Abstract
Optimizing molecules to improve their properties is a fundamental challenge in drug design. For a fine-tuning of molecular properties without losing bio-activity validated in advance, the concept of bioisosterism has emerged. Many in silico methods have been proposed for discovering bioisosteres, but they require expert knowledge for their applications or are restricted to known databases. Here, we introduce DeepBioisostere, a deep generative model to design suitable bioisosteric replacements. Our model allows a fully automated bioisosteric replacement by intelligently selecting fragments for both removal and insertion. Through various scenarios of multiple property control, we showcase the model's capability to delicately tune specific properties, addressing the challenge in molecular optimization. Our model's innovation lies in its capacity to design a bioisosteric replacement reflecting the compatibility with the surroundings of the modification site, facilitating the control of sophisticated properties that depend on the overall molecular structures such as drug-likeness. DeepBioisostere can also provide previously unseen bioisosteric replacements, highlighting its capability for exploring diverse chemical modifications rather than just mining them from known databases. Lastly, we used DeepBioisostere to improve the sensitivity of a known SARS-CoV-2 main protease inhibitor to the E166V mutant that exhibits drug resistance to the inhibitor, demonstrating its potential application in drug discovery.
Introduction
In the complex landscape of drug discovery, identifying a molecule that binds effectively to a biological target is merely the first step. Often, these "lead" molecules suffer from suboptimal physicochemical properties, such as low solubility, metabolic instability, or poor synthetic accessibility. To address these issues without compromising biological activity, medicinal chemists employ bioisosterism—the strategic replacement of chemical groups with others that maintain similar spatial or electronic characteristics while improving the drug's overall profile.
Historically, bioisosteric replacement has been a manual, intuition-driven endeavor. While computational tools have emerged, they often rely on static database lookups, limiting their ability to suggest novel structures. Furthermore, these tools frequently require human experts to pre-select the modification site and often fail to account for how a specific fragment swap influences the properties of the entire molecule.
Methodology
DeepBioisostere is a deep generative model designed to automate molecular optimization through a sequential three-step decision process:
- Removal: The model autonomously identifies which substructure within the molecule should be excised.
- Insertion: It selects an appropriate replacement fragment from a diverse library.
- Orientation: It determines the optimal geometric configuration for re-attaching the new fragment to the remaining molecular scaffold.
Technical Architecture
The model utilizes a hierarchical architecture featuring two levels of Message Passing Neural Networks (MPNNs). An Atom-level network (AMPN) captures fine-grained chemical details, while a Fragment-level network (FMPN) evaluates how these pieces integrate into the broader molecular context. To ensure chemical feasibility, fragmentation follows BRICS rules (Bridges Retrosynthetically Interesting Chemical Substructures), ensuring that suggested modifications are synthetically accessible.
Training and Data
DeepBioisostere was trained using a massive dataset derived from the ChEMBL database. The researchers utilized over 24.1 million Matched Molecular Pairs (MMPs)—pairs of molecules differing by a single small fragment—to teach the model the statistical relationships between structural changes and property shifts. The model draws from a library of 206,008 diverse fragments for its insertion step.
Results and Discussion
A critical insight from this research is that the impact of a bioisosteric swap is highly dependent on the molecular "environment." While simple properties like lipophilicity (logP) may change additively, complex properties like drug-likeness (QED) exhibit non-linear relationships. DeepBioisostere’s fragment-level embeddings allow it to account for these contextual nuances, enabling more sophisticated optimizations than traditional similarity-based methods.
Multi-Property Optimization
The model's performance was validated across several real-world scenarios. It demonstrated high precision in tuning properties like molecular weight (MW), logP, and QED. For instance, when tasked with increasing or decreasing logP by 1.0 while maintaining a constant molecular weight, the model achieved mean changes of +0.81 and -0.86, respectively.
Case Study: SARS-CoV-2
To demonstrate practical utility, DeepBioisostere was applied to optimize Ensitrelvir, a known SARS-CoV-2 main protease inhibitor, to combat the drug-resistant E166V mutant. The model suggested a cyclohexanol replacement that was predicted to have a tenfold increase in binding affinity while preserving drug-likeness and synthetic ease.
Conclusion
DeepBioisostere represents a significant advancement in automated molecular design. By moving beyond simple database mining and incorporating contextual awareness through deep learning, it provides a powerful tool for the delicate task of lead optimization. Future work will focus on integrating 3D protein-ligand interaction data to further refine the orientation of replacement fragments within specific binding pockets.
Summary updated