AI drug-discovery skeptics have a point: binding simulation accuracy is just 49%, data costs 30x longer

citrini · x · 2026-10-25

Investor citrini lays out the deep divide between "AI people" (scaling-law believers who assume better AI means novel medical discoveries) and "biology people," who argue biology has no bitter lesson: AI/ML can't yet reliably simulate in vivo studies like binding affinities or functional assays, and reliable in silico human toxicology is 5-10 years away. Even the most AI-pilled biologists reject the idea that near-term AI can find novel targets or causal understanding — and call the assumption that even an "ASI" could cure disease from existing knowledge without new biological data "patently false."
Firsthand experiment: AI generated new molecules at stunning speed, but its in silico binding simulation hit only 49% accuracy. Improving it requires feeding back lab data that took 30x longer to produce than generating the molecules — and that was test-tube work; human trials would be far slower.
Human biology adds nth-order interaction effects we lack data to predict; virtual cells cover only a small fraction of possible perturbations. Unique human brain cell identities make animal studies nearly useless for some diseases.
AI can speed up data analysis, trial design, and approvals — but "can't make a 6-month trial complete in 3 months." Coding-style speedups don't map onto a field bottlenecked by generating new biological data.

More from AGI Musings