[Paper Review] Projecting Molecules into Synthesizable Chemical Spaces

Shitong Luo, Wenhao Gao+4 more|arXiv (Cornell University)|Jun 7, 2024

Innovative Microfluidic and Catalytic Techniques Innovation4 citations

TL;DR

This paper introduces a transformer-based framework that projects unsynthesizable molecules from generative models into synthesizable chemical space using a postfix notation for synthetic pathways. By encoding molecular graphs and decoding them into reaction-based postfix sequences, the model generates structurally similar, experimentally feasible analogs while preserving key properties, achieving up to 0.68 JNK3 score in hit expansion with 67% Tanimoto similarity.

ABSTRACT

Discovering new drug molecules is a pivotal yet challenging process due to the near-infinitely large chemical space and notorious demands on time and resources. Numerous generative models have recently been introduced to accelerate the drug discovery process, but their progression to experimental validation remains limited, largely due to a lack of consideration for synthetic accessibility in practical settings. In this work, we introduce a novel framework that is capable of generating new chemical structures while ensuring synthetic accessibility. Specifically, we introduce a postfix notation of synthetic pathways to represent molecules in chemical space. Then, we design a transformer-based model to translate molecular graphs into postfix notations of synthesis. We highlight the model's ability to: (a) perform bottom-up synthesis planning more accurately, (b) generate structurally similar, synthesizable analogs for unsynthesizable molecules proposed by generative models with their properties preserved, and (c) explore the local synthesizable chemical space around hit molecules.

Motivation & Objective

• To address the critical gap in generative molecular design where models produce synthetically infeasible molecules, limiting experimental validation.
• To develop a method that ensures generated molecules are not only property-optimized but also synthesizable using purchasable building blocks and known reactions.
• To enable practical application of generative models in drug discovery by transforming unsynthesizable outputs into viable, structurally similar analogs.
• To explore and expand local chemical space around hit compounds while maintaining desired bioactivity and synthetic feasibility.

Proposed method

• Representing molecules via a postfix notation of synthetic pathways, enabling a linear, scalable, and unambiguous encoding of reaction sequences.
• Designing a transformer-based encoder-decoder model where the encoder processes molecular graphs and the decoder generates postfix synthetic pathways.
• Training the model to map molecular structures to valid, step-by-step synthetic routes using purchasable building blocks and expert-defined reaction rules.
• Using the generated postfix sequences to reconstruct molecules that are guaranteed to be synthesizable through known reaction pathways.
• Applying the model to project unsynthesizable molecules from generative models into synthesizable analogs while preserving structural and functional similarity.
• Validating the method on hit expansion tasks using a JNK3 inhibition oracle function to assess both synthetic feasibility and bioactivity retention.

Experimental results

Research questions

• RQ1 Can a deep learning model effectively project unsynthesizable molecules into synthesizable chemical space while preserving key molecular properties?
• RQ2 How accurately can a model perform bottom-up synthesis planning using a postfix notation of synthetic pathways?
• RQ3 To what extent can the model generate structurally similar, synthesizable analogs of molecules proposed by goal-directed generative models?
• RQ4 Can the model explore and expand the local chemical space around hit molecules to identify improved, feasible candidates?

Key findings

• The model successfully generated 21 JNK3 inhibitors with higher scores than the original hit (0.68), achieving an average Tanimoto similarity of 0.67.
• In hit expansion, 500 analogs were generated with diverse structures and JNK3 scores, demonstrating broad coverage of synthesizable chemical space.
• For molecules from goal-directed generative models, the average Morgan fingerprint similarity of the analogs was 0.43, indicating structural divergence due to original non-synthesizability.
• Despite structural differences, the analogs retained significant property performance, with maximum objective scores close to the original unsynthesizable molecules.
• The model improved predicted binding affinity in 14 out of 15 targets, with slight improvements in binding energy (in kcal/mol) observed in most cases.
• The approach enables efficient exploration of synthesizable chemical space without relying on combinatorial optimization or virtual screening, offering a scalable alternative for drug discovery.