGCCCTGACGAGGTCTGAGCACCTAGGCGGAGGCGGCGC
TACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAACCGTAGACCAGATAGCATAGACATACCGTAGACCAATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGATAGCATAGACATACCGTAATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACA
TACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAACCGTAGACCAGATAGCATAGACATACCGTAGACCAATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGATAGCATAGACATACCGTAATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACA
TACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAACCGTAGACCAGATAGCATAGACATACCGTAGACCAATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGATAGCATAGACATACCGTAATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACA
TACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAACCGTAGACCAGATAGCATAGACATACCGTAGACCAATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGATAGCATAGACATACCGTAATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACATACCGTAGACCAGATAGCATAGACA
Essays
Through a Glass Darkly: Mechanistic Interpretability as the Bridge to End-to-End Biology
November 2024
Single-Cell Encoder Models: Toward a Statistical Physics of Emergent Agency
June 2023
Hyperbolic Science: Or, Playing Games with Growth
August 2022
A Future History of Biomedical Progress
August 2022