Is AI Now Good Enough To Cure Cancer? | YC Paper Club
AI for biology is reaching a tipping point: cheap sequencing makes cancer a per-patient data problem, agents act as research labs, and DNA language models are steerable by prompting and design loops.
Watch on YouTube ↗Charlie's summary
- ▶ 10:28
Cancer is now a data problem, so each patient can be treated as their own dataset
Alfredo says sequencing is now cheaper than many Bay Area dinners and insurance covers it for advanced tumors. The Cancer Genome Atlas sequenced only a couple hundred patients per tumor type, so researchers looked for averages. Now each patient has about 20,000 gene data points, plus copy numbers, RNA expression and splicing. The shift is to mine one patient's data rather than generalize across cohorts.
- ▶ 14:06
Agents can replace the 30-person genomics lab with a personalized research lab for every patient
Landmark studies needed algorithm experts, infrastructure staff, geneticists and MDs, which only places like the Broad could afford. Alfredo's company, Omante, treats agents as those bioinformatics experts, running in real time. A discovery made Tuesday can reach a patient by Wednesday morning. Agents specialize in trials, evidence and safety, and toxicity, shrinking a years-long timeline to real time.
- ▶ 17:10
Biology doesn't respect tumor labels, so clustering by pathway can find therapies the literature doesn't
Alfredo's team clustered a rare brain tumor across many datasets and found it resembled mesenchymal tumors, which opened therapies not suggested in the literature. In another patient, tumor evolution analysis showed the last three lines of chemotherapy didn't fit the tumor's biology. Alfredo argues this approach also improves safety, since it favors compounds with known safety profiles and avoids ineffective toxic lines.
- ▶ 36:38
Scaling raw cell data fails; scale AI-guided experimentation in a loop instead
Yousef's team used agents to reprocess public NIH data, going from 36 million to nearly 500 million cells. But the data was mostly immune cells, with T-cells heavily overrepresented, so diversity was poor. Yousef argues for scaling AI-guided experimentation rather than data generation: use models like Claude to choose which data to generate, in a loop. He says this is likely faster to a useful virtual cell.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.