Nvidia’s ReaSyn Uses Chain‑of‑Reaction Reasoning to Boost Molecule Reconstruction and Path Diversity

Modern drug discovery faces a "dual dilemma": the chemical space is astronomically large (≈10^60 possible molecules with just ten atoms) and candidate molecules must satisfy activity, toxicity, solubility, etc., leading to development cycles over ten years, costs of billions of dollars, and success rates below 10%.

While molecular generation models promised to shorten discovery cycles, they often produce molecules that are impractical to synthesize, limiting real‑world impact.

Academic attempts to address this have taken two routes: (1) treating synthesizability as an optimization objective with a scoring function—effective only when the complex structure‑to‑synthesizability relationship can be captured; (2) restricting generation to known synthesizable molecules—improving feasibility but sacrificing structural novelty. The "synthetic projection" strategy emerged to modify unsynthesizable molecules into similar, synthesizable analogs, balancing innovation and practicality.

In this context, Nvidia’s research team introduced ReaSyn, a high‑efficiency, synthesis‑aware molecular projection framework. ReaSyn adopts a Chain‑of‑Reaction (CoR) representation that casts a synthesis path as a Large Language Model (LLM) chain‑of‑thought (CoT), opening a new avenue for tackling realistic synthesis challenges.

Key claims : ReaSyn achieves the highest reconstruction rate and path diversity among existing methods, delivers the best optimization performance on target‑directed tasks, and significantly outperforms prior approaches in hit‑expansion experiments.

ReaSyn framework and CoR representation convert synthesis paths into interpretable reasoning chains.

Custom RL fine‑tuning and computational scaling dramatically improve exploration efficiency and optimization performance.

Multi‑task experiments confirm the framework’s effectiveness and versatility in synthesizable molecule generation and optimization.

The authors first built a dataset that mirrors real‑world drug‑discovery scenarios. It comprises 115 common reaction types and 212,000 purchasable building blocks sourced from the Enamine inventory, jointly defining a synthesizable chemical space exceeding 10^60 molecules. The benchmark focuses on the "synthesizable molecule reconstruction" task, evaluating a model’s ability to generate feasible synthesis routes for given target molecules.

Test‑set design includes several challenging collections: (1) 1,000 random molecules from the Enamine REAL diversity set and ChEMBL database as a baseline; (2) an extended set simulating inventory updates, where 37,000 molecules with fewer than 18 heavy atoms are drawn from ZINC250k to create new building blocks and generate another 1,000 test molecules; (3) the ChEMBL test set from Luo et al. for comparability with prior work.

ReaSyn Technical Pipeline

The CoR representation defines the synthesizable space as the product of a building‑block set and a reaction‑rule set, each rule expressed in SMARTS. A synthesis path is a sequence of functional blocks sharing a unified vocabulary: molecular blocks are encoded as specially marked SMILES strings, while reaction blocks are represented by single tokens. Concatenating these tokens yields the full path sequence.

CoR offers three breakthroughs: (1) it embeds chain‑of‑thought reasoning at the reaction level; (2) it predicts complete paths without hierarchical classification; (3) it eliminates dependence on molecular fingerprints.

Model training follows a two‑stage strategy. The supervised stage uses paired target‑molecule and synthesis‑path data, training a Transformer to predict the next token while applying a token‑type weighted loss and leveraging intermediate products for richer supervision. The reinforcement‑learning fine‑tuning stage employs an online RL algorithm that rewards high‑quality paths and promotes stable model behavior, compensating for the limited exploration capability of pure supervised learning.

During inference, ReaSyn combines a stack structure with beam search. The stack dynamically manages reactants and intermediates, supporting step‑wise reasoning, while beam search maintains multiple high‑scoring candidate paths to preserve diversity. Scoring strategies are task‑specific: reconstruction tasks prioritize structural similarity and reaction feasibility; optimization and activity‑expansion tasks incorporate a reward model that evaluates building‑block and intermediate properties, guiding the search toward molecules with desirable attributes within the synthesizable space.

Experimental Results

Across several key tasks, ReaSyn consistently surpasses existing methods such as SynNet and SynFormer. In the synthesizable molecule reconstruction benchmark, it attains the highest reconstruction rate and path diversity.

For target‑directed molecule optimization, the authors integrate ReaSyn into a Graph Genetic Algorithm (Graph GA‑ReaSyn). On 15 TDC benchmark tasks, Graph GA‑ReaSyn achieves superior "AUC top‑10" scores and markedly higher Synthesizability (SA) scores than all synthesis‑constrained baselines, demonstrating that ReaSyn improves synthesizability without sacrificing optimization performance.

In a hit‑expansion study using JNK3 inhibitors, ReaSyn generates 100 analogs per seed molecule via beam search. Evaluation on "analog rate", "improvement rate", and "success rate" shows ReaSyn outperforms prior approaches on all metrics.

Broader AI‑Driven Synthesis Landscape

Beyond ReaSyn, the community explores diverse AI‑enabled synthesis solutions. The University of Toronto’s Organa robot combines computer vision with LLMs to translate natural‑language commands into chemical description language (χDL) code, automating laboratory tasks. Liverpool University’s Mobile Robotic Chemist completed 688 experiments in eight days, exploring 1,000 catalytic formulations and discovering a new catalyst.

Industry efforts focus on integrating advanced AI into production pipelines. BenevolentAI’s partnership with Merck leverages an end‑to‑end AI platform that uses LLM‑based synthesis reasoning to deliver synthesizable, high‑activity compounds, dramatically shortening the lead‑time from concept to candidate. Insilico Medicine’s end‑to‑end drug‑discovery workflow incorporated a ReaSyn‑like synthesizable projection module, achieving 100% synthesis success for the INS018_055 candidate and reducing the overall development cycle by 60%.

These academic and industrial initiatives, while differing in focus and implementation, share the common goal of enhancing our ability to design and synthesize useful molecules efficiently, thereby accelerating progress in drug discovery, material science, and related fields.