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

Across biology, AI capabilities are rapidly compounding, from genome language models (EVO2/Manurva) that learn DNA's 'language' well enough to predict mutati

Is AI Now Good Enough to Cure Cancer? A Synthesis of Genome Language Models, Virtual Cells, and Autonomous Research Agents

Abstract

This synthesis examines the convergence of foundation-model methods and biological discovery, drawing on reported results from genome language models, virtual cell modeling, autonomous research agents, and patient-specific oncology pipelines. The central claim is that biology is being reframed as a search problem over an enormous hypothesis space, and that a recursive self-improvement (RSI) loop - sample candidates, select, test, retrain - is now available for any task admitting verifiable rewards, including biological design. Evidence includes EVO 2, a genome language model trained on 9 trillion base pairs that predicts mutation pathogenicity and has been used generatively to design bacteriophage capsids now under patient testing; emergent in-context learning enabling few-shot design from as few as four examples; and autonomous research systems outperforming general coding agents on molecular design tasks. A contrasting negative result - failure to pass an informal Turing test - calibrates these advances against persistent limits in general machine cognition.

1. Introduction

Two developments underpin the current moment in computational biology. First, sequencing costs have fallen sharply, producing biological data at exponential rates; genomic records deposited in NCBI reportedly doubled over a three-year period. Second, self-supervised sequence modeling has proven transferable from natural language to biological polymers, treating DNA as a language - a symbolic encoding of RNA and protein products shaped by natural selection, but one that humans cannot natively read. The combination has produced systems whose outputs are entering clinical workflows: AlphaFold is reported to be used at some stage in every drug now brought to market, and AI-designed phages from the Arc Institute are being tested in patients.

A parallel shift is occurring in oncology. Where population-level studies such as TCGA once sequenced only a few hundred patients per tumor type, full individual multi-omic profiles now make each patient a dataset unto themselves. This has reframed cancer as, in the sponsors' words, "a data problem," creating the conditions for personalized, agent-assisted research pipelines that replicate the function of large academic research teams for a single patient in real time.

This synthesis addresses a single research question: what mechanisms allow AI to accelerate biological discovery, and where do those mechanisms currently fail? Key terminology is defined as follows. A genome language model is a sequence model trained on nucleotide data, typically by next-token prediction. A virtual cell model predicts the behavior of cells, tissues, or organisms rather than molecular sequences alone. An RLVable task is one whose outputs can be automatically verified, permitting reinforcement learning from verifiable rewards. Section 2 establishes theoretical context. Section 3 analyzes four research programs. Section 4 consolidates technical findings, Section 5 discusses implications, and Section 6 concludes.

2. Background and Related Work

Biological discovery is usefully framed as search over a hypothesis space too large for exhaustive experimentation. Structure prediction systems such as AlphaFold and ESM demonstrated the value of this framing by mapping structural landscapes without running physical experiments. The epistemic gap motivating this work has been stated directly: "We understand the structure of the universe better than the workings of living cells. We know the full genome sequence of many species, but we still cannot predict how a cell will behave if we mutate a previously unstudied gene."

Two conceptual frameworks recur throughout the analysis. The Waddington landscape from developmental biology holds that cell states occupy a low-dimensional manifold, a claim that parallels representational-space arguments in machine learning, including the embedding geometry learned by models such as UCE. The RSI loop, imported from automated ML research, supplies the procedural template: generate, verify, retrain. Its cost profile differs sharply between in-silico verification, which is cheap and fast, and physical (wet-lab) verification, which remains slow and expensive - a constraint that shapes which biological problems are currently tractable for closed-loop automation. Benchmark maturity enables hill-climbing: CASP and the Virtual Cell Challenge provide measurement infrastructure for structure and cell-state prediction, while PostTrainBench performs an analogous role for autonomous research agents.

3. Core Analysis

3.1 Genome Language Models and the EVO 2 Architecture

EVO 2 is trained with next-token prediction on nucleotide sequences, using a hybrid attention-and-convolution architecture to support context windows up to 1 million base pairs, at 7B and 40B dense parameter scales. The training corpus was curated down to 9 trillion base pairs by retaining one representative genome per species and removing near-duplicate sequences, then trained in two phases - first on short, gene-dense conserved regions, then with extended context. This curation step proved critical: omitting annotation-focused data augmentation and instead sampling 50% of the genome at random reduced Spearman correlation on a BRCA1 loss-of-function benchmark by 0.16, a margin the sponsors characterize as the difference between state-of-the-art and bottom-of-pack performance. Despite training on only a single human reference genome, EVO2's likelihood outputs correlate with mutation pathogenicity competitively with supervised, human-specific models. Interpretability work using sparse autoencoders identified a single learned feature associated with phage biology, firing across prophages, CRISPR arrays, and repurposed phage genes - including previously unannotated elements - suggesting the model has internalized a biologically coherent abstraction rather than merely memorizing sequence statistics. A successor bidirectional model, Manurva, addresses EVO2's weakness in modeling co-evolutionary interactions between distant genome regions, enabling discovery of previously unknown non-coding RNAs and a novel CRISPR-associated reverse transcriptase (UG27), subsequently validated in the laboratory.

3.2 Emergent In-Context Learning for Sequence Design

Traditional biological sequence design required large labeled datasets - EVO1, for instance, was fine-tuned on tens of thousands of CRISPR-Cas loci. EVO2 instead exhibits emergent in-context learning (ICL) from pure next-token pretraining on nucleotides, permitting few-shot design without weight updates. Demonstrated cases include generating novel rare-earth-binding proteins from roughly fourteen example lanmodulin sequences, and generating CRISPR-Cas9 systems from only four in-context loci. A hypothesized mechanism attributes this capability to the long-tail, bursty structure of pretraining data combined with induction-head-like circuits analogous to those documented in language models. In one validated case, a reinforcement-learning post-training step (GRPO) using a sequence-to-function predictor as reward improved EVO2's in-context capability for cell-type-specific enhancer design in mouse brain tissue, with additional reward terms constraining sequence entropy and GC content to prevent reward hacking.

3.3 Virtual Cells and Autonomous Research Agents

The Universal Cell Embedding (UCE) model, trained on 35 million cells via masked gene-expression prediction, produces structure matching known cell-type biology without explicit supervision - an unsupervised recovery of the Waddington manifold. Agent-driven reprocessing later scaled this corpus to roughly 500 million cells but exposed severe class imbalance, with immune and T-cell populations overrepresented by orders of magnitude, motivating a shift from linear data-scaling toward loop-based, AI-guided experimentation. Separately, the Locus system led the PostTrainBench leaderboard for autonomous post-training and, when redirected toward molecular property prediction and generative sampling, outperformed general-purpose coding agents such as BoltzGen and Codex as compute scaled - evidence that research-specific harness engineering, not model scale alone, differentiates agent performance on biological tasks.

3.4 Patient-Specific Oncology Pipelines

Full individual genomic and multi-omic profiling, increasingly covered by insurance for advanced tumors, allows AI agents to replicate functions once requiring 20-30-person research teams, producing data packets that inform - rather than replace - clinician judgment. Reported cases include reclassifying a rare brain tumor's subtype via clustering analysis and identifying that three prior chemotherapy lines were not supported by observed clonal evolution.

4. Technical Insights

Several implementation patterns recur across the systems reviewed. Data curation dominates raw scale: EVO2's annotation-focused augmentation outperformed naive random sampling by a 0.16 Spearman-correlation margin, indicating that representation quality, not corpus size, bounds genomic model performance. Architecture choices track domain constraints: hybrid attention-convolution designs extend context to the million-base-pair range needed for gene-scale dependencies. In-context learning generalizes across modalities when pretraining exhibits long-tail structure, enabling few-shot design with as few as four examples - a substantial reduction in labeled-data requirements relative to fine-tuning. Reward-hacking mitigation (entropy and GC-content constraints) is necessary whenever RL post-training is layered on generative sequence models. Finally, agent performance on scientific tasks depends more on research-specific harness engineering than on swapping underlying LLMs, as evidenced by Locus outperforming general coding agents on identical model backbones.

5. Discussion

These findings suggest that biological AI is approaching a regime where in-silico verification loops can be closed cheaply, while physical verification remains the binding constraint on throughput - an asymmetry likely to persist as long as wet-lab assays and organoid culture remain resource-intensive and bottlenecked by practical issues such as hospital tissue-handling protocols. The gap between structural and functional understanding is also notable: clinical efficacy failures are attributed primarily to poor target identification rather than poor molecule design, implying that virtual-cell and target-discovery research may yield larger near-term returns than further binder optimization.

The informal Turing test result, in which an AI persona achieved only roughly 30% confusion against a 70% threshold, cautions against extrapolating narrow biological competence to general machine cognition. Domain-specific sequence models can internalize structured biological abstractions, as the sparse-autoencoder phage feature suggests, without implying broader conversational or reasoning parity with humans.

6. Conclusion

The evidence reviewed indicates that genome language models, virtual cell embeddings, and autonomous research agents are jointly compressing the biological discovery cycle, with concrete clinical artifacts - AI-designed phage capsids under patient testing, clinician-facing personalized data packets - already in use. The practical takeaway for practitioners is that data curation and harness engineering, rather than raw model or dataset scale, currently determine returns on investment in this domain. Future work should prioritize closing the physical-verification bottleneck through lab-in-the-loop design and improved organoid pipelines, while maintaining realistic expectations about generalization beyond narrow, well-benchmarked biological tasks.

Sources

• Is AI Now Good Enough To Cure Cancer? | YC Paper Club - Original Creator (YouTube)
• Analysis and summary by Sean Weldon using AI-assisted research tools

About the Author

Sean Weldon is an AI engineer and systems architect specializing in autonomous systems, agentic workflows, and applied machine learning. He builds production AI systems that automate complex business operations.

LinkedIn | Website | GitHub