Multimodal AI Agent Framework for Integrative Pharmacology and Toxicology Evaluation

Yi Ni^*, Liwei Zhu, Shengwei Wang, Chenghao Liu, Debin Zou

Bio LIMS INC, a digital intelligence company for life science labs, Massachusetts, United States

^*Corresponding author: yi.ni@biolims.net

AMP 2025

Date — November, 2025

Location — Boston

Abstract

Traditional pharmacology and toxicology evaluation faces fundamental limits when dealing with massive heterogeneous data, complex biological mechanisms, and individual variability — only 1 in 5,000 to 1 in 10,000 candidate compounds reaches approval, and roughly 60% of failures trace back to inadequate evaluation at the pharmacology and toxicology stages. We present a multimodal AI Agent framework that addresses these bottlenecks through six specialized agents collaborating across document analysis, content extraction, image understanding, knowledge association, quality control, and report generation. The system unifies text, tabular, and image modalities, links extracted entities into a Neo4j drug–target–disease knowledge graph, and produces GLP-compliant safety evaluation reports with full audit trails. Across toxicity data extraction, structural alert identification, and entity relationship extraction, the framework outperforms traditional methods by 52–77% in accuracy and F1, processes literature roughly seven times faster than manual review, and reduces expert intervention workload by 68%. A hierarchical verification mechanism spanning data, model, and application layers addresses hallucination, explainability, and regulatory compliance challenges.

Six-agent collaboration architecture

Document Analysis Agent

Identifies data types — PDF reports, scanned records, images — and triggers parallel processing.

Content Extraction Agent

Extracts IC50, LD50, and experimental conditions with multi-language support.

Image Understanding Agent

ViT-based gel electrophoresis quantification (CV < 5%) and SMILES encoding generation.

Knowledge Association Agent

Builds drug–target–experiment networks; links CAS numbers to the knowledge base.

Quality Control Agent

Data consistency verification, outlier detection, manual-review flagging.

Report Generation Agent

Produces GLP-compliant safety evaluation reports with automated visualizations.

Backing storage: Neo4j (drug–target–disease graphs), PostgreSQL (experimental records and PK parameters), Elasticsearch (full-text and image vectors).

Headline performance

IC50 extraction accuracy

94.3%

vs. traditional baseline

Toxicity prediction accuracy

+31%

vs. traditional QSAR

Decision-making speed

+54%

faster

Processing time

−47%

unstructured data

Manual intervention

−68%

expert review workload

Literature review throughput

~7×

3,000 papers in 2 days

Detailed comparison vs. traditional methods

Task	Metric	AI Agent	Traditional	Improvement
Toxicity data extraction	LD50 accuracy	94.3%	62.1%	+52%
Structural alert ID	F1 score	87.6%	55.3%	+58%
Entity relationship extraction	F1 score	83.5%	47.2%	+77%

Hierarchical verification mechanism

Data layer. Multi-source cross-validation across PubMed literature, the CTD database, and internal experimental records to ground each extracted fact in independent sources.

Model layer. Domain knowledge constraints from chemical rules and toxicology standards bound model outputs and prevent contradictions with scientific consensus.

Application layer. Triple expert cross-validation on key results, Neo4j-backed reasoning path visualization for transparency, and evidence-level grading (certainty / possibility / speculation) on every claim.

Regulatory and compliance posture

The framework is designed against current FDA guidance on AI model traceability and explainability, EMA expectations around expert review of AI-assisted outputs, and ICH directions for global GLP harmonization. Standard operating procedures, Neo4j-based audit trails for full lifecycle traceability, hash-based desensitization of patient and experimental data, and role-based access control round out the compliance surface.

Keywords

Multimodal AI AgentPharmacologyToxicology evaluationLarge language modelsKnowledge graphGLP complianceDrug safetyNeo4j

How to cite

Ni Y, Zhu L, Wang S, Liu C, Zou D. Multimodal AI Agent framework for integrative pharmacology and toxicology evaluation. Poster presented at: [AMP 2025 — full conference name, date, location to be confirmed]; 2025.