Graph Convolutional Network for Lysosomal Protein Binding Affinity Prediction

Overview

This solution applies graph convolutional network (GCN) architectures to predict small molecule binding affinity for lysosome-specific proteins. By training on general protein-ligand binding data and fine-tuning for lysosomal targets, the platform enables rapid identification of potential small molecule agonists and binders. The system integrates with BioBricks.ai, a FAIR-compliant data registry that aggregates billions of rows from over 40 life science databases, and Aspis4J for cheminformatic model distribution. The approach is designed to accelerate early-stage drug discovery for lysosome-related diseases by prioritizing candidate molecules through computational prediction rather than expensive wet-lab screening.

The platform serves pharmaceutical researchers, toxicologists, and drug discovery teams seeking to identify novel therapeutic compounds targeting lysosomal proteins. By combining daily-updated binding data, GCN-based prediction models, and an active learning feedback loop, the solution reduces the time and cost associated with traditional high-throughput screening approaches.

Technical specifications

Core capabilities:

• Graph convolutional neural networks trained on large-scale public binding affinity data from ChEMBL, BindingDB, and other public/private datasets
• Fine-tuned models specifically adapted for lysosomal protein targets
• Integration with BioBricks.ai data registry providing daily-updated curated protein binding datasets
• Multiple GCN architectures evaluated with inputs including protein structure, chemical structure, and chemical biological features
• Active learning framework enabling researchers to accept or reject predictions with reasoning, continuously improving model accuracy
• Deployable classification models runnable from standard data science environments

Key features:

• Automated data collection and distribution pipeline
• Model registry with daily updates of the best-performing architectures
• Human-in-the-loop validation supporting regulatory-grade documentation
• Compatible with existing cheminformatic workflows

Technology readiness level

The underlying GCN methodology for protein-ligand binding prediction has been validated in published research and demonstrated by other teams. The BioBricks.ai data infrastructure is operational and currently distributes data from 40+ life science databases. Future validation plans include building a comprehensive protein binding data registry, developing and benchmarking multiple GCN architectures, and deploying an active learning system with researcher feedback. The Insilica team brings experience from the EU ONTOX project for generalized cheminformatic modeling and prior NCATS ASPIRE awards for computational toxicology prediction. The platform is positioned for pilot deployment and collaborative refinement with pharmaceutical partners.