Project Description
Focus Areas: Computational Biology, Data Science, Agentic AI, Machine Learning
Personnel:
Advisor/Principal Investigator: Md Amiruzzaman, West Chester University
CS Research Assistant (Agentic AI & ML Focus): Huy Nguyen, West Chester University
Problem Statement
Current computational toxicology research suffers from two primary structural limitations:
Lack of Hierarchical Structural Modeling: Traditional approaches rely on relational SQL tables or flat vector databases. However, biological toxicity data is intrinsically hierarchical and directional (flowing logically from chemical structures to bio-assays, molecular targets, and organ-level toxicity outcomes). Despite this natural alignment, there is a distinct lack of computational models leveraging "Directed Acyclic Graphs (DAGs)" to capture these nested dependencies.
Data Siloing and Fragmentation: Most existing toxicity prediction tools and retrieval systems focus exclusively on isolated, small-scale datasets, restricting their generalizability and scope across multi-endpoint toxicity screening.
Without an automated structure that maps compound-to-endpoint dependencies alongside an intelligent retrieval mechanism, computational biologists spend disproportionate effort on data wrangling rather than downstream analysis.
Project Description & Technical Methodology
This project proposes a hybrid framework combining "Directed Acyclic Graphs (DAGs)" for structured data flow and dependency modeling with an "Agentic AI Framework" for dynamic information retrieval.
Hierarchical DAG Modeling and Multi-Dataset Ingestion: We model toxicological entity relationships as an acyclic dependency graph G = (V, E) to preserve biological causality (for example, Compound -> Assay Protocol -> Target Receptor -> Toxicity Endpoint). We aggregate full release dumps from the ChEMBL database (https://ftp.ebi.ac.uk/pub/databases/chembl/ChEMBLdb/latest/) and cross-link external benchmark parameters from TOXRIC (https://toxric.bioinforai.tech/home) using canonical SMILES strings and InChIKeys.
Agentic AI Retrieval Engine: We deploy an autonomous AI agent equipped with dynamic tool-use capabilities. Given an open-ended natural language query (for example, "Retrieve all compounds demonstrating hepatic cytotoxicity below 10 uM with target affinity for enzyme X"), the agent deconstructs search constraints, traverses directed paths across multi-hop dependencies without expensive SQL joins, and aggregates consolidated toxicity profiles rapidly.
Timeline & Execution Strategy
Session: Phase 1 (Sept – Dec)
Focus Area: DAG Design & Core Retrieval Pipeline
Key Milestones & Activities:
Parse and ingest primary raw SQLite/SDF datasets directly from the ChEMBL database (https://ftp.ebi.ac.uk/pub/databases/chembl/ChEMBLdb/latest/).
Construct the core DAG schema mapping chemical structures to toxicity endpoints.
Build and evaluate the initial Agentic AI retrieval toolset on ChEMBL data.
Session: Phase 2 (Jan – Mar)
Focus Area: Dataset Integration & Open-Source Release
Key Milestones & Activities:
Expand and enrich the DAG model by integrating multi-source benchmark toxicity datasets from TOXRIC (https://toxric.bioinforai.tech/home).
Benchmark retrieval speed, accuracy, and latency against standard relational and graph queries.
Finalize code, documentation, and release the complete framework publicly.
Expected Research Outcomes & Deliverables
Open-Source Agentic Toxicity Retrieval Framework: A publicly hosted repository containing the Python-based agentic pipeline, toolset, and DAG schema.
Graph-Structured Toxicity Schema: A standardized, reusable DAG structure mapping primary data from ChEMBL (https://ftp.ebi.ac.uk/pub/databases/chembl/ChEMBLdb/latest/) alongside benchmark toxicity parameters.
Multi-Source Integrated Benchmark Dataset: A consolidated knowledge layer combining raw ChEMBL bioassays with curated endpoints from TOXRIC (https://toxric.bioinforai.tech/home).
Benchmarking & Performance Report: Quantitative evaluation demonstrating retrieval accuracy, latency improvements, and multi-hop query success rates.