A recurring challenge in molecular design, whether for pharmaceutical, chemical, or material applications, is creating synthesizable molecules. Synthesizability assessment often requires mapping the synthesis pathway for a molecule: the sequence of chemical reactions needed to rework precursor molecules into the goal product molecule. This post introduces ReaSyn, a generative model from NVIDIA designed for predicting molecular synthesis pathways that also addresses limitations in current approaches.
Why chain-of-thought reasoning matters for AI in chemistry
Large language models (LLMs) have change into integral to each day life, powering applications from virtual assistants to complex problem-solving. Modern LLMs solve complex problems by generating a sequence of thought (CoT), which is a series of intermediate reasoning steps that result in a final answer. Combining CoT and test-time search methods, reminiscent of generating multiple CoT paths, are critical to the improved accuracy of recent LLMs.
Chemistry faces the same challenge in molecular synthesis pathway prediction, where a pathway incorporates a series of intermediate synthesis steps. Pathway prediction is a critical step in drug, chemical, and materials development because a molecule, nevertheless promising, is simply invaluable if it will probably‌ be synthesized. ReaSyn is a novel generative framework that efficiently predicts molecular synthesis pathways. It uses a singular chain of response (CoR) notation, inspired by the CoT approach in LLMs, combined with a test-time search algorithm.
An artificial pathway follows a bottom-up tree structure: easy molecules, or constructing blocks (BB), are combined through chemical reactions (RXN) to provide intermediate products (INT), which in turn undergo further reactions to form increasingly complex molecules (Figure 1). This process is multi-step, with each response applied to reactants which may be either constructing blocks or intermediates. In practice, chemists deduce such pathways step-by-step, reasoning through each transformation to succeed in the ultimate goal molecule.
ReaSyn captures this step-by-step reasoning through its CoR notation, inspired by the CoT approach in LLMs. In CoR, a whole synthetic pathway is represented as a linear sequence where each step explicitly includes the reactants, the response rule, and the resulting product. Reactants and products are encoded as SMILES (strings, wrapped with special tokens marking their boundaries), while each response is denoted by a single reaction-class token. This representation not only mirrors how chemists take into consideration synthesis but in addition enables the model to receive intermediate supervision at every step for richer learning of chemical response rules and more reliable multi-step pathway generation.
Constructing on the CoR notation and its sequential design, ReaSyn operates as an autoregressive generative model (Figure 2) that treats synthetic pathways as reasoning chains—each step corresponding to a single chemical response. Just as CoT reasoning in LLMs produces intermediate steps before a final answer, ReaSyn incrementally constructs a pathway ranging from easy constructing blocks to a goal molecule. This allows it to reconstruct pathways for synthesizable molecules and to project unsynthesizable molecules into synthesizable chemical space, producing close synthesizable analogs that could be made in practice.
After the model predicts the reactants and response rules at each step, intermediate products could be easily obtained using a response executor reminiscent of RDKit. This intermediate information not only provides richer training signals for the model to learn chemical rules, but in addition guides the synthetic pathway generation process.
Viewing synthetic pathway generation as LLM reasoning also enables us to adopt additional LLM reasoning techniques reminiscent of reinforcement learning (RL) finetuning and test-time search. ReaSyn explores these two techniques for synthetic pathway generation.
How outcome-based RL finetuning improves exploration
There could be multiple pathways that result in the identical product molecule. ReaSyn advantages from sampling various synthetic pathways and getting feedback via GRPO. The result-based reward for a pathway is about to the molecular similarity between its end-product molecule and the input molecule. Because the reward is simply applied to the end result, as an alternative of the reasoning steps, the model is endowed with the power to explore diverse synthetic pathways.
Goal-directed search: guiding pathways
During generation, ReaSyn uses beam search, which maintains a pool of sequences being generated and expands them block-by-block (BB or RXN). The search enables ReaSyn to generate diverse pathways for a single input molecule, and guides the generation in a preferred direction by scoring the sequences through a reward function. In retrosynthesis planning, the reward function could be the similarity to the input molecule. In goal-directed optimization tasks, the reward function could be the specified chemical property.
Generating synthetic pathways with ReaSyn
ReaSyn’s synthesizable projection is extremely versatile: it enables retrosynthesis planning, suggests analogs for unsynthesizable molecules, supports goal-directed molecular optimization, and facilitates synthesizable hit expansion. Below, we examine its performance on these tasks.
Retrosynthesis planning
Even given an unlimited synthesizable space, previous synthesizable molecule generation methods have struggled to cover this space extensively. ReaSyn shows a high success rate in generating synthetic pathways given synthesizable molecules, demonstrating its powerful explorability within the synthesizable chemical space.
Synthesizable goal-directed molecular optimization
ReaSyn can project molecules generated by an off-the-shelf molecular optimization method to perform synthesizable goal-directed optimization. Combined with Graph GA, Graph GA-ReaSyn shows higher optimization performance than previous synthesis-based methods.
Synthesizable hit expansion: exploring molecule neighborhoods
The search scheme enables ReaSyn to suggest multiple synthesizable analogs for a given goal molecule by projecting them in a different way. ReaSyn explores the neighborhood of given molecules in synthesizable space, and could be applied to hit expansion to seek out diverse synthesizable analogs of hit molecules (Figure 5).
Most generative models create molecules that aren’t synthesizable in practice. ReaSyn builds on recent reasoning advances in LLMs, equipping scientists with an efficient generative tool to project small molecules into the synthesizable chemical space. With its enhanced reasoning capabilities, diversity, and flexibility, ReaSyn shows promise as a way for navigating combinatorially large synthesizable chemical space in real-world drug discovery.
Discover more about ReaSyn by reading our paper on arXiv and the code is accessible on GitHub.
