Deep Lead Optimization: Leveraging Generative AI for Structural Modification

Odin Zhang, Haitao Lin, Hui Zhang, Huifeng Zhao, Yufei Huang, Yuansheng Huang, Dejun Jiang, Chang-Yu Hsieh, Peichen Pan, Tingjun HouJournal of the American Chemical Society 2024 · DOI · OpenReview

Abstract

The idea of using deep-learning-based molecular generation to accelerate discovery of drug candidates has attracted extraordinary attention, and many deep generative models have been developed for automated drug design, termed molecular generation. In general, molecular generation encompasses two main strategies: de novo design, which generates novel molecular structures from scratch, and lead optimization, which refines existing molecules into drug candidates. Among them, lead optimization plays an important role in real-world drug design. For example, it can enable the development of me-better drugs that are chemically distinct yet more effective than the original drugs. It can also facilitate fragment-based drug design, transforming virtual-screened small ligands with low affinity into first-in-class medicines. Despite its importance, automated lead optimization remains underexplored compared to the well-established de novo generative models, due to its reliance on complex biological and chemical knowledge. To bridge this gap, we conduct a systematic review of traditional computational methods for lead optimization, organizing these strategies into four principal sub-tasks with defined inputs and outputs. Additionally, we introduce a unified perspective based on constrained subgraph generation to harmonize the methodologies of de novo design and lead optimization.

Introduction: Bridging the "Valley of Death"

In the drug development lifecycle, the transition from a "hit"—a molecule showing initial activity—to a viable clinical candidate is often referred to as the "Valley of Death." While a hit molecule interacts with a target, it frequently suffers from toxicity, poor absorption, or weak binding affinity. The process of iteratively refining these chemical structures to overcome such flaws is known as Lead Optimization.

Overview of the AI-aided Drug Discovery (AIDD) pipeline, highlighting the role of lead optimization in transitioning from hits to drug candidates.

Traditionally, this phase relies heavily on medicinal chemistry intuition and manual experimentation. While AI has significantly advanced de novo generation (creating entirely new structures), these models often produce "molecular chimeras" that are synthetically inaccessible or ignore the structural context of the starting lead. This paper addresses the need for a unified framework for AI-aided Lead Optimization (AILO), categorizing fragmented algorithmic approaches into a cohesive engineering discipline.

A Unified Perspective: Constrained Graph Generation

We propose that lead optimization can be mathematically framed as constrained graph generation. Unlike de novo design, which seeks to learn the distribution of all possible molecules $P(G)$, lead optimization focuses on a conditional distribution $P(G_{\mu} | G_{\setminus \mu})$. Here, $G_{\setminus \mu}$ represents the "sketch" or the preserved part of the molecule, while $G_{\mu}$ represents the "fill" or the modified substructure. This "sketch and fill" approach is inherently more efficient as it leverages the massive amount of prior chemical information contained in the starting lead.

Principal Sub-tasks in Lead Optimization

The field of AILO can be organized into four pivotal sub-tasks, each addressing specific structural modification needs:

  1. Scaffold Hopping: Replacing the central core of a molecule to improve properties or navigate patent landscapes while maintaining key functional groups.
  2. Linker Design: Constructing the optimal chemical bridge to connect two active fragments.
  3. Fragment Replacement/Growing: Substituting or adding small chemical moieties to better fit a protein's binding pocket.
  4. Side-chain Decoration: Fine-tuning the peripheral "limbs" of a molecule to optimize potency and safety profiles.
Visualization of the four principal sub-tasks in lead optimization: fragment elaboration, linker design, side-chain decoration, and scaffold hopping.

Recent advancements show a significant shift from 2D topological models to 3D-aware generation. Models such as DiffLinker and D3FG utilize equivariant diffusion to generate fragments that respect the physical geometry of the protein binding pocket. Furthermore, the integration of protein pocket information (structure-based design) has been shown to increase the predicted binding probability of generated ligands from approximately 0.25 to 0.75.

Despite these technical strides, data scarcity remains a primary hurdle. Successful optimization paths are often proprietary. To mitigate this, researchers employ molecular decomposition techniques—such as RECAP, BRICS, and Matched Molecular Pairs (MMP)—to create synthetic training pairs from public databases like ChEMBL.

Conclusion and Future Directions

While AI-aided lead optimization is maturing, several challenges persist. Most current models assume a rigid protein target, ignoring the natural flexibility of biological systems. Future work must incorporate protein dynamics and move beyond simple atom-and-bond graphs to consider electronic properties (bioisosteres). Finally, the field requires new, task-specific evaluation metrics that go beyond standard AI benchmarks to measure true medicinal chemistry improvement.

Summary updated

Abstract

The idea of using deep-learning-based molecular generation to accelerate discovery of drug candidates has attracted extraordinary attention, and many deep generative models have been developed for automated drug design, termed molecular generation. In general, molecular generation encompasses two main strategies: de novo design, which generates novel molecular structures from scratch, and lead optimization, which refines existing molecules into drug candidates. Among them, lead optimization plays an important role in real-world drug design. For example, it can enable the development of me-better drugs that are chemically distinct yet more effective than the original drugs. It can also facilitate fragment-based drug design, transforming virtual-screened small ligands with low affinity into first-in-class medicines. Despite its importance, automated lead optimization remains underexplored compared to the well-established de novo generative models, due to its reliance on complex biological and chemical knowledge. To bridge this gap, we conduct a systematic review of traditional computational methods for lead optimization, organizing these strategies into four principal sub-tasks with defined inputs and outputs. This review delves into the basic concepts, goals, conventional CADD techniques, and recent advancements in AIDD. Additionally, we introduce a unified perspective based on constrained subgraph generation to harmonize the methodologies of de novo design and lead optimization. Through this lens, de novo design can incorporate strategies from lead optimization to address the challenge of generating hard-to-synthesize molecules; inversely, lead optimization can benefit from the innovations in de novo design by approaching it as a task of generating molecules conditioned on certain substructures.
BibTeX
@article{zhang2024deep,
  author = {Odin Zhang and Haitao Lin and Hui Zhang and Huifeng Zhao and Yufei Huang and Yuansheng Huang and Dejun Jiang and Chang-Yu Hsieh and Peichen Pan and Tingjun Hou},
  title = {{Deep Lead Optimization: Leveraging Generative AI for Structural Modification}},
  journal = {Journal of the American Chemical Society},
  year = {2024},
  doi = {10.1021/jacs.4c11686},
}
Figure