Spaces:
Sleeping
Sleeping
Upload agent/toxicologist_agent.py with huggingface_hub
Browse files- agent/toxicologist_agent.py +85 -0
agent/toxicologist_agent.py
ADDED
|
@@ -0,0 +1,85 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import json
|
| 2 |
+
import os
|
| 3 |
+
from groq import Groq
|
| 4 |
+
from pydantic import BaseModel, Field, TypeAdapter
|
| 5 |
+
from typing import List, Optional
|
| 6 |
+
from config import GROQ_API_KEY, GROQ_MODEL
|
| 7 |
+
|
| 8 |
+
|
| 9 |
+
class ToxicologyReport(BaseModel):
|
| 10 |
+
target_assay: str = Field(description="The target assay endpoint evaluated.")
|
| 11 |
+
risk_level: str = Field(description="Overall risk level: Low, Moderate, High, or Critical.")
|
| 12 |
+
toxicophore_assessment: str = Field(description="Analysis of which structural features contribute to predicted toxicity.")
|
| 13 |
+
biochemical_mechanism: str = Field(description="Explanation of the biochemical mechanism linking structure to endpoint.")
|
| 14 |
+
structural_alerts_found: List[dict] = Field(description="List of structural alerts matched and their relevance.")
|
| 15 |
+
bioisosteric_replacements: List[str] = Field(description="Suggested safer bioisosteric replacements.")
|
| 16 |
+
confidence: str = Field(description="Confidence in this assessment based on model prediction and alert matching.")
|
| 17 |
+
|
| 18 |
+
|
| 19 |
+
class ToxicologistAgent:
|
| 20 |
+
def __init__(self):
|
| 21 |
+
self.client = Groq(api_key=GROQ_API_KEY)
|
| 22 |
+
self.model = GROQ_MODEL
|
| 23 |
+
self.schema_adapter = TypeAdapter(ToxicologyReport)
|
| 24 |
+
self.json_schema = self._build_strict_schema()
|
| 25 |
+
|
| 26 |
+
def _build_strict_schema(self) -> dict:
|
| 27 |
+
raw = self.schema_adapter.json_schema()
|
| 28 |
+
raw["additionalProperties"] = False
|
| 29 |
+
if "properties" in raw:
|
| 30 |
+
for prop_val in raw["properties"].values():
|
| 31 |
+
if isinstance(prop_val, dict) and "properties" in prop_val:
|
| 32 |
+
prop_val["additionalProperties"] = False
|
| 33 |
+
if "$defs" in raw:
|
| 34 |
+
for def_val in raw["$defs"].values():
|
| 35 |
+
if isinstance(def_val, dict):
|
| 36 |
+
def_val["additionalProperties"] = False
|
| 37 |
+
return raw
|
| 38 |
+
|
| 39 |
+
def generate_report(self, smiles: str, target_assay: str, predictions: list,
|
| 40 |
+
shap_attributions: list, high_attr_indices: list,
|
| 41 |
+
alerts_found: list) -> ToxicologyReport:
|
| 42 |
+
prompt = f"""You are an expert computational toxicologist. Analyze the following compound's toxicological profile.
|
| 43 |
+
|
| 44 |
+
COMPOUND SMILES: {smiles}
|
| 45 |
+
TARGET ASSAY: {target_assay} ({predictions[0].get('target_class', 'Unknown')})
|
| 46 |
+
PREDICTED PROBABILITY: {predictions[0].get('probability', 0.5):.4f}
|
| 47 |
+
PREDICTED CLASS: {predictions[0].get('predicted_class', 'Unknown')}
|
| 48 |
+
|
| 49 |
+
SHAP HIGH-ATTRIBUTION ATOM INDICES: {high_attr_indices}
|
| 50 |
+
FULL SHAP VECTOR: {shap_attributions}
|
| 51 |
+
|
| 52 |
+
STRUCTURAL ALERTS DETECTED: {json.dumps(alerts_found, indent=2)}
|
| 53 |
+
|
| 54 |
+
Based on this data:
|
| 55 |
+
1. Identify which specific structural fragments (toxicophores) are driving the {predictions[0].get('predicted_class', 'Unknown')} prediction for {target_assay}.
|
| 56 |
+
2. Explain the biochemical mechanism linking these fragments to the assay endpoint.
|
| 57 |
+
3. Suggest 2-3 specific bioisosteric replacements that could mitigate toxicity while preserving core scaffold.
|
| 58 |
+
4. Assess overall risk level.
|
| 59 |
+
5. Provide confidence in the assessment.
|
| 60 |
+
|
| 61 |
+
Output a structured report adhering to the provided schema.
|
| 62 |
+
"""
|
| 63 |
+
chat_completion = self.client.chat.completions.create(
|
| 64 |
+
model=self.model,
|
| 65 |
+
messages=[
|
| 66 |
+
{
|
| 67 |
+
"role": "system",
|
| 68 |
+
"content": "You are a professional computational toxicologist. Output reports strictly adhering to the provided JSON schema."
|
| 69 |
+
},
|
| 70 |
+
{"role": "user", "content": prompt}
|
| 71 |
+
],
|
| 72 |
+
response_format={
|
| 73 |
+
"type": "json_schema",
|
| 74 |
+
"json_schema": {
|
| 75 |
+
"name": "Toxicology_Report",
|
| 76 |
+
"strict": True,
|
| 77 |
+
"schema": self.json_schema
|
| 78 |
+
}
|
| 79 |
+
},
|
| 80 |
+
temperature=0.1,
|
| 81 |
+
)
|
| 82 |
+
|
| 83 |
+
raw_json = chat_completion.choices[0].message.content
|
| 84 |
+
report = self.schema_adapter.validate_json(raw_json)
|
| 85 |
+
return report
|