Abstract
Lead optimization (LO) is a critical yet complex phase in drug discovery, requiring the refinement of molecular structures to enhance potency and drug-like properties. Traditional deep learning approaches to LO are often fragmented, focusing on isolated subtasks and frequently ignoring the three-dimensional (3D) context of the protein-ligand interaction. We present Delete (Deep lead optimization enveloped in protein pocket), a novel framework that addresses these limitations through a unified deleting (masking) strategy and a structure-aware geometric network. By treating tasks such as fragment growing, linking, and replacement as a "mask-and-recover" problem, Delete generates molecules with high binding affinity and realistic 3D conformations. Evaluations on benchmark datasets and case studies demonstrate that Delete outperforms existing state-of-the-art models, providing a versatile tool for both "me-better" drug development and first-in-class hit-to-lead optimization.
Introduction
The journey from a biological "hit" to a clinical "lead" is one of the most resource-intensive stages of drug discovery. Lead optimization involves the iterative modification of a molecule's chemical structure to optimize its pharmacological profile. Despite the proliferation of AI-driven de novo molecular generation, LO has remained a challenge due to two primary factors: task fragmentation and the dimensionality gap.
Most existing AI models for LO are specialized for single tasks—such as scaffold hopping or side-chain addition—forcing researchers to manage multiple disparate workflows. Furthermore, many models operate primarily in 2D space, failing to account for the intricate 3D "lock-and-key" geometry of the protein binding pocket. This often leads to the generation of molecules that are chemically valid but biologically inactive due to steric clashes or poor fit.
The Delete Framework
To overcome these hurdles, the Delete framework introduces a unified approach to lead optimization. The model is built on two core pillars: a versatile masking strategy and a 3D-aware geometric architecture.
Unified Deleting Strategy
Inspired by the success of masked language models in natural language processing, Delete treats molecular optimization as a "fill-in-the-blanks" problem. The authors developed seven distinct masking strategies to train the model:
- General Maskers: Random, Spatial, and Topological masks allow the model to learn fundamental chemical rules and spatial relationships.
- Task-Specific Maskers: Dedicated masks for Linkers, Fragments, Scaffolds, and Side-chains enable the model to handle the specific requirements of various LO subtasks within a single architecture.
Structure-Aware Architecture
Unlike 2D-based models, Delete is designed to "see" the protein environment. The protein pocket is represented as a surface of triangle patches, and the model employs geometric neural networks to perform message passing between protein residues and ligand atoms. This allows the model to predict 3D coordinates, atomic identities, and bonding relationships simultaneously, effectively performing "in-place" generation within the binding site.
Experimental Evaluation
The model was trained and validated using two primary datasets: CrossDock (a large-scale dataset of ~100,000 docked pairs) and BindingMoad (a high-quality dataset of ~35,000 X-ray crystallized structures). Three variants were evaluated: Delete-C (trained on CrossDock), Delete-M (trained on BindingMoad), and Delete-CM (pre-trained on CrossDock and fine-tuned on BindingMoad).
Performance in Linker Design and PROTACs
In comparative studies on the BindingMoad dataset, Delete-CM significantly outperformed previous state-of-the-art models like DiffLinker. The model achieved success rates exceeding 90% and generated molecules with binding affinities superior to the original reference ligands.
The framework's utility was further demonstrated in the design of Proteolysis Targeting Chimeras (PROTACs). In a challenging case study involving PROTAC linkers, 98.1% of the compounds generated by Delete exhibited docking scores superior to the original reference molecules, highlighting its potential for complex bifunctional molecule design.
Real-World Recovery and 3D Fidelity
Delete demonstrated a remarkable ability to recover known bioactive conformations. In case studies targeting the D2 dopamine receptor and Adrb1, the model successfully re-discovered established drugs such as risperidone and cyanopindolol. Moreover, the generated molecules maintained high 3D shape similarity (often >0.8 or 0.9) to experimental crystal structures, confirming that the model respects the physical constraints of the protein pocket.
Discussion and Conclusion
The "Delete" framework represents a significant step toward a more integrated and physically grounded approach to lead optimization. By unifying multiple LO subtasks into a single model and incorporating 3D structural awareness, it reduces the computational and operational burden on medicinal chemists.
One current limitation is the assumption of a rigid protein-ligand environment during generation, as proteins and ligands are naturally flexible. Future work will likely focus on incorporating conformational flexibility to better simulate the dynamic nature of biological systems. Nevertheless, Delete provides a powerful tool for efficiently navigating the chemical space between a hit and a high-quality lead, potentially accelerating the development of both "me-better" and first-in-class therapeutics.
Summary updated