ToxD4C Documentation

ToxD4C is an advanced deep learning framework for molecular toxicity prediction. This framework innovatively integrates Graph Neural Networks, Transformer architecture, geometric information processing, and chemical prior knowledge to provide accurate and reliable toxicity prediction capabilities for drug discovery and chemical safety assessment.

• Multi-task toxicity prediction: Simultaneously predict 31 different toxicity endpoints (26 classification + 5 regression tasks) • Multi-modal information fusion: Integrate 2D graph structure, 3D geometric information, molecular fingerprints and chemical descriptors • Uncertainty quantification: Provide confidence estimates for each prediction • Enhanced interpretability: Provide model interpretability through attention mechanisms and hierarchical representation learning

The ToxD4C framework consists of four core components: 1. Multi-Modal Encoder Core - GNN-Transformer hybrid architecture - Dynamic fusion module with adaptive weight learning - Cross-attention mechanism for feature enhancement 2. Geometric Information Processing - SE(3) equivariant layers for 3D molecular structure - Distance-aware message passing - Geometric-topological dual encoder 3. Hierarchical Representation Learning - Four-level hierarchy: Atom → Functional Group → Scaffold → Molecule - Multi-scale GCN architecture with different receptive fields - Chemical feature encoding at multiple levels 4. Multi-task Prediction Architecture - Task-specific heads for 31 toxicity endpoints - Uncertainty quantification with Bayesian inference - Contrastive learning for enhanced representation quality

Multi-Modal Deep Fusion: • Integrates four complementary molecular representation modalities • Dynamic weight generation based on molecular features • Cross-attention mechanism for deep information exchange Hierarchical Representation Learning: • Four-level hierarchical architecture mimicking chemist cognition • Multi-scale receptive fields (2/4/8-layer GCN) • Complete modeling from microscopic to macroscopic Intelligent Uncertainty Quantification: • Bayesian deep learning integration • Aleatoric and epistemic uncertainty modeling • Calibrated confidence intervals for risk assessment End-to-End Multi-Task Learning: • Simultaneous prediction of 31 toxicity endpoints • Shared representation learning and task knowledge transfer • Unified toxicity prediction platform

Classification Tasks (26): • Carcinogenicity, Ames Mutagenicity, Cardiotoxicity • CYP Inhibition, Hepatotoxicity, Nephrotoxicity • Neurotoxicity, Skin Sensitization, Eye Irritation • Respiratory Toxicity, Reproductive Toxicity, Developmental Toxicity • Endocrine Disruption, Immunotoxicity, Genotoxicity • Hematotoxicity, Plasma Protein Binding, BBB Penetration • P-gp Substrate, hERG Blocking, Nuclear Receptor Activation • Stress Response Pathway, DNA Damage, Cell Cycle Toxicity • Mitochondrial Toxicity, Oxidative Stress Regression Tasks (5): • Acute Oral Toxicity LD50 • Aquatic Toxicity LC50 • Bioconcentration Factor BCF • Soil Adsorption Coefficient Koc • Octanol-Water Partition Coefficient LogP

Computational Efficiency: • Total parameters: ~50M • Hidden dimension: 512 • Attention heads: 8 • Maximum sequence length: 512 • Parallel computation design with GPU optimization Prediction Accuracy: • Classification tasks average AUC: 0.85-0.92 • Regression tasks average R²: 0.75-0.88 • Calibration error (ECE): < 0.05 • Uncertainty correlation coefficient: > 0.80 Interpretability Analysis: • GAT attention weights show important chemical bonds and atoms • Transformer attention reveals long-range molecular interactions • Cross-attention reveals information fusion between modalities • Feature importance analysis across hierarchical levels

Drug Discovery: • Early toxicity screening before synthesis • Lead compound optimization guidance • ADMET prediction integration Chemical Safety Assessment: • New chemical registration support (REACH compliance) • Environmental risk assessment • Occupational health protection Regulatory Science: • Computational toxicology advancement • Risk assessment modernization • International chemical safety standardization Technical Innovation Value: • First systematic multi-modal molecular toxicity prediction framework • Original combination of SE(3) equivariant processing + hierarchical learning • First deep application of contrastive learning in molecular toxicology