Mosaic: a knowledge-graph + MCP-tool layer for pre-clinical R&D intelligence

Pre-clinical R&D triage requires joining evidence across six disconnected resources — ChEMBL for compound activity, PubMed for literature, the patent offices for competitive position, AlphaFold and the PDB for structure, DepMap for genetic dependency, and STRING/Reactome for pathway context. None of these alone answers the question that governs a portfolio decision: which targets are simultaneously biologically supported, under-developed, and tractable for the modality the team can execute?

A generalist LLM with web search can summarise the surface of those sources, but it cannot return the structured scaffold every candidate needs to be useful — per-row evidence counts, an explicit ranking criterion, and source identifiers an analyst can audit. Mosaic is the join layer: a curated knowledge graph exposed as MCP tools so the agent traverses typed relations across all six sources at once. The output is rows with provenance, not paragraphs that need re-verification.

Pre-clinical R&D intelligence is fragmented across PubMed, ChEMBL, patent offices, AlphaFold, and DepMap. A generalist LLM can summarise the surface of these sources, but the question that drives a portfolio decision — which targets are biologically supported, under-developed, and tractable for a chosen modality — requires typed traversal across all of them at once, and a prose answer cannot carry the structured scaffold such a decision needs.

We present Mosaic, an open Model-Context-Protocol (MCP) tool surface over a curated knowledge graph (Postgres + Apache AGE + pgvector) shipping with an oncology starter corpus chosen to validate the architecture on one tractable therapy area before widening: 363 targets, 12,756 compounds, 35,430 papers, 14,305 patents, and roughly 736K typed edges. The schema and pipeline are therapy-area-agnostic; cardiovascular and neurology expansion are in flight.

The tools are the product; the web UI is one delivery surface among several. Methods include zero-shot GLiREL relation extraction (precision evaluated on a stratified human-reviewed sample), Foldseek structural similarity over the AlphaFold corpus, and a heuristic SMILES modality classifier. On twelve first-order benchmark questions the MCP-armed agent cites roughly four times as many concrete, mechanically valid identifiers as a no-tool baseline with zero invalid IDs in either arm; a date-aware LLM judge gives a 33% win-rate (~58% after verified judge-error correction).

Across five case studies, Mosaic's advantage is consistently a structured-output one — per-row evidence, an explicit ranking criterion, and auditable identifiers — that holds even where a web-armed frontier LLM recovers comparable prose recall. The production knowledge graph is released as a versioned public data archive, and the live MCP server and reproducibility package are publicly available; the production source code is held privately and made available to peer reviewers on request.

The full pipeline is reproducible on a fresh VM with only Docker installed, via a separate public repository (mosaic-reproducibility). Docker Compose pulls a versioned hosted image carrying the schema, query layer, and the run_repro driver — no local build step. A hand-curated fixture mini-KG (10 oncology targets, 24 compounds, 30 landmark papers) demonstrates architecture reproducibility end-to-end; the driver applies the schema, loads fixtures, refreshes kg_metadata, and emits representative tool outputs into ./output/MANIFEST.md.

Full-data reproducibility runs against the public archive (mosaic-data, tagged paper-v1, Zenodo DOI 10.5281/zenodo.20529685): point DATABASE_URL at a local Postgres loaded from the published dumps to re-derive the case studies, or query the live MCP server at mcp.getmosaic.dev with a free-tier Bearer token to verify any number in the paper without standing up infrastructure.