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

Main Ideas

1. Opening Experiment: Can People Tell AI from Humans?

François opened with a live chat experiment inspired by Alan Turing’s prediction that computers might convincingly imitate people. Attendees chatted with either another person or an AI system and tried to identify which was which.

The results were not presented as a definitive test, but François noted that some human participants were mistaken for AI—roughly a third, by his account. His takeaway was that AI is getting close to convincing people in some short interactions, though the experiment did not establish that the Turing test has been passed.

2. Why AI for Biology Now?

François argued that biology is becoming a particularly promising area for AI because:

• Biological data is growing rapidly. Falling sequencing costs mean researchers can collect far more genomic information than before.
• Models are improving. He cited large DNA models, AlphaFold’s growing use in drug development, improved AI protein-binder results, and AI-designed bacteriophages beginning to reach patients.
• AI can help prioritize experiments. When predictions can be checked computationally or in the lab, researchers can use a repeated cycle: generate hypotheses, select promising candidates, test them, incorporate the results, and repeat. Computational checks can be fast; physical experiments are slower and more expensive, but still fit the same general process.
• The field needs more than bigger datasets. The speakers emphasized models and workflows that help researchers decide what data to generate and which experiments are most informative.

François also announced that this was his final YC Paper Club before leaving YC to start another project.

3. Alfredo Gonzalez: Cancer as a Patient-Specific Data Problem

Alfredo Gonzalez argued that cancer care can increasingly be approached as a problem of collecting and interpreting data for each individual patient.

• Move beyond averages. Earlier cancer-genomics studies often had to look for common patterns across relatively small groups of patients. With more data available, it may be possible to analyze a person’s tumor as its own rich dataset rather than relying only on broad tumor categories.
• Use many layers of tumor information. A patient-level picture can include genomic changes, copy-number changes, RNA expression, alternative splicing, cell states, and the patient’s treatment history.
• Apply AI agents to the analysis workflow. Gonzalez described the prospect of automating work that previously required a large team of bioinformaticians, infrastructure specialists, geneticists, and clinicians. The goal is to assemble and interpret relevant information quickly enough to inform decisions in real time.
• Look for biology that crosses diagnostic categories. A rare tumor may resemble another tumor type biologically, even if it arises in a different organ. Gonzalez described a rare brain tumor whose molecular patterns appeared closer to mesenchymal tumors, which helped identify potential treatment directions that would not have been obvious from its standard classification.
• Track tumor evolution and treatment response. Re-examining earlier samples can help reveal why prior therapies may not have worked and guide later decisions.

Gonzalez stressed that this is not a claim that AI has already cured cancer. He described personalized analysis as a way to improve treatment decisions and outcomes, especially when standard options have failed or evidence is scarce.

Practical Considerations and Limitations

• Organoids could support patient-specific testing and experiments, but their use is constrained by how hospitals handle tissue. Samples preserved in formalin generally cannot be used to grow organoids.
• Clinicians remain responsible for treatment decisions. AI-generated analyses need to be delivered in a form clinicians can interpret and use.
• Safety is central. For existing drugs, known safety profiles can inform choices; genuinely new treatments would need established clinical-research safeguards, such as dose escalation.
• Claims that a treatment can reverse malignancy need clear definitions and patient-outcome evidence.

4. Yusuf Roohani: Virtual Cells and Autonomous Research

Yusuf Roohani described a “virtual cell” as an AI model of biological systems at different levels—from cells and tissues to whole organisms. The aim is to predict how biological systems behave and identify interventions that could shift disease states toward healthier ones.

• Why virtual-cell models matter: Many drug candidates fail because researchers cannot reliably predict what will work in the human body. Models of molecules have advanced, but predicting relevant cellular and disease outcomes remains difficult.
• Learn representations of cell states: One approach is to train models on gene-expression data, treating cells as collections of gene measurements. In one example, a model trained on about 35 million cells learned representations that grouped known cell types together, despite not being given those labels directly.
• More data alone is not enough: Agents were used to collect hundreds of millions of cell records from a public repository. However, the data were heavily skewed toward immune cells, particularly T cells. Scaling up a dataset can therefore increase its size without providing the breadth needed for a useful general model.
• Use models to guide data collection: Roohani proposed an iterative alternative: use AI to decide which experiments and measurements would best answer a specific biological question, then use their results to improve the models.
• Transfer methods from AI research automation: His team’s system, called Locus in the subtitles, was used to automate model research and then applied to scientific tasks such as molecular-property prediction and selecting candidates from model-generated possibilities. The reported results were computational and still require laboratory validation.
• Account for noisy evaluation: Biological experiments can vary even under similar conditions. Research systems therefore need to handle small validation sets, noisy measurements, expensive experiments, and long-running tasks.

The broader lesson was to build a loop linking AI models with physical laboratories, so experiments are selected for their usefulness rather than collected indiscriminately.

5. Garyk Brixi: What DNA Language Models Learn

Garyk Brixi discussed DNA models that learn by predicting the next nucleotide, using a training approach analogous to next-token prediction in language models.

• DNA is difficult to model. It has a small alphabet but very long sequences, repeated and redundant regions, and biological patterns operating at multiple scales—from single mutations to genes spanning large stretches of DNA.
• Data selection matters. The Evo 2 model was trained on a curated set of roughly 9 trillion base pairs. The team selected one representative genome per species and removed sequences that were too similar to others.
• Training emphasized informative regions. The team prioritized genes and other evolutionarily conserved elements, then increased the model’s context length up to about one million base pairs.
• Use predictions to assess mutations. If a mutation makes a sequence less likely under the model, that may correspond to a damaging change. Brixi showed an example involving BRCA1 and said Evo 2 performed competitively with specialized models on several mutation-prediction benchmarks.
• Inspect what the model has learned. Sparse autoencoders were used to find internal features associated with biological concepts. One feature was active in multiple phage-related regions, including some that had not been annotated as prophages by previous methods.
• Generate and combine sequences. DNA generated by the model can be assessed with other tools, such as functional classifiers. Brixi cited work on designing bacteriophages and noted that AI-designed phages are beginning to be used in treatment contexts.
• Use a different model for interactions across DNA regions. Minerva was presented as a model intended to better capture interactions between separate sequence regions. It helped identify candidate noncoding RNAs and other genomic systems, including a reverse-transcriptase system that the team confirmed in the lab.

Brixi cautioned that DNA-only models have limitations, including difficulties with some long-distance enhancer relationships and the absence of many kinds of population or cellular data. His account emphasized that models can reveal useful patterns without making DNA directly readable to humans.

6. Divya Nori: Few-Shot Prompting for Biological Design

Divya Nori described using in-context learning (ICL) to guide a genomic model with a small set of examples, instead of conducting extensive task-specific retraining.

• Provide examples in the prompt. A model receives a few biological sequences with a shared property and is asked to generate new sequences with similar properties.
• Apply the approach across scales. Nori cited examples involving rare-earth-binding proteins, antibody design, CRISPR-Cas systems, and whole bacteriophage genomes.
• Use few-shot prompting when labeled data is scarce. Biological tasks often lack the large, well-labeled datasets that conventional task-specific training would require.
• Combine prompting with reinforcement learning. For enhancer design, the team used a sequence-function prediction model as a reward signal and updated Evo 2 using GRPO. Additional reward terms constrained properties such as DNA entropy and GC content.
• Test generated designs experimentally. An enhancer generated with the approach was tested in mouse brain tissue and showed activity and specificity in the reported experiment.

Nori emphasized that the model’s ability to learn from examples appeared during general genomic pretraining; it was not explicitly trained to perform in-context learning. She suggested that this capability may arise from patterns in genomic data and mechanisms similar to those that support ICL in language models.

The longer-term vision is a cycle in which researchers prompt the model, test promising designs in the lab, and use the results to refine subsequent designs.

Speakers and Sources Featured

Presenters

• François — YC Paper Club host and moderator; led the opening chat experiment and introduced the AI-for-biology theme.
• Alfredo Gonzalez — Cancer bioinformatics researcher and company co-founder; spoke about patient-specific cancer analysis.
• Yusuf Roohani — Spoke about virtual cells, autonomous research, and AI-guided experimental design.
• Garyk Brixi — Discussed Evo 2, Minerva, and genomic language models.
• Divya Nori — Discussed few-shot prompting and biological sequence design.

Other Participants or Sources Mentioned

• Audience members asked questions during the talks.
• Greg from ARC-AGI was mentioned by François as the person who prompted the opening discussion about the Turing test; he was not presented as a speaker.
• The talks referenced research and tools including AlphaFold, Evo 2, Minerva, Locus, the Cancer Genome Atlas (TCGA), and GRPO.