YC Paper Club: speakers argue AI is turning cancer into a solvable data problem

A live Turing test experiment opens the night

The host ran an audience experiment, pairing attendees with either another human or an AI chatbot in blind chat sessions to test whether Alan Turing's 1950 prediction about the imitation game has been met. Results varied by model: "Astra won, Astra, then Claude, then GLM, then Gemini... Then Grock and Kimmy are tied," the host reported, noting that 33% of humans were mistaken for AI. He concluded that even with a strongly bolstered AI persona, humans still outperformed at being identified as human, so he does not believe the Turing test is fully solved.

Why AI-for-bio is arriving now

The host framed the moment as driven by collapsing sequencing costs and growing model scale, noting EVO2 was trained on 9 trillion base pairs and that AlphaFold is reportedly used "at some stage in every drug brought to market." He cited "10 to 100% success rates of AI protein binders" on a benchmark he called the Minecraft benchmark, and pointed to a company, Radical Numerics, moving AI-designed phages into patients. He argued biology now has something like the reinforcement self-improvement loop used in frontier AI labs: sample candidates, test them, retrain, repeat, for any task that can be verified computationally or physically.

Alfredo Gonzalez: treating each cancer patient as their own dataset

Alfredo Gonzalez, co-founder of Omnia and a former UCLA cancer bioinformatics PhD, argued that sequencing costs have fallen so far that genomic sequencing for cancer patients is now covered by insurers including Medicare and Medicaid. Where a landmark study like The Cancer Genome Atlas a decade ago could only sequence "a couple hundred patients" per tumor type, forcing researchers to rely on population averages, each patient's full genome (around 20,000 genes, plus RNA expression, splicing and copy number data) can now be treated as its own dataset to mine for that individual.

He described Omnia's approach as building an "agenticized" discovery pipeline — a personalized research lab run by AI agents instead of the 30-40 person bioinformatics teams previously needed. He gave examples of patients helped: a rare brain tumor patient whose data clustered closer to mesenchymal tumors, opening therapy options not supported by existing literature, and a patient whose prior three chemotherapy lines were, in retrospect, not supported by the underlying tumor biology. Asked about safety, he argued the approach increases safety by avoiding ineffective therapies, while acknowledging that fully novel compound design would require dose-escalation-style safeguards.

Yusuf Roohani: virtual cells and loops borrowed from AI research automation

Yusuf Roohani, now head of science at Intology, described "virtual cell" research as modeling biology at the cell, tissue and organism level rather than just molecules. He presented a prior project, the Universal Cell Embedding model trained on 35 million cells, which recovered known cell-type structure via self-supervised learning alone. A follow-up effort used agents to scrape and reprocess NIH transcriptomic data, scaling from 36 million to roughly 500 million cells — but revealed the data was overwhelmingly immune cells, with T-cells "over represented," limiting biological diversity.

His proposed fix is to shift from scaling raw data to "scaling AI-guided experimentation": using loop-based discovery borrowed from automated ML research. At Intology, his system Locus, built for autonomous post-training research, was applied to molecular binding prediction and reportedly outperformed off-the-shelf coding agents from Anthropic and OpenAI as compute scaled, though he stressed these are in-silico results requiring lab validation.

"We understand the structure of the universe better than the workings of living cells... we still cannot predict how a cell will behave if we mutate a previously unstudied gene."

Garyk Brixi: what EVO2 learned from training purely on DNA

Garyk Brixi, who led the EVO2 team and now works on the Manurva project, explained that EVO2 is a next-token prediction model trained on DNA across species, using a 9-trillion-base-pair dataset built by discarding redundant sequences and overweighting gene-dense, evolutionarily conserved regions during pretraining. He said this data-mixing choice was decisive: removing that augmentation cost "0.16 lower" Spearman correlation on a BRCA1 mutation benchmark — "the difference between being state-of-the-art and bottom of the pack."

He showed EVO2 can be used as a likelihood model to predict whether mutations are disease-causing, competitive with supervised models trained specifically on human data, and as an interpretability tool: sparse autoencoders revealed a neuron feature that fires on phage-associated DNA in E. coli genomes, including previously unannotated phages and CRISPR arrays. EVO2 has also been used generatively, fine-tuned to help design new phages, with Brixi noting such AI-designed phages are "making its way into patients." The newer Manurva model targets a weakness in EVO2 around predicting co-evolving, interacting DNA regions, and was used to identify new non-coding RNA systems, including an uncharacterized reverse transcriptase his team is still investigating in the lab.

Divya Nori: few-shot prompting works on a genome model, unexpectedly

Divya Nori, working in Brian Hie's lab, presented work showing EVO2 exhibits in-context learning (ICL) — generating novel, functional-seeming biological sequences from just a handful of examples in the prompt, without fine-tuning, despite this capability never being explicitly trained for. She said this was surprising because ICL "has not been widely documented to occur in regimes other than natural language pre-training." Prior approaches like EVO1 and EVO2 required fine-tuning on thousands of labeled examples; her team instead prompted EVO2 with as few as 14 known lanthanide-binding proteins to generate novel candidates, and with four CRISPR-Cas loci to generate new Cas9 systems.

In a wet-lab-validated case, her team combined few-shot prompting with reinforcement learning (using a sequence-to-function reward model) to design cell-type-specific enhancers, later tested in mouse brain tissue and reportedly shown to be active and specific in the targeted region. Asked why ICL emerges at all, she said one leading hypothesis is that "the burstiness" of long-tailed genomic data drives it, and that analysis in the preprint suggests an induction-head mechanism similar to that seen in language models.

What remains open

Several questions went unresolved. Roohani did not specify exactly how virtual cell simulations are validated against physical reality beyond general evaluation against experiments. Brixi acknowledged EVO2 still struggles with distal enhancer elements and that mechanistic-interpretability-based steering has so far underperformed other approaches. Nori said it remains unclear how much of EVO2's enhancer design ability reflects pretrained priors versus genuine in-context recombination of motifs. Gonzalez did not address a question from the audience about a Korean company's cancer-reversal claims, saying he lacked data to evaluate it.