Why fluent answers are not enough for biotech decisions

A fluent answer can be scientifically literate and still be useless at an investment committee. The failure is easy to miss because the prose sounds finished. What has usually disappeared is the machinery of the decision: which evidence was accepted, how much authority it carried, which assumption it changed, and whether the change was large enough to alter the action.

Evidence is not a bag of facts

A human genetic association, an animal model, an assay result, a competitor press release and a registry update are not interchangeable observations. They answer different questions and fail in different ways. The distinction has economic weight. A 2024 analysis of 29,476 target and indication pairs found that drug mechanisms with human genetic support had a 2.6-fold higher probability of clinical success, with important variation by disease area and development phase (Minikel et al.). Clinical success rates also change with indication, phase, sponsor and biomarker strategy (Wong, Siah and Lo). A system that compresses those differences into “strong evidence” has removed information the decision maker needs.

ARiDA keeps evidence families separate. Target genetics, pathway biology, translational evidence, modality readiness, clinical precedent and safety can contribute to a target case, but absence does not become a neutral score. Missing positive evidence reduces coverage and therefore reduces the computed result. Contradictory or adverse evidence is subtractive. A high-severity, source-backed safety liability is treated as a veto. It can force a stop or pass outcome even when the average looks attractive.

Science must be allowed to move the model

The usual handoff between research and finance is a meeting. Scientists discuss the evidence, somebody chooses a probability of success, and the number enters a spreadsheet with little of its history attached. ARiDA makes that handoff computational.

For an asset built from a reference-class benchmark, governed molecular descriptors can produce a conservative druggability modifier. The adjusted transition probability then flows through the same stage-gated valuation and Monte Carlo rNPV calculations as every other assumption. ARiDA deliberately leaves analyst-authored probabilities untouched. The code changes only benchmark-derived stages and only when the molecular record has both source references and ledger identifiers. No usable science means no adjustment.

At portfolio level, molecular similarity can inform a correlation prior. It is not treated as proof that two programs will fail together. It is an explicit, reviewable assumption about shared exposure. The matrix must cover every asset, align to the portfolio order and pass a positive-semidefinite check before it can change covariance, diversification or allocation. Partial chemistry coverage is a no-op.

The real standard

The question for an enterprise buyer is not whether a model can answer a scientific prompt. It is whether a changed assay, safety finding or target-validation result can travel through an approved method and produce a different decision without losing its provenance.

That requires retrieval, scientific computation, decision mathematics and review state to remain connected. The recommendation is only the visible end. The product is the recoverable chain beneath it.