Is AI Now Good Enough To Cure Cancer? | YC Paper Club

Event Overview & Turing Test Experiment

• Francois hosted the 10th YC AI for Bio Club, his final event before leaving YC to launch a new venture, having built a community with over 1 million views across nine prior sessions.
• A live Turing test was conducted using a custom website "vibe coded" in 12 hours by Claude Code, pitting human participants against AI agents and other humans.
  • Participants: Astra, Claude, GLM, Gemini, Grok, and Kimi were tested.
  • Rankings: Astra ranked first, followed by Claude, GLM, then Gemini, with Grok and Kimi tied.
  • Human Accuracy: Participants correctly identified AI with high confidence, estimating their success rate at over 70% (exceeding Turing's 1950 prediction of 70% chance for the interrogator to fail identification).
  • False Negatives: 33% of humans were mistaken for AI, attributed to the speakers' natural communication style.
  • Conclusion: The host concluded the Turing test is not yet solved, noting that even with persona injection (a 32-year-old male context), AI identification rates remained high.

Macro Trends in AI for Biology

• Data Scale: The Evo 2 model was trained on 9 trillion base pairs, with the next iteration planned for larger scale; genomic data volume is doubling approximately every three years (NCBI).
• Cost Trajectory: Sequencing costs have plummeted to the point where individual genome sequencing is cheaper than a Bay Area dinner, with insurance coverage (Medicare/Medicaid) now including tumor sequencing for advanced and recurring cases.
• Drug Discovery Adoption: AlphaFold is now utilized at some stage in every drug brought to market; 10 to 100% success rates have been observed for AI-designed protein binders on the BindCraft benchmark.
• Clinical Validation: AI-developed phages are entering human trials via Radical Numerics, a company founded by lab member Eric Nagoyan.
• Foundation Models: Significant investments are flowing into companies like Lilla, Chai, and Isomorphic; the community has developed RSI (Reward-Model-Scale-Inference) procedures to verify and iterate on RLVR-able tasks in both in silico and physical domains.

Speaker Presentations

Alfredo (Co-founder, Amanta): Personalized Cancer Labs

• Paradigm Shift: The speaker argues cancer research has shifted from an average-based approach (TCGA style) to a "data problem" where each patient (20,000+ gene data points) is treated as an independent, high-dimensional dataset.
• Amanta's Approach: Amanta replaces traditional 30-40 person research teams with AI agents acting as bioinformatics experts, infrastructure managers, and geneticists to mine patient data in real-time.
• Case Study 1 (Rare Tumors): A patient with a rare brain tumor was reclassified via clustering as a mesenchymal tumor, allowing the identification of therapies previously unavailable for that specific diagnosis.
• Case Study 2 (Treatment Correction): Analysis of a patient's clonal evolution revealed that three prior lines of cytotoxic chemotherapy were ineffective; AI analysis identified a more suitable therapy, correcting the clinical course.
• Future Vision: The goal is to compress the drug discovery timeline from years to real-time, including de novo protein design and delivery mechanism generation for specific patients.
• Self-Improving Models: Cited Google DeepMind's "self-improving agents" as a model for self-improving biology, where models save thousands of research hours by automating the hypothesis generation loop.
• Q&A Highlights:
  • Organoids: Scientifically feasible for personalized labs, but limited by hospital infrastructure (formalin fixation prevents organoid culture).
  • Safety: Safety is ensured by leveraging known safety profiles of existing compounds or following dose escalation protocols similar to clinical trials when designing new therapies.
  • Clinical Adoption: Clinicians are receptive to AI outputs because they lack the capacity to perform real-time analysis themselves; the AI serves as a decision-support tool for complex cases (e.g., post-second-line failure).

Yusuf Rouhani (Head of Science, Entology): Virtual Cells & Autonomous Research

• Definition: "Virtual cells" aim to model biological systems from the cell level up to the organism to predict human phenotypes, addressing the current gap in translating molecular design to clinical efficacy.
• Model Development:
  • Universal Cell Embedding: Trained on 35 million cells using self-supervised learning (masking genes) to create a Waddington landscape representation of cell states.
  • Data Scaling: Agents reprocessed NIH transcriptomic data to scale from 36 million to 500 million cells, though the dataset was found to be heavily skewed toward immune cells (specifically T cells).
• Strategy Shift: The field is moving from "scaling data generation" to "scaling AI-guided experimentation" (active learning), where LLMs optimize the data generation process rather than passively consuming all available data.
• Entology's LOCUS System:
  • Applied an autonomous AI research system (optimized for LLM post-training) to molecular design and property prediction.
  • Performance: The system outperformed standard coding agents (Claude Code, Codex) and generative models (Boltzgen) in predicting molecular binding efficiency and generating sequences as compute scaled.
  • Robustness: Demonstrated ability to handle noisy evaluation metrics and small validation sets common in biological experimentation.
• Q&A Highlights:
  • RL vs. Harness: The team prioritizes "harness engineering" (prompting/context tuning) over weight fine-tuning for generalization, though parameter updates are used when fast retrieval is required.
  • Compute Metrics: Performance was measured in GPU compute hours rather than token costs, showing superior compute efficiency compared to frontier labs' off-the-shelf agents.

Garrick (Lead, Evo 2/Minerva): Genomic Language Models

• Evo 2 Architecture:
  • A hybrid attention-convolution model trained on 9 trillion base pairs (one representative per species) with a focus on genic regions and evolutionary conservation.
  • Trained in two phases: short context (genic regions) followed by extended context (up to 1 million base pairs) to learn long-range interactions.
  • Capabilities: Functions as a "decision model" where likelihood scores correlate with disease-causing mutations and loss-of-function variants; outperformed supervised models on BRCA1 benchmarks despite no specific human training.
• Interpretability: Sparse autoencoders identified a "phage-associated" feature, revealing unannotated prophages, CRISPR arrays, and repurposed phage genes within E. coli.
• Minerva: A bidirectional model designed to solve Evo 2's weakness in predicting distal interactions (e.g., base pairing), successfully identifying non-coding RNAs and a novel UG27 reverse transcriptase validated in the lab.
• Q&A Highlights:
  • Tokenization: Tokenization at the nucleotide level was preferred for mutation prediction accuracy, though byte-level tokenization is a future target.
  • Human Specificity: Training on one genome per organism avoids redundancy; human-specific variants are predicted via conservation signals across primates, though multi-omic data is needed for fine-grained human-specific insights.

Divya (HeLab): Few-Shot Prompting & ICL

• Paradigm Shift: Introduced a shift from task-specific fine-tuning to few-shot In-Context Learning (ICL) for biological sequence design, enabling design with as few as 4-14 examples.
• Mechanism: Evo 2 exhibits emergent ICL capabilities from pre-training, allowing it to infer tasks and recombine motifs without weight updates.
• Key Applications Validated:
  • Proteins: Designed novel lanthanide-binding proteins from 14 known examples.
  • Antibodies: Generated hits from structure-based model outputs without explicit structural conditioning.
  • CRISPR-Cas: Designed functional Cas9 systems (array, tracrRNA, protein) from 4 examples.
  • Phage Genomes: Generated whole phage genomes with correct architecture at a fraction of the compute cost of fine-tuning.
• Post-Training with RL: Applied RL (using GRPO and a sequence-to-function reward model) to post-train Evo 2, significantly improving few-shot performance for enhancer design.
  • Validation: Designed enhancers active in the mouse brain were synthesized and validated via fluorescence.
• Future Direction: Proposes a "lab-in-the-loop" design process where AI prompts generate candidates, lab testing provides feedback to update prompts, and inference-time techniques (harness engineering) are adapted for biology.
• Q&A Highlights:
  • ICL Origin: Attributed to long-tail power laws and "burstiness" in genomic data distribution, similar to mechanisms in language models.
  • Generalization: Models rely on a mix of pre-trained motifs and context-specific recombination; enhancer specificity patterns are largely driven by the prompt examples due to their sparsity in the pre-training distribution.