The evening's answer to its title is a qualified yes, resting on one idea all four speakers return to: biology is becoming a loop rather than a pipeline, in which models propose, the lab or a predictor tests, and results feed back. They show it at the scale of single patients, cell models, DNA language models and prompted sequence design. Most of the evidence is in silico; wet-lab confirmations are a handful.
The premise. Host Francois, at his tenth and last YC Paper Club, frames the night around a loop he says frontier labs have worked out for any task whose output can be verified: sample candidates, select the best, test them, retrain. Verification in biology costs money and time, but he argues the procedure is the same and believes "it's just going to work." He cites Evo 2's 9 trillion training base pairs and Demis Hassabis's claim that AlphaFold now touches every drug brought to market.
The patient as dataset. Alfredo, a UCLA cancer-bioinformatics PhD and co-founder of a YC company that builds a personalised research lab around each patient, argues that cheap, insurance-covered sequencing has made cancer a data problem. TCGA could sequence a few hundred patients per tumour type, enough only for averages; today each patient carries roughly 20,000 genes of data plus copy number, expression, splicing and cell state. The 20–40-person teams that once made this possible only at the Broad or Sanger can, he says, be rebuilt as agents. His cases: a rare brain tumour that clustered with mesenchymal tumours, opening therapies the literature did not suggest, and a patient whose three cytotoxic chemotherapy lines were, by his reading of the data, ill-suited and avoidable. Asked about safety, he argues the approach raises it, since most options have known safety profiles.
Loops for virtual cells. Yusuf Roohani defines a virtual cell as AI modelling biology at the cell, tissue and organism level, aimed at clinical efficacy, which he names as a leading cause of failure. His universal cell embedding model, trained self-supervised on 35 million cells, recovered known cell-type structure unprompted. Agents then spent a month reprocessing NIH's public transcriptomic data, reaching 500 million-plus cells, but the corpus was mostly immune cells, T cells over-represented by an order of magnitude. His conclusion is to scale AI-guided experimentation rather than data generation. At Intology, the system Locus, built for automated language-model post-training, beat off-the-shelf coding agents on molecular binding prediction as GPU compute scaled, and sampled foundation-model latent spaces more efficiently than BoltzGen and Codex. He stresses these results are a day old and still need lab validation.
Reading DNA. Garyk, who led Evo 2, describes next-token prediction on DNA at 7B and 40B parameters. The training set kept one genome per species and up-weighted genes and conserved regions before context was extended to 1 million base pairs. That curation, he argues, was decisive: without it Spearman correlation on a BRCA1 mutation benchmark drops by 0.16, from state-of-the-art to bottom of the pack. Evo 2 matches or beats human-specific supervised models on variant effects despite seeing one human reference genome. Sparse autoencoders surfaced a phage feature that also fired on unannotated prophages and CRISPR arrays. A fine-tuned Evo 2 designed phage genomes, and a newer bidirectional model predicts co-evolving base pairs, revealing hundreds of missed non-coding RNAs and a lab-validated reverse transcriptase carrying a CRISPR-like array. He concedes Evo 2 is weak on distal enhancers.
Prompting instead of fine-tuning. Divya, in Brian Hie's lab, reports that in-context learning emerged in Evo 2 from pre-training alone. Earlier designs needed tens of thousands of CRISPR-Cas loci or thousands of Microviridae genomes for fine-tuning; now 14 known lanmodulins, four CRISPR-Cas loci, or structure-model antibody hits in the prompt yield novel candidates predicted to function. GRPO post-training against a sequence-to-function reward improved how well the model used its prompts, and a resulting enhancer proved active and specific in mouse brain. She attributes the emergence, as a hypothesis, to bursty, long-tailed genome data and induction-head-like mechanisms, and expects prompt pools updated from lab results to close the loop.