ReactionTeam: Teaming Experts for Divergent Thinking Beyond Typical Reaction Patterns

TL;DR

ReactionTeam addresses the inherent stochasticity of chemical reactions by moving beyond likelihood-maximization to capture multiple plausible outcomes for the same set of reactants. It introduces sequential Mixture-of-LoRA Experts (Seq MOLE) with a chief model plus lightweight LoRA adapters, plus inference-time dropout to induce diverse yet plausible predictions, and a ranking mechanism to select the best candidates. Across USPTO-MIT and USPTO-STEREO datasets, ReactionTeam improves Top-K accuracy, particularly for atypical reactions, and increases output diversity and novelty without sacrificing Top-1 performance. The approach offers a scalable, parameter-efficient way to mimic divergent chemical thinking, with broader implications for discovery in chemistry and related disciplines.

Abstract

Reaction prediction, a critical task in synthetic chemistry, is to predict the outcome of a reaction based on given reactants. Generative models like Transformer have typically been employed to predict the reaction product. However, these likelihood-maximization models overlooked the inherent stochastic nature of chemical reactions, such as the multiple ways electrons can be redistributed among atoms during the reaction process. In scenarios where similar reactants could follow different electron redistribution patterns, these models typically predict the most common outcomes, neglecting less frequent but potentially crucial reaction patterns. These overlooked patterns, though rare, can lead to innovative methods for designing synthetic routes and significantly advance synthesis techniques. To address these limitations, we build a team of expert models to capture diverse plausible reaction outcomes for the same reactants, mimicking the divergent thinking of chemists. The proposed framework, ReactionTeam, is composed of specialized expert models, each trained to capture a distinct type of electron redistribution pattern in reaction, and a ranking expert that evaluates and orders the generated predictions. Experimental results across two widely used datasets and different data settings demonstrate that our proposed method achieves significantly better performance compared to existing state-of-the-art approaches.

Figures (11)

• Figure 1: Number of training samples whose $loss > 0$ over training epochs.
• Figure 2: Reactants that follow typical or atypical electron redistribution patterns are encoded into similar representations due to structural similarity. The decoder learns to favor high-probability tokens (e.g., C and N), while underrepresenting tokens necessary for atypical reactions (e.g., Li).
• Figure 3: The overall framework of our proposed solution ReactionTeam.
• Figure 4: The Top-K Accuracy (%) of evaluated models in the USPTO-MIT and USPTO-STEREO datasets.
• Figure 5: Our Method V.S. Previous Best Baseline on predicted candidate products novelty and diversity, as judged by GPT-5.