mcp/orchestrator.py

#!/usr/bin/env python3
"""
MedGenesis – dual-LLM orchestrator (v5)
---------------------------------------
• No external 'pytrials' dependency.
• Uses direct HTTP for clinical trials.
• Clean async fan-out, dual-LLM support.
"""

from __future__ import annotations
import asyncio, itertools, logging
from typing import Dict, Any, List, Tuple

from mcp.arxiv           import fetch_arxiv
from mcp.pubmed          import fetch_pubmed
from mcp.ncbi            import search_gene, get_mesh_definition
from mcp.mygene          import fetch_gene_info
from mcp.ensembl         import fetch_ensembl
from mcp.opentargets     import fetch_ot
from mcp.umls            import lookup_umls
from mcp.openfda         import fetch_drug_safety
from mcp.disgenet        import disease_to_genes
from mcp.clinicaltrials  import fetch_clinical_trials
from mcp.cbio            import fetch_cbio
from mcp.openai_utils    import ai_summarize, ai_qa
from mcp.gemini          import gemini_summarize, gemini_qa

log = logging.getLogger(__name__)
_DEFAULT_LLM = "openai"


def _llm_router(engine: str = _DEFAULT_LLM) -> Tuple:
    """Choose summarization and QA functions based on engine name."""
    if engine.lower() == "gemini":
        return gemini_summarize, gemini_qa, "gemini"
    return ai_summarize, ai_qa, "openai"


async def _safe_gather(*tasks, return_exceptions: bool = False):
    """
    Await multiple coroutines, log any exceptions, and optionally return them.
    """
    results = await asyncio.gather(*tasks, return_exceptions=True)
    cleaned: List[Any] = []
    for r in results:
        if isinstance(r, Exception):
            log.warning("Task failed: %s", r)
            if return_exceptions:
                cleaned.append(r)
        else:
            cleaned.append(r)
    return cleaned


async def _gene_enrichment(keys: List[str]) -> Dict[str, Any]:
    """
    Fan-out gene-related endpoints for each seed keyword:
      - NCBI gene lookup
      - MeSH definition
      - MyGene.info
      - Ensembl cross-refs
      - OpenTargets associations
    Returns a dict of results.
    """
    jobs: List[asyncio.Task] = []
    for k in keys:
        jobs.extend([
            asyncio.create_task(search_gene(k)),
            asyncio.create_task(get_mesh_definition(k)),
            asyncio.create_task(fetch_gene_info(k)),
            asyncio.create_task(fetch_ensembl(k)),
            asyncio.create_task(fetch_ot(k)),
        ])
    results = await _safe_gather(*jobs, return_exceptions=True)

    def bucket(idx: int) -> List[Any]:
        return [res for i, res in enumerate(results) if i % 5 == idx and not isinstance(res, Exception)]

    return {
        "ncbi": bucket(0),
        "mesh": bucket(1),
        "mygene": bucket(2),
        "ensembl": bucket(3),
        "ot_assoc": bucket(4),
    }


async def orchestrate_search(query: str, llm: str = _DEFAULT_LLM) -> Dict[str, Any]:
    """
    Main entry point. Performs:
      1. Literature fetch (PubMed + arXiv)
      2. Keyword seed extraction
      3. Bio-enrichment (UMLS, OpenFDA, gene services)
      4. Clinical trials lookup
      5. cBioPortal variants
      6. AI LLM summary
    Returns a unified dict for the UI.
    """
    # 1) Literature
    pubmed_t = asyncio.create_task(fetch_pubmed(query, max_results=7))
    arxiv_t  = asyncio.create_task(fetch_arxiv(query, max_results=7))
    papers_raw = await _safe_gather(pubmed_t, arxiv_t)
    papers = list(itertools.chain.from_iterable(papers_raw))[:30]

    # 2) Seed keywords
    seeds = {
        w.strip()
        for p in papers
        for w in p.get("summary", "")[:500].split()
        if w.isalpha()
    }
    seeds = list(seeds)[:10]

    # 3) Bio-enrichment fan-out
    umls_tasks = [asyncio.create_task(lookup_umls(k)) for k in seeds]
    fda_tasks  = [asyncio.create_task(fetch_drug_safety(k)) for k in seeds]
    gene_task  = asyncio.create_task(_gene_enrichment(seeds))
    trials_t   = asyncio.create_task(fetch_clinical_trials(query, max_studies=10))
    cbio_t     = asyncio.create_task(
        fetch_cbio(seeds[0]) if seeds else asyncio.sleep(0, result=[])
    )

    umls_list, fda_list, gene_data, trials, variants = await asyncio.gather(
        _safe_gather(*umls_tasks, return_exceptions=True),
        _safe_gather(*fda_tasks, return_exceptions=True),
        gene_task,
        trials_t,
        cbio_t,
    )

    # 4) Deduplicate gene symbols from enrichment
    genes = {
        g["symbol"]
        for src in (gene_data["ncbi"], gene_data["mygene"], gene_data["ensembl"], gene_data["ot_assoc"])
        for g in src if isinstance(g, dict) and g.get("symbol")
    }
    genes = list(genes)

    # 5) Deduplicate variants by genomic coordinates
    seen: set = set()
    unique_vars: List[dict] = []
    for v in variants or []:
        key = (
            v.get("chromosome"),
            v.get("startPosition"),
            v.get("referenceAllele"),
            v.get("variantAllele"),
        )
        if key not in seen:
            seen.add(key)
            unique_vars.append(v)

    # 6) LLM-driven summary
    summarize_fn, _, engine_used = _llm_router(llm)
    combined = " ".join(p.get("summary", "") for p in papers)
    ai_summary = await summarize_fn(combined[:12000])

    return {
        "papers": papers,
        "umls": [u for u in umls_list if not isinstance(u, Exception)],
        "drug_safety": list(
            itertools.chain.from_iterable(dfa for dfa in fda_list if isinstance(dfa, list))
        ),
        "clinical_trials": trials or [],
        "variants": unique_vars,
        "genes": gene_data["ncbi"] + gene_data["ensembl"] + gene_data["mygene"],
        "mesh_defs": gene_data["mesh"],
        "gene_disease": gene_data["ot_assoc"],
        "ai_summary": ai_summary,
        "llm_used": engine_used,
    }


async def answer_ai_question(question: str, context: str, llm: str = _DEFAULT_LLM) -> Dict[str, str]:
    """
    Follow-up QA: uses the chosen LLM’s QA function.
    """
    _, qa_fn, _ = _llm_router(llm)
    prompt = f"Q: {question}\nContext: {context}\nA:"
    try:
        answer = await qa_fn(prompt)
    except Exception as e:
        answer = f"LLM error: {e}"
    return {"answer": answer}