Is AI Now Good Enough To Cure Cancer? | YC Paper Club
The YC AI for Bio Paper Club event explores how artificial intelligence is transforming cancer research and biological discovery through foundation models, personalized genomic analysis, and autonomous research systems. Speakers demonstrate that cancer is becoming a data problem solvable through AI agents, DNA language models like EVO2 are enabling efficient protein and genome design through few-shot prompting, and validated lab results show these approaches are moving from theory into clinical practice.
Summary
The event begins with a playful Turing test demonstration where audience members attempt to distinguish between AI chatbots and humans in chat conversations, finding that Astra and Claude perform best but still fall short of the 70% threshold for passing the original Turing test. The host then transitions to the core theme: AI's potential to cure cancer through massive advances in biological AI.
Alfredo Romero from AMA presents cancer as fundamentally a data problem. Historically, landmark cancer genomics studies like TCGA could only sequence a few hundred patients per tumor type. Today, with plummeting sequencing costs (now cheaper than dinner in many Bay Area restaurants), cancer genomes can be sequenced affordably and are covered by insurance for advanced tumors. This generates ~20,000 gene data points per patient across multiple modalities including copy numbers, RNA expression, and cell states. Romero describes a paradigm shift: instead of analyzing populations for common patterns, AI agents can now treat each patient's complete genomic profile as an individual dataset to mine. AMA builds 'personalized research labs' using AI agents—replacing 30-40 person research teams from past landmark studies—to analyze tumor evolution, identify optimal therapies in real-time, and correct treatment courses. Case studies show rare brain tumors being reclassified by biological pathways rather than tumor type, enabling novel therapies, and failed chemotherapy patients being redirected to better-matched treatments based on tumor heterogeneity analysis. The future involves designing proteins and delivery mechanisms in real-time for each patient.
Yousef Ruhani discusses virtual cells and autonomous AI research. He frames biological discovery as a search problem over hypotheses, where AI accelerates this search. While models like AlphaFold excel at sequence level, the challenge is moving to physiologically relevant levels—cells, tissues, organisms—where multiple interconnected variables create complexity. Ruhani's universal cell embedding model trained on 35 million cells creates a representation space showing that cell types cluster without explicit labels, validating self-supervised learning. However, when processing 500+ million cells from public repositories using AI agents, the team discovered severe data imbalance: T cells are overrepresented by an order of magnitude, limiting diversity. Rather than simply scaling data first, Ruhani proposes 'AI-guided experimentation'—using language models like Claude to optimize which experiments to run, creating a discovery loop rather than linear data generation. He draws parallels to autonomous ML research at frontier labs, where similar challenges (costly evaluations, noisy measurements, vast search spaces) exist. Using a system called Locus optimized for scientific discovery, his team shows that models fine-tuned on automated ML research also perform well on molecular design tasks, improving property prediction and generative sampling efficiency. This suggests shared primitives between AI research automation and scientific discovery automation.
Garrick Lowe presents EVO2, a DNA language model trained on 9 trillion base pairs using a hybrid attention-convolution architecture. DNA presents unique modeling challenges: only 4-letter vocabulary, multi-scale evolution from single base pairs to hundred-thousand base pair genes, and highly redundant sequences from parasitic copying. Rather than using all available data, the team carefully curates by taking one representative genome per species and focuses training on annotated genic regions with less repetitive junk. The model is trained in two phases: first on shorter contexts emphasizing conserved regions, then extended to 1 million base pair contexts. EVO2 learns to predict the likelihood of DNA sequences, and this likelihood correlates with disease impact—mutations the model rates as unlikely often cause disease or loss of function. Remarkably, without seeing human-specific training data, EVO2 is competitive with or superior to supervised models on benchmarks like BRCA1 cancer risk mutations. Through sparse autoencoders, the model has discovered internal representations of biological concepts: one feature fires on phages, prophages, CRISPR arrays, and phage-derived genes—diverse functions connected by their phage origin that the model learned purely from sequence prediction. A new bidirectional model called Manurva specifically learns co-evolution between genome regions, enabling discovery of non-coding RNAs and CRISPR-like systems previously missed in well-studied bacterial genomes, including the validated discovery of UG27 reverse transcriptase with programmable arrays.
Divia Agarwala demonstrates few-shot prompting of EVO2 for biological design, showing an emergent in-context learning (ICL) capability that parallels language models but was unexpected in genome models trained purely on next-token prediction. With just 14 examples of rare-earth-metal-binding lanmodulins, EVO2 generates novel predicted lanmodulins; with 4 CRISPR-Cas9 examples, it generates complete functional systems including arrays and tracr RNAs. For whole phage genomes, few-shot prompting achieves predicted functionality at a fraction of fine-tuning compute cost. The team validates one application—cell-type-specific enhancer design—in the wet lab: after prompt-based generation, reward model training using a sequence-to-function predictor, and RL-based post-training on EVO2, a designed enhancer shows predicted specificity when tested in mouse brain tissue. This establishes a lab-in-the-loop paradigm where initial prompts from natural sequences are iteratively refined through testing, design, and prompt updates.
Key Insights
- Alfredo Romero argues that cancer is now fundamentally a data problem: sequencing costs have become so low that patients can be treated as individual datasets with ~20,000 gene data points each, rather than aggregating across populations to find common patterns as was necessary when only hundreds of samples existed.
- Yousef Ruhani demonstrates that when processing all publicly available transcriptomic data with AI agents, T cells are overrepresented by an order of magnitude, revealing that simply scaling data collection without guidance produces imbalanced datasets, suggesting AI-guided experimentation is more effective than bulk data generation.
- Garrick Lowe shows that EVO2, trained only on nucleotide prediction without human-specific training data, achieves competitive or superior performance to supervised models on clinical mutation effect prediction by focusing training on annotated conserved regions rather than repetitive genome sequences.
- Garrick Lowe reports that sparse autoencoders reveal EVO2 learned an internal feature that fires on phages, prophages, CRISPR arrays, and phage-derived bacterial genes—conceptually related biological entities that are sequentially diverse but were unified by the model through pure next-token prediction.
- Divia Agarwala demonstrates that EVO2 exhibits in-context learning capabilities for biological design: prompting with just 4 CRISPR-Cas9 examples generates complete functional systems, and post-training with reinforcement learning on a designed enhancer produces a sequence validated to be specific in mouse brain tissue.
Topics
Transcript
[0:05] [music] How's everyone doing today? [screaming] Uh, lots to talk about tonight. YC AI for Bio Club. Very exciting night tonight. Some of my favorite uh, speakers and a couple of my favorite um, folks in the in the lab that are coming and even some of my favorite folks um, that I worked with at YC are coming tonight to speak. So, I'm super excited. But before that, I kind of have a crazy idea. I need everyone to take out their phone and scan this. And then I'll tell [0:36] you why in a second. Okay. So, I had uh kind of an argument with um Greg at Arc AGI who we had this dinner and he…
Full transcript available for MurmurCast members
Sign Up to AccessMore from Y Combinator
We’ll Be Inferior to Our Grandchildren—and That’s Great
A discussion about scaling societies and the relative ratio of good versus bad behavior, with one participant expressing concern that future generations will be intellectually superior to current humans, similar to how humans are vastly superior to ants.
YC Debates: Should We Slow Down AI?
Jacob Coxin argues governments should substantially intervene to slow frontier AI development, citing rapid capability growth (superhuman AI researchers by 2027), alignment risks, and misaligned corporate incentives. Robin Hanson counters that typical technological adoption patterns suggest slower real-world impact, that regulation is dangerous without concrete problems, and that liability-based approaches are preferable to centralized global regulation.
How To Build The Future: Eric Glyman, Co-CEO of Ramp
Eric Glyman, co-CEO of Ramp, discusses how the company evolved from Parabus to build a $60 billion financial operations platform, and how Ramp is leveraging AI agents and internal tools like Inspect, Research, and Glass to transform organizational productivity and reshape the future of work.
Grok Bot: What Changes When You Can Build AI Teammates
Lauren Tan and Roman Ugarte discuss Grokbot, a platform for creating AI teammate agents with their own identities and tool access that can work autonomously. The conversation covers how AI agents are transforming software engineering and non-technical work, with Lauren showcasing her 4,000+ PRs/month workflow using specialized bots in Slack and describing the shift from execution-constrained to execution-abundant organizations.
Grok Bot: What Changes When You Can Build AI Teammates
Lauren Tan and Roman Ugarte discuss Grockbot, a product enabling users to create AI teammates with their own identities and capabilities. They share how AI agents are transforming software engineering and business operations, demonstrating workflows where engineers ship thousands of PRs monthly by leveraging specialized bots integrated into team communication tools.