From Cancer Data to Designed DNA: What AI Biology Can Actually Demonstrate

1. Guest Background

The question in this YC Paper Club episode’s title—whether AI is now good enough to cure cancer—brings together several research problems that operate at different distances from patient care. Across roughly 82 minutes, the Y Combinator discussion moves among personalized cancer analysis, cellular representations, autonomous research systems, and biological sequence design. It is a meeting of complementary research perspectives, rather than a presentation of one treatment or a single clinical result.

Alfredo, a co-founder of Amanta, approaches the subject through cancer bioinformatics. His background includes doctoral research at UCLA. He describes building computational pipelines for processing tumor and normal genomes at scale, working with rare sarcoma data, and contributing to early cancer-cell CRISPR screen validation at the Broad Institute. That experience explains his emphasis on the practical distance between obtaining molecular data and making it useful for an individual patient.

Yusuf, Entology’s head of science, previously established the machine learning team at the Arc Institute. His contribution connects models of cell state with systems that automate scientific research. Garrick, a Stanford doctoral student who previously led the Evo2 team and now participates in Minerva, explains what genomic foundation models can learn and how researchers interrogate them. Divya studies generative AI for biological design and examines how a handful of examples can steer sequence generation. Their respective roles matter because a patient evidence package, a cellular embedding, and a designed sequence answer different questions, even when they belong to the same broader research agenda.

2. What the Episode Covers

Alfredo’s starting point is that an individual patient can become the organizing unit for a substantial research effort. He contrasts older searches for common patterns across tumor cohorts with analysis that combines roughly 20,000 genes, copy number, RNA expression, alternative splicing, cell states, and patient background. His assertion that earlier academic work used about 0.1% of the data is a speaker’s characterization, not an independently established measurement in the episode. The underlying argument is that generating information and having the infrastructure to exploit it are different achievements.

He proposes an agent-based workflow spanning algorithms, computing infrastructure, variant calling, genetic interpretation, visualization, and medical interpretation. Additional agents would examine trials, evidence, safety, and toxicity. The ambition is to incorporate relevant discoveries rapidly, subject to their being feasible and applicable to the patient. This makes the research process continuous: new evidence can be reconsidered against an existing patient dataset instead of waiting for an entirely new analysis effort.

Two patient stories illustrate the approach. In one, a rare brain tumor clustered more closely with mesenchymal tumors across the datasets the team assembled. Alfredo says pathways, surface markers, and disease drivers then suggested therapies. In another, retrospective analysis of initial samples helped reconstruct tumor evolution, leading his team to judge three previous treatment lines poorly matched to the underlying biology and change direction. Neither account supplies the drugs, response criteria, follow-up, or comparison needed to quantify benefit.

Yusuf broadens the discussion from molecular information to cellular behavior. Structure prediction and target binding can help prioritize experiments, he argues, but researchers still need to identify targets whose manipulation changes a disease phenotype. His Universal Cell Embedding example uses masked prediction over a cell-by-gene transcript-count matrix. He reports that self-supervised training on approximately 35 million cells produced structures consistent with known broad and finer cell types without supplying those labels. A subsequent month-long agent effort reprocessed NIH transcriptomic data and expanded the collection beyond 500 million cells, only to reveal heavy immune-cell representation, particularly T cells.

Locus addresses another part of this search problem. Yusuf describes transferring an autonomous system developed for language-model post-training research into molecular design. He reports improvements with additional research compute on the presented binding-efficiency prediction and latent-space sampling tasks, while explicitly identifying the results as in silico.

Garrick then explains the foundation beneath genomic applications. According to his presentation, Evo2 combines attention and convolution, trains on a nine-trillion-base-pair dataset selected using species representatives and similarity filtering, and emphasizes annotated regions. Training progresses from shorter contexts to one million base pairs, with seven-billion- and forty-billion-parameter dense models. These choices describe how the model is built; the episode’s subsequent examples determine what scientific uses those capabilities can support.

3. Core Views: Reasoning, Examples, and Limits

The most consequential distinction in the episode is between making a hypothesis more plausible and demonstrating that an intervention works. Alfredo’s brain-tumor example shows why crossing conventional tumor categories could be valuable: molecular pathways may provide useful comparisons where a rare diagnostic category offers little guidance. Yet proximity to mesenchymal tumors in an analysis does not establish that a therapy effective in another setting will help this patient. Similarly, his team’s retrospective judgment about three earlier treatment lines does not independently establish what would have happened under an alternative course. The cases reveal a reasoning method, while leaving its clinical effect insufficiently specified.

Yusuf’s argument about targets explains this gap. Designing a molecule that binds a selected target and selecting a target that changes disease are connected tasks, but success at the first cannot settle the second. Cellular models are intended to help bridge that distance. However, embeddings that recover recognizable cell types first demonstrate useful representation learning. Predicting how an intervention moves a cell from one state to another requires further evidence. The episode does not show that recovering existing cellular structure automatically supplies that capability.

The data examples also challenge a simple equation between volume and understanding. Expanding a collection beyond 500 million cells can reinforce already common categories if immune cells dominate. Garrick’s BRCA1 example makes the complementary point that composition can materially affect performance: he reports that retaining a 50% whole-genome random-sampling component would lower Spearman correlation by 0.16 on the displayed loss-of-function benchmark. That is a benchmark-specific claim, not a universal prescription. Selecting one representative genome per species also trades reduced redundancy for missing within-species variation, while leaving substantial population, cellular, and multi-omic information outside Evo2.

Locus demonstrates why resource accounting and attribution deserve equal attention. The comparisons use the same sandbox and GPUs, and the horizontal axis measures experimental GPU hours. Token costs are absent from that plot. Yusuf says token quantities were approximately similar in most comparisons and acknowledges that both contribute to cost. The evidence therefore supports a bounded argument about compute use on the presented computational tasks, rather than a complete accounting of discovery economics. Moreover, the team emphasizes harness engineering and relatively little weight fine-tuning. Its disclosed results cannot be assigned exclusively to an unspecified reinforcement-learning recipe.

Interpretability offers a different route to scientific value. Garrick describes an Evo2 phage-associated feature that responds to prophages, CRISPR arrays containing phage DNA, and genes bacteria have repurposed from phages. The interesting result is a relationship across distinct sequences and biological contexts. Minerva extends discovery through bidirectional modeling of distant interactions. A laboratory-validated UG27 reverse-transcriptase system, with a CRISPR-like RNA array and a DNA-to-RNA-to-DNA process producing small hairpins, provides a concrete example. Its function remains unknown. Discovering a system and explaining what it does are separate accomplishments.

Divya’s few-shot examples similarly require careful labels. She describes Evo2’s in-context learning as emerging from nucleotide prediction pretraining, with demonstrations spanning roughly 100 to 100,000 nucleotides. Fourteen natural lanmodulin examples guide generation of proteins predicted to belong to that family. Four CRISPR-Cas loci guide generation of systems containing the array, tracrRNA, and Cas9. The phage-genome example emphasizes predicted gene architecture. These are useful forms of design control under scarce supervision; they do not establish experimental functionality for every generated candidate.

The enhancer work advances further into physical evidence. GRPO updates Evo2 using a sequence-to-function predictor as the reward source, with entropy and GC-content constraints intended to reduce reward hacking. AlphaGenome is mentioned as an example of that predictor category, not explicitly identified as the implemented model. Divya reports that one generated enhancer showed activity and specificity through fluorescence in mouse brain. Without sample counts, quantitative effects, controls, or statistical detail, the presentation cannot establish broad reliability. Her explanations for ICL—long-tailed data, burstiness, parallel genomic structures, and an induction-head-like mechanism—also remain hypotheses or preliminary analysis. Asked about previously unseen combinations of cell specificity, she offers an account involving prior knowledge and motif recombination, rather than direct validation.

These distinctions give substance to the proposed experimental loop. The host imagines generating candidates, selecting them, testing efficacy, and retraining. Yusuf argues for choosing experiments around scientific questions and stresses noisy evaluation. Such a loop can expose and correct model errors, but its value depends on what the evaluator measures and what physical verification costs. The host’s confidence that it will work is a forecast; each application still has to earn that conclusion.

4. Learning and Application

A clinical research team can apply the episode’s ideas by treating the patient analysis as an evidence package that a physician can interrogate. In Alfredo’s account, this is particularly relevant after first- and second-line treatment failures. A useful package would connect a molecular finding to a treatment hypothesis, explain the supporting evidence, and retain the uncertainty. Cross-tumor similarity should remain visible as a rationale to investigate. The physician retains the final treatment decision, and assembling this package does not resolve the separate challenge of establishing an N-of-1 trial.

For evaluating patient stories, begin with the missing outcome information: which drugs were used, when treatment changed, how response was defined, how long follow-up lasted, and what adverse events occurred. These questions make it possible to distinguish a promising analytical approach from reproducible patient benefit. Alfredo’s separation of existing therapies from newly designed treatments is also operationally important. Existing safety information can inform review of the former; the latter still require arrangements such as dose escalation. His request for explicit definitions and extended patient outcomes when discussing reversal of malignancy provides a useful standard for assessing similarly ambitious claims.

Organoid projects need coordination before tissue processing forecloses the intended experiment. Alfredo identifies formalin fixation and hospital handling practices as barriers to subsequent culture. A team planning perturbation screens should therefore establish sample requirements early. Organoids can expand scarce tissue and help compare hypotheses, but their differences from the primary tumor should be recorded when interpreting results. An experimental response in that system remains one input into a patient-level decision.

For model developers, audit coverage against the actual scientific question. A cell project should inspect relevant cell types and tissues rather than report only aggregate counts. A genomic project should distinguish functional-region coverage, within-species variation, and missing multi-omic information. Garrick’s BRCA1 comparison makes data mixture a worthwhile experimental variable, but its reported improvement should be tested on the intended task before becoming a general rule. A composition that improves one benchmark may still leave another application poorly served.

Research-agent evaluations should separately record experimental GPU hours, token costs, evaluation conditions, and physical validation status. Locus makes clear why a compute-efficiency result is not automatically a total-cost result. Yusuf’s emphasis on long tasks, progress tracking, idea ranking, and experiment cost suggests concrete harness improvements to examine before assuming weight updates are necessary. Recording these changes separately also helps identify where a gain actually came from.

For sequence design, Divya’s proposed prompt-pool loop offers a practical research structure: start with natural sequences or candidates from an existing generator, expand the pool with Evo2, test selected designs, and update the prompts using the results. If GRPO post-training is added, distinguish the weight update from the examples used for a particular generation task. Treat the sequence-to-function predictor as a proxy, and keep its scores separate from measured activity. Entropy and GC-content constraints may restrict some undesirable optimization, but do not certify function. The mouse-brain enhancer example shows why physical testing adds a different kind of evidence. The next experiment should address the remaining scientific uncertainty, rather than merely reward another increase in the same prediction score.

Source

• Original episode: Is AI Now Good Enough To Cure Cancer? | YC Paper Club