File size: 7,474 Bytes
089d665 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 | """KG sparsification — extract the patient-specific reasoning subgraph.
Inspired by:
- "Knowledge Graph Sparsification for GNN-based Rare Disease Diagnosis"
(arXiv 2510.08655, Oct 2025)
- KARE (ICLR 2025) — KG community retrieval for reasoning
- MedGraphRAG (ACL 2025) — triple-graph for grounded medical QA
Output is a small (~30-200 nodes) graph centered on the patient that can be:
(a) rendered in the front-end with `react-force-graph-3d`
(b) fed to the LLM as a structured triple list ("Patient ─[HAS_PHENOTYPE]→ HP:X ←[ANNOTATES]─ Disease ORPHA:Y")
(c) used to extract narrated paths Patient→...→Disease
"""
from __future__ import annotations
import logging
from typing import Optional
from .types import Subgraph, SubgraphNode, SubgraphEdge
logger = logging.getLogger("gemeo.subgraph")
async def _safe_query(cypher: str, params: dict = None) -> list:
try:
from space_graph import _safe_query as q
return await q(cypher, params or {}, timeout=15.0)
except Exception as e:
logger.debug(f"cypher failed: {e}")
return []
async def extract(
*,
patient_id: str,
hpo_ids: list[str],
gene_symbols: list[str] = None,
target_orpha: str = None,
max_nodes: int = 80,
) -> Subgraph:
"""Extract reasoning subgraph for this patient.
If `target_orpha` is given, extract paths Patient→...→that disease.
Otherwise extract a 1-hop neighborhood centered on the patient's HPOs/genes
and the top diseases that share phenotypes with the patient.
"""
gene_symbols = gene_symbols or []
nodes: dict = {}
edges: list = []
# 1) the patient
pid = f"patient:{patient_id}"
nodes[pid] = SubgraphNode(
id=pid, label="Patient", name="Patient",
weight=1.0, extra={"is_center": True},
)
# 2) phenotypes
for hpo in hpo_ids[:30]:
nid = f"hpo:{hpo}"
# enrich with name
rows = await _safe_query(
"MATCH (p:Phenotype {hpoId: $hpo}) RETURN p.name AS name, p.definition AS def LIMIT 1",
{"hpo": hpo},
)
name = (rows[0]["name"] if rows else hpo)
nodes[nid] = SubgraphNode(id=nid, label="Phenotype", name=name, code=hpo, weight=0.9)
edges.append(SubgraphEdge(source=pid, target=nid, rel="HAS_PHENOTYPE", weight=1.0))
# 3) genes
for sym in gene_symbols[:10]:
nid = f"gene:{sym}"
nodes[nid] = SubgraphNode(id=nid, label="Gene", name=sym, code=sym, weight=0.9)
edges.append(SubgraphEdge(source=pid, target=nid, rel="HAS_GENE_VARIANT", weight=1.0))
# 4) candidate diseases
if target_orpha:
candidate_orphas = [target_orpha]
else:
if hpo_ids:
rows = await _safe_query(
"""
MATCH (p:Phenotype)<-[:HAS_PHENOTYPE]-(d:Disease)
WHERE p.hpoId IN $hpos
WITH d, count(p) AS overlap
ORDER BY overlap DESC
LIMIT 8
RETURN d.orphaCode AS orpha, d.name AS name, overlap
""",
{"hpos": hpo_ids[:30]},
)
candidate_orphas = [r["orpha"] for r in rows if r.get("orpha")]
else:
candidate_orphas = []
for orpha in candidate_orphas[:6]:
rows = await _safe_query(
"""
MATCH (d:Disease {orphaCode: $orpha})
OPTIONAL MATCH (d)-[:HAS_PHENOTYPE]->(p:Phenotype)
WHERE p.hpoId IN $hpos
OPTIONAL MATCH (d)-[:ASSOCIATED_WITH]->(g:Gene)
WHERE g.symbol IN $genes
RETURN d.name AS name,
d.cid10 AS cid10,
collect(DISTINCT p.hpoId) AS shared_hpos,
collect(DISTINCT g.symbol) AS shared_genes
""",
{"orpha": orpha, "hpos": hpo_ids[:30], "genes": gene_symbols[:10]},
)
if not rows:
continue
r = rows[0]
did = f"disease:{orpha}"
nodes[did] = SubgraphNode(
id=did, label="Disease", name=r.get("name") or orpha,
code=orpha, weight=0.95,
extra={"cid10": r.get("cid10")},
)
for hpo in (r.get("shared_hpos") or []):
hid = f"hpo:{hpo}"
if hid in nodes:
edges.append(SubgraphEdge(source=did, target=hid, rel="DISEASE_HAS_PHENOTYPE", weight=0.8))
for sym in (r.get("shared_genes") or []):
gid = f"gene:{sym}"
if gid in nodes:
edges.append(SubgraphEdge(source=did, target=gid, rel="ASSOCIATED_WITH", weight=0.85))
else:
# gene mentioned by disease but not in patient — still informative
gid = f"gene:{sym}"
nodes[gid] = SubgraphNode(id=gid, label="Gene", name=sym, code=sym, weight=0.6)
edges.append(SubgraphEdge(source=did, target=gid, rel="ASSOCIATED_WITH", weight=0.85))
# 5) optional: drugs targeting candidate diseases (1-hop)
if candidate_orphas:
rows = await _safe_query(
"""
MATCH (d:Disease)-[:TREATED_BY|TARGETED_BY]->(drug:Drug)
WHERE d.orphaCode IN $orphas
RETURN d.orphaCode AS orpha, drug.name AS name, drug.rxcui AS rxcui
LIMIT 20
""",
{"orphas": candidate_orphas},
)
for r in rows:
drug_name = r.get("name")
if not drug_name:
continue
did = f"disease:{r['orpha']}"
drug_id = f"drug:{r.get('rxcui') or drug_name}"
nodes[drug_id] = SubgraphNode(
id=drug_id, label="Drug", name=drug_name,
code=r.get("rxcui"), weight=0.7,
)
edges.append(SubgraphEdge(source=did, target=drug_id, rel="TREATED_BY", weight=0.7))
# cap node count by weight
if len(nodes) > max_nodes:
kept = sorted(nodes.values(), key=lambda n: n.weight, reverse=True)[:max_nodes]
kept_ids = {n.id for n in kept}
nodes = {n.id: n for n in kept}
edges = [e for e in edges if e.source in kept_ids and e.target in kept_ids]
# 6) build narrated paths Patient→...→Disease
paths = []
for orpha in candidate_orphas[:3]:
did = f"disease:{orpha}"
if did not in nodes:
continue
steps = []
steps.append({"node": pid, "rel": "is", "label": "Patient"})
# find a shared phenotype
for e in edges:
if e.source == pid and e.rel == "HAS_PHENOTYPE":
hpo_node = nodes.get(e.target)
if not hpo_node:
continue
# is this phenotype linked to the disease?
shared = any(
ee.source == did and ee.target == e.target
for ee in edges
)
if shared:
steps.append({"node": e.target, "rel": "HAS_PHENOTYPE", "label": hpo_node.name})
steps.append({"node": did, "rel": "DISEASE_HAS_PHENOTYPE_REVERSE", "label": nodes[did].name})
break
paths.append({
"target_orpha": orpha,
"target_name": nodes[did].name,
"steps": steps,
})
return Subgraph(
nodes=list(nodes.values()),
edges=edges,
paths=paths,
method="cypher_sparsify",
target_disease=target_orpha,
)
|