ITForge: A Generative AI-Driven Drug Discovery Pipeline for Structure-Based De Novo Drug Design and Molecular Ranking with Open-Source Frameworks

Poster presented at the Machine Learning Summer School on Drug and Materials Discovery (MLSS^D 2025)

Abstract:

Structure-Based De Novo Drug Design (SBDNDD) is a promising class of Computer-Aided Drug Design (CADD) methods leveraging deep generative Artificial Intelligence (AI) models to create novel, valid, and synthesizable ligands based on design constraints and the target protein’s 3D structure. The key objective is designing molecules that bind with high affinity and selectivity to a given target while ensuring optimal pharmacokinetic, toxicity, and drug-like physico-chemical properties. This approach can significantly accelerate early preclinical drug discovery, particularly hit/lead identification and optimization. 

Despite recent progress driven by increased data availability, enhanced hardware performance, and advancements in Deep Learning (DL), state-of-the-art SBDNDD models still face limitations hindering their industrial application. Common issues include low explainability, limited chemical diversity, and the lack of integrated post-processing and ranking systems. Furthermore, defining multi-objective scoring functions to effectively guide the generative process remains complex, requiring a careful balance between exploring chemical space, adhering to constraints, and maintaining creativity.  

In this study, we present ITForge, an end-to-end, AI-driven pipeline developed to support hit-to-lead workflows through generative fragment growing. ITForge integrates and optimizes multiple open-source frameworks to address significant SBDNDD limitations. The pipeline combines a generative workflow based on the scaffold decoration model LibINVENT, pre-trained on the ChEMBL database and optimized via Reinforcement Learning (RL), with a comprehensive post-processing module designed to progressively filter and rank compounds using increasingly accurate scoring stages. 

The generative chemistry process starts from a promising SMILES input fragment with defined growing points. RL guides the molecule generation toward high-scoring regions of chemical space based on a custom multi-objective function incorporating docking scores, an empirical synthetic accessibility score, and physico-chemical and structural properties related to drug-likeness and flexibility. 

Generated molecules are subsequently ranked by progressively accurate filtering and scoring methods, including: docking pose-template RMSD filtering, DL-based synthetic feasibility assessment, commercial building block search in an in-house database, ADME-Tox property prediction, refined docking scores, and binding free energy estimation. 

ITForge provides a flexible, scalable, and open-source solution for AI-driven drug design, integrating generative modeling, multi-objective optimization, and rigorous post-processing to prioritize high-quality candidate molecules for experimental validation.