AI drug discovery is no longer hypothetical
AI drug discovery is the use of computational models, including machine learning and physics-based methods, to identify therapeutic targets, design AI-designed drug candidates, and prioritize them for clinical trial validation more quickly and systematically than traditional discovery pipelines. The key shift today is not that algorithms can propose molecules—we have known that for years—but that those molecules and targets are now entering the unforgiving reality of human studies. The emerging story is that AI is moving from clever prediction engines to engines of experimental commitment. If a candidate can win an investigational new drug clearance and face patients in Phase I/II, the technology is no longer an experiment; it has joined the machinery of drug development. That is where the hype will either be confirmed or corrected.
REC-7735: AI-designed chemistry meets the clinic
Recursion’s AI-designed PI3Kα H1047R inhibitor REC-7735 receiving IND clearance and moving into a Phase I/II trial is a line in the sand for machine learning therapeutics. This is not a retrospective success story; it is a live experiment testing whether an AI-built molecule can deliver in patients with PIK3CA H1047R‑mutant solid tumors. The opinion that matters now is the one written in clinical data, not pitch decks. By pushing an AI-designed drug candidate into human testing, Recursion is forcing a higher standard for the entire field: either the compound’s safety and efficacy justify the algorithmic complexity, or AI remains a fancy way to rediscover what medicinal chemists could have found more slowly. The real disruption is not speed alone; it is whether AI can expand chemical space into clinically meaningful territory that conventional workflows would miss.
AI maps and the first validated neuro target
If oncology is where AI chemistry hits the clinic, neuroscience is where AI target discovery is challenging dogma. Recursion and Genentech’s decision to co-develop a program based on the first validated neuro target from an AI-generated map is a bigger deal than its single $3 million (approx. RM13.8 million) milestone might suggest. According to the Biotechnology Innovation Organization, only 8.4% of neurology drug candidates that enter Phase I studies ultimately win approval, an indictment of how limited our target set has been. Recursion’s whole-genome CRISPR knockout map in iPSC-derived neurons, spanning more than 17,000 genes and thousands of small molecules, represents a deliberate rejection of pathway-by-pathway guesswork. The important opinion here is that unbiased, genome-wide AI maps are not academic toys; they are becoming the front door for new neurobiology. The unknown target name is less important than the fact that Genentech accepted the validation package at all.
Closed-loop discovery: Receptor.AI and Sethera shorten the feedback cycle
The most underappreciated shift in AI drug discovery is the move from one-off predictions to closed-loop experimentation. Receptor.AI and Sethera’s partnership is a testbed for that philosophy. Sethera’s polymacrocyclic peptide libraries push into chemical architectures that standard design rarely touches, while Receptor.AI applies physics-based modeling and AI to interpret sequence, architecture, enrichment, and activity data. The point is not to let models dictate the molecules, but to lock discovery into repeat cycles of design, synthesize, test, and learn. This kind of closed-loop workflow makes time, not chemistry, the main variable: how many optimized cycles can you fit before a program hits the clinic? If the approach improves hit confirmation and lead optimization for hard-to-drug targets, it will challenge the old assumption that screening campaigns must be linear, slow, and largely manual.
From structure prediction to clinical trial validation
AI systems for structural prediction, exemplified by tools like AlphaFold 3, are reshaping the earliest decisions about what to test at all. When you can estimate protein structures and complexes at scale, the bottleneck moves from “can we see the target?” to “which targets and modalities deserve the next experiment?” That matters because every target promoted into clinical development carries enormous cost and opportunity risk. In this new model, machine learning therapeutics live upstream and downstream: upstream in selecting and validating targets, downstream in designing molecules like REC-7735, with closed-loop systems tightening the feedback between both. The crucial judgment is that AI should not be treated as a magical shortcut; it is a triage engine. Its real value will be measured in how many dead-end programs are never started and how many once-unthinkable candidates reach the point where patients, not algorithms, deliver the final verdict.
