Large Universe Models in pharmaceutical R&D

What it ingests

Trial registries and protocol amendments, adverse-event databases, competitor filings and conference abstracts, internal assay and screening results, patent filings, regulatory guidance, and publication and preprint feeds in the target area.

The belief that matters

Probability of technical success, held per programme and revised continuously rather than at stage gates.

A stage gate is a scheduled re-examination of a belief that has been drifting since the last one. Between gates, evidence accumulates that nobody formally incorporates: a competitor's Phase II readout implying something about the target class, three internal assays trending the wrong way, a safety signal in a related compound.

What it prevents

The characteristic failure is discovering at a gate what was knowable months earlier. A Large Universe Model that weights each arriving result into a maintained posterior makes the decline visible while the decision to reallocate is still cheap.

Competitive intelligence as a stream

Competitor programmes are visible through registries, conference abstracts, patent filings and hiring patterns. Individually these are noise. Held as evidence bearing on a specific belief — is someone ahead of us on this target — they are a signal that arrives continuously and is almost never tracked continuously.

Limits

Probability of technical success is notoriously badly calibrated across the industry, and a maintained posterior inherits whatever biases go into its weighting. The value is in surfacing movement and its causes, not in the absolute number.