Design systems need more than measurements and structures: they need machine-usable knowledge — how chemistry sets permeability, how peptides sit in their targets, what earlier campaigns did and why. BioMate builds a provenance-linked knowledge layer where every statement is tied to its exact source, and extracts new knowledge with a pipeline rigorous enough to trust.

A knowledge-aware layer for cyclic-peptide design — study overview
Study overview — a provenance-linked knowledge layer turns literature into machine-usable design knowledge, validated de-circularized on fresh papers by an independent five-family committee.

Results in brief

107 · 825
provenance-linked entries · structures spanning 437 targets
0.73 (κ 0.68)
independent five-family faithfulness on held-out fresh papers
0.97
precision at the committee-agreement gate

From text to usable rules

Existing collections give a designer data and structures, but not a design statement — "adding a hydrogen-bond donor and enlarging the ring lower permeability" — linked to its source, its evidence strength and the molecular feature it constrains. BioMate's layer provides exactly that, joining structure, property and prior-campaign knowledge in one schema with provenance on every entry.

Rigor is the hard part

Matching an exact quote proves where a sentence came from, not what it means. Scaling extraction naively produces thousands of records that mostly do not support the statement beside them. BioMate's pipeline pairs a frontier model with explicit, deterministic rules — one measured relation per statement, core developability properties only, no results cited from earlier work, no predictions, no off-scope endpoints — and then a multi-model committee gate for the meaning the rules cannot check.

Measured improvement

Under this rigorous, decompose-then-verify pipeline, statement-level extraction precision reaches a level that makes a scaled knowledge layer dependable rather than noisy — a substantial, measured improvement over a naive first generation, with inter-model agreement to match. Reliability comes from removing the checkable failures first and gating the rest.

Provenance you can audit

Every statement in the layer carries its source identity, exact quotation, evidence tier and the feature it constrains. That is what lets a generative system consume design knowledge as a prior it can trace — and what lets a scientist audit it. BioMate builds knowledge that is usable by a program and accountable to a person.

Read more A knowledge-aware layer for cyclic-peptide design

The knowledge layer is detailed in a companion manuscript (in preparation). Our published cyclic-peptide work — the decoder, the value-of-information scheduler and the benchmark framework — is open-access.

See all BioMate publications →