Siloam Therapeutics MCP Server

Every atom you see on the right is being placed by a generative model conditioned on a live binding pocket. Rotate. Explore. This is one iteration of thousands drug.AI runs per hour — every one of them reviewable by a human chemist.

Tools

• search_diseases

  Search PrimeKG disease nodes by case-insensitive substring. Use this to browse/disambiguate disease names. You usually do NOT need to call it first: rank_causal_drivers / get_subgraph_summary / analyze_disease all resolve free-text names themselves via a curated alias map (e.g. "kidney cancer" -> "renal cell carcinoma (disease)", "bladder cancer" -> "urinary bladder cancer") plus an exact/substring fallback. Reach for this tool when a name is ambiguous or you want to see

• list_subtypes

  List subtype-divergent diseases and their available subtype labels. Some diseases have driver sets that diverge sharply by subtype, so a single pan-disease ranking blurs them (e.g. ERBB2 only tops breast cancer under HER2-enriched). A disease may expose more than one *taxonomy* — a named way of partitioning it. Breast cancer, for example, is registered under both a molecular taxonomy ("PAM50 (molecular)": Luminal A/B, HER2-enriched, Basal-like) and a clinical taxonomy ("

• get_subgraph_summary

  Extract the N-hop gene/pathway subgraph around a disease and return its summary (node/edge counts, node-type breakdown, sample relationships). Args: disease_name: disease name or substring (resolved against the KG). hops: traversal depth (default 2).

• rank_causal_drivers

  Rank causal driver genes/pathways for a disease. Pipeline: extract the disease subgraph, then score nodes with the chosen ranking mode and a global-degree hub penalty. Args: disease_name: disease name or substring. hops: subgraph traversal depth (default 2). top_k: number of ranked drivers to return (max 20 recommended). mode: "evidence_rrf" (DEFAULT, recommended for pan-disease) ranks genes by Reciprocal Rank Fusion of the Open Targets

• fetch_opentargets_evidence

  Fetch disease->target association scores from the Open Targets Platform. Returns a {gene_symbol: score} map (scores in [0,1]) plus the resolved EFO id. No API key required. Args: disease_name: free-text disease name. datatype: "somatic_mutation" (cancers), "genetic_association" (GWAS), or "" for the overall aggregated score. size: number of associated targets to pull.

• build_synthesis_prompt

  Rank drivers, then return the structural-biology synthesis prompt(s) for the *calling* LLM to complete (rationale, publications, small-molecule role, confidence). The LLM step stays client-side so the server is model-agnostic. Args mirror rank_causal_drivers (including `subtype` for the opt-in subtype-divergent branch — see rank_causal_drivers / list_subtypes). Returns `prompts` (one per batch) and the `drivers` table they were built from. After the client fills each pro

• analyze_disease

  One-call orchestrator that ranks causal drivers for ANY disease, using the subtype-divergent branch when one exists and a pan-disease ranking otherwise. Chains the existing tools so a client doesn't have to: 1. Look up the disease's registered subtypes (across all taxonomies — molecular and clinical/IHC). 2a. If subtypes exist -> for *each* label run the subtype-resolved ranking AND build its structural-biology synthesis prompt(s), so every subtype

• analyze_disease_agentic

  Agentic (LangGraph) causal-driver ranking. Runs the orchestrated DAG: intake → ground → route → strategize → probe → rank → critic ⇄ (acquire/revise) → synthesize → report. Unlike the fixed-pipeline tools, this *decides* per disease: pan vs subtype-divergent routing, which evidence channels to use (and whether to fuse them, via a complementarity probe), and self-corrects against hub-artifact and buried-watchlist controls. Args: disease_name: disease name or