FSscore: A Machine Learning-based Synthetic Feasibility Score Leveraging Human Expertise (2312.12737v2)

Abstract: Determining whether a molecule can be synthesized is crucial in chemistry and drug discovery, as it guides experimental prioritization and molecule ranking in de novo design tasks. Existing scoring approaches to assess synthetic feasibility struggle to extrapolate to new chemical spaces or fail to discriminate based on subtle differences such as chirality. This work addresses these limitations by introducing the Focused Synthesizability score~(FSscore), which uses machine learning to rank structures based on their relative ease of synthesis. First, a baseline trained on an extensive set of reactant-product pairs is established, which is then refined with expert human feedback tailored to specific chemical spaces. This targeted fine-tuning improves performance on these chemical scopes, enabling more accurate differentiation between molecules that are hard and easy to synthesize. The FSscore showcases how a human-in-the-loop framework can be utilized to optimize the assessment of synthetic feasibility for various chemical applications.

• The paper presents FSscore, a novel approach that integrates expert pairwise preferences with fine-tuning to assess synthetic feasibility.
• It employs a two-stage process, starting with GNN-based pre-training on reactant-product pairs to capture synthetic complexity.
• The model seamlessly integrates into generative frameworks, demonstrating enhanced performance in complex chemical domains like natural products and PROTACs.

Introduction

Synthetic feasibility is crucial in various fields of chemistry and drug discovery, ensuring practicality and efficiency in experimental efforts. Traditionally, assessing the synthesizability of molecules relies on the expertise of chemists. In the digital era, computational models offer an automated means to score synthetic feasibility; however, these models often struggle with generalizing beyond their original domain and recognizing subtle yet critical differences like chirality. Addressing these challenges, a new model known as the Focused Synthesizability score (FSscore) has been developed.

Methodology

The FSscore utilizes a two-stage training process. Initially, it is pre-trained on a large dataset of reactant-product pairs, leveraging a graph neural network (GNN) to create an unbiased baseline. This first stage avoids the need for a predefined ground truth score and instead relies on observed chemical reactions, effectively capturing synthetic complexity.

Enhancing its capabilities involves fine-tuning this baseline model with data labeled by expert chemists, focusing on specific chemical domains of interest. A particularly novel approach includes employing pairwise preferences to refine the score, identifying one molecule as more or less synthesizable than another. This method is fully differentiable, which offers seamless application in various computational chemistry tools including generative models and reinforcement learning settings.

Contributions and Applications

The researchers present key contributions of the FSscore that include:

• A novel synthetic feasibility approach combining pairwise preferences with fine-tuning based on expert knowledge.
• Successful integration into generative models due to its differentiable nature.
• Enhanced focus on specific chemical spaces, including challenging domains like natural products and PROTACs, while requiring minimal labeled data.
• Demonstrated ability to generate synthesizable molecular structures with high docking scores in generative model applications.

Performance and Future Direction

The FSscore, designed to be refined with specific chemical knowledge, is not a one-size-fits-all solution but one meant for tailored applications. While pre-training establishes a robust baseline, fine-tuning to particular chemical spaces could result in the loss of generality. Despite these challenges, the model shows improved performance across various case studies over baseline models and established scores.

Future endeavors could explore integrating rule-based features to strengthen the baseline performance before fine-tuning or assessing the model's efficacy across a broader range of chemical spaces. Moreover, extended applications with different generative architectures and objectives could further solidify the FSscore's potential as a tool in molecular design and virtual screening workflows.