AI Didn't Just Help Big Pharma Beat Small Biotech — It Crowded Everyone Into the Same Targets
A few months ago, the expected story was top pharma plus AI squeezing out small players. Conversations with several US pharma BD teams point to a different, messier reality: the number of active projects they're evaluating has multiplied, while homogeneity has gotten strikingly worse — most of what crosses their desk is still PD-1s and monoclonal antibodies chasing heavily overlapping indications, since protein/antibody design happens to be exactly where AI gains the most ground and patent protection is comparatively weak.
Why AI made the crowding worse, not better
AI's precision in molecular design and preclinical research cuts both ways. Internal testing showed AI can cut pipeline experimental cost by roughly 70% for hot targets — but only where the target and its pathway already have deep public data behind them.
Agentic Co-scientist Benchora doing scaffold-hopping:
A self-reinforcing loop, and where it leads
That creates a strange loop: hot targets have more data → AI works better on them → more teams pile in → even more data gets generated → AI gets stronger there still. Everyone Fast-Following off the same public data can genuinely reach IND quickly, but the result is dozens of companies racing down the identical track, and products that struggle to sell once they're actually on the market — EGFR is the clearest example, with tens of thousands of public data points and AI predictions accurate to the point of near-perfect overlap, virtually guaranteeing a brutal commercial fight among near-identical entrants.
The hot/cold target divide
Oncology drug development already accounts for roughly 40% of global pipelines, heavily concentrated in kinase inhibitors (TKIs) — the point where experienced medicinal chemists can practically recite pocket structures from memory. Genuinely cold targets, by contrast, rarely surface in public view at all: once PDB and UniProt have only a handful of entries for a target, AI's predictive power drops close to zero. Training a reliable potency model needs at least 300 or so high-quality data points — and by the time a small or midsize biotech can afford to generate 300 data points just for training, it's already spent enough to have pushed an actual drug candidate into Phase 1. The only real path past this is exploiting homologous-target knowledge or using physics-based simulation to generate synthetic training data through iterative cycles.
RiemannMol: reshaping molecular space around potency
One approach QuantaBricks has been developing for the cold-target data problem is RiemannMol, which re-endows molecular latent space with a potency-aware metric rather than treating it as a plain Euclidean embedding. In the raw latent space, high-pIC50 molecules are scattered with no clear structure, but under RiemannMol's reshaped 128-dimensional metric space, those same high-activity molecules concentrate into a single connected region — turning "find the active compounds" into a much more tractable local-search problem instead of a needle-in-a-haystack one.
The commercial reality AI can't touch
None of this crowding and duplication actually benefits patients. Every company's competition still dies at the IND gate, FDA approval throughput is capped regardless, and payer budgets are a real, external constraint — no amount of AI-driven IND speed changes a one-year regulatory delay, a multi-year patient recruitment slog, or a drug nobody can actually sell once approved.
Five practical moves for breaking out of the crowd
(1) Return to first commercial principles before committing technical resources; (2) don't pivot blindly — domain depth is the real moat; (3) run a genuine "1+n" strategy — a lower-regulatory-risk parallel track keeps cash flow alive when the lead program stalls; (4) treat IND as a milestone, not an exit — model Phase 2 success probability before IND; (5) add an AI layer now — cost reduction means more shots on goal from the same budget.
Feel free to schedule a meeting with me if you have any specific questions/requests: https://calendly.com/yizhouma/chemorchestra-user-session
Join our community to learn more: https://discord.gg/u3Mb6C2aHv
Benchora with RiemannMol: https://benchora.quanta-bricks.com/
ChemOrchestra: https://www.quantabricks.xyz/