Generative Molecular Design with Steerable and Granular Synthesizability Control

TL;DR

This work tackles the bottleneck of synthesizability in small-molecule design by enabling steerable, granular control over predicted synthesis routes. It couples a generalist generator (Saturn) with a retrosynthesis interface (Syntheseus) and reaction-labeling tools (Rxn-INSIGHT or NameRxn) within a reinforcement-learning framework to enforce or avoid user-specified reactions and building blocks, including route-length incentives. The framework yields property-optimized molecules with predicted synthesis routes under diverse constraints, demonstrates waste-to-drugs and ultra-large library applicability, and provides a thorough benchmark against recent synthesizability-constrained models, showing superior sample efficiency with a trade-off in diversity. Practically, this enables more efficient, constraint-aware molecular design and rapid in silico screening within large reaction spaces and make-on-demand libraries, with open-source accessibility for broader adoption and extension.

Abstract

Synthesizability in small molecule generative design remains a bottleneck. Existing works that do consider synthesizability can output predicted synthesis routes for generated molecules. However, there has been minimal attention in addressing the ease of synthesis and enabling flexibility to incorporate desired reaction constraints. In this work, we propose a small molecule generative design framework that enables steerable and granular synthesizability control. Generated molecules satisfy arbitrary multi-parameter optimization objectives with predicted synthesis routes containing pre-defined allowed reactions, while optionally avoiding others. One can also enforce that all reactions belong to a pre-defined set. We show the capability to mix-and-match these reaction constraints across the most common medicinal chemistry transformations. Next, we show how our framework can be used to valorize industrial byproducts towards de novo optimized molecules. Going further, we demonstrate how granular control over synthesizability constraints can loosely mimic virtual screening of ultra-large make-on-demand libraries. Using only a single GPU, we generate and dock 15k molecules to identify promising candidates in Freedom 4.0 constituting 142B make-on-demand molecules (assessing only 0.00001% of the library). Generated molecules satisfying the reaction constraints have > 90% exact match rate. Lastly, we benchmark our framework against recent synthesizability-constrained generative models and demonstrate the highest sample efficiency even when imposing the additional constraint that all molecules must be synthesizable from a single reaction type. The main theme is demonstrating that a pre-trained generalist molecular generative model can be incentivized to generate property-optimized small molecules under challenging synthesizability constraints through reinforcement learning.

Figures (13)

• Figure 1: Overview of the framework and all reaction constraints that can be imposed. All generated molecules have a predicted synthesis route. Within these routes, reaction constraints include enforcing specific building blocks are used, enforcing specific reactions, avoiding specific reactions, and minimizing the number of synthetic steps. Molecular generation can mix-and-match these constraints which enables steerable and granular. The molecule SVG in the middle was created using Molecule Icon Generator molecule-icon.
• Figure 2: Docking scores of generated molecules filtered for QED > 0.5 across various reaction constraints with and without also enforcing blocks. The molecules are pooled across all 5 seeds (0--4 inclusive) and de-duplicated. N denotes the total number of molecules and annotated on the x-axis labels. The total number of molecules with docking scores < -10 are annotated just above the x-axis.
• Figure 3: Example synthesis routes of generated molecules under various reaction presence constraints. Left: Enforcing reaction presence. Right: Enforcing reaction and building block presence. A box is used to denote the specific reaction satisfying the constraint. If also enforcing building block presence, the specific enforced building block is highlighted and its substructure also shown in the generated molecule. Property values of the generated molecules are annotated.
• Figure 4: Example synthesis routes of generated molecules containing only the specified reactions. Left: Enforcing all reactions. Right: Enforcing all reactions and building block presence. All reactions are boxed to denote that they belong to the specified reaction set. If also enforcing building block presence, the specific enforced building block is highlighted and its substructure also shown in the generated molecule. Property values of the generated molecules are annotated.
• Figure 5: Co-emergence of reaction tandems and the effect of avoiding specific reaction classes. When generating property-optimized molecules, we observe that certain reactions co-emerge in frequency with the enforced reaction. These co-emergent reactions may be undesired and we demonstrate the ability to avoid them. a and b enforce the presence of the Mitsunobu and Wittig reaction, respectively, while avoiding other reactions. In the Enforce presence of Mitsunobu. Avoid deprotections example, the reaction is classified as "Mitsunobu esterification" by Rxn-INSIGHT, although specific reaction conditions would be needed to define the transformation (see Appendix \ref{['appendix:rxn-insight-reaction-name']} for more details about reaction name classification nuances).