Alfredo, co-founder of AMA, told Y Combinator that solving cancer with AI is not hyperbole. He said his team builds a personalized research lab around each patient and argued the field is right at the cusp of helping far more people.
He framed cancer as a data problem.
Sequencing now costs less than dinner in the Bay Area, Alfredo said, and sequencing for advanced tumors is covered by insurance including Medicare and Medicaid, according to Alfredo. The Cancer Genome Atlas sequenced only a couple hundred patients per tumor type a decade ago, Alfredo said, while each patient today yields about 20,000 gene-level data points plus copy number, RNA expression, splicing and cell state that hospitals are not set up to mine.
AMA wants to replace what once took a 30 to 40 person genomics team with agents that act as trial, evidence and toxicity experts, Alfredo said. He cited a rare brain tumor that clustered closer to mesenchymal tumors across data sets, which pointed to therapies the literature did not suggest, according to Alfredo, and a second patient where automated reanalysis showed the last three chemotherapies were a poor fit for the tumor biology, according to Alfredo. He pointed to real-time protein design and Google DeepMind co-scientist work from Mountain View as the path to compress years of lab work, while the host cited EVO 2 trained on 9 trillion base pairs, in-context steering with 4 to 64 examples, and heavy investment from Lola to Chai to Isomorphic Labs.
That optimism met a harder record outside the room. No drug designed by AI has won Food and Drug Administration clearance, per Benzinga reporting on Wall Street Journal interviews with drugmaker chiefs, who said AI helps pick targets and patients but cannot resolve biology alone. AstraZeneca chief Pascal Soriot rejected the claim that AI will cure cancer on its own, per the same reporting, which leaves Alfredo's account of patients reaching remission with AI guided care sitting against an industry with zero AI designed approvals to date.
Alfredo did not say who pays for a full time machine lab per patient, what regulators will require to accept an agent's repurposing call, or when a designed protein actually treats a person rather than a benchmark.
