MAELLE framework re-engineers chemical reaction prediction by tracking electron movement with discrete flow matching to reveal reaction pathways and byproducts

Computational models for predicting chemical reactions have long treated molecules either as static graphs subjected to approximate structural modifications or as text generated from scratch, overlooking the fundamental physical reality that a chemical reaction is essentially a redistribution of charges and electron movement in molecular space. While this common simplification has yielded acceptable results on some benchmarks, it has created a deep divide between AI outputs and genuine chemical understanding, leaving those models unable to explain how transformations occur or predict unwanted byproducts.

New research led by a team of researchers including Nguyen Xuan Phu, Octavian Susanu, Daniel Armstrong, and Philippe Schwaller introduces a new framework namedMAELLE, a model that relies on discrete flow matching to track electron rearrangements. The model formulates the transformation from reactants to products as a continuous-time Markov chain over an integer-valued electron occupancy space, built on a graph structure that covers bonding sites, non-bonding sites, and hydrogen atoms combined.

The methodological strength of the framework lies in generalizing discrete flow mixture paths to electron rearrangements using optimal transport theory, which allows it to generate a sequential trajectory of chemically interpretable transition steps without requiring labeled training data for each individual elementary step.When evaluated on the USPTO-480K benchmark, MAELLE achieved competitive performance against leading reaction prediction models, with a decisive advantage in out-of-distribution robustness tests across both molecular structural complexity and unfamiliar reaction types, maintaining its accuracy in settings where conventional models falter.

Because the learned flow operates across the entire electron redistribution pathway, the framework naturally recovers mechanistic pathways that conform to established chemical rules, offering a practical capability to predict side products accompanying the primary reaction, a critical factor that typically determines economic viability and purity in advanced pharmaceutical and chemical manufacturing routes.

This methodological shift delivers a tangible impact to R&D teams, biochemistry laboratories, and petrochemical and pharmaceutical industries in the Gulf, Egypt, and the wider region, where the cost of failed laboratory experiments and raw material waste represents a primary operational bottleneck. Relying on models grounded in actual physics that predict byproducts reduces the need for expensive wet validation cycles, providing engineers and researchers with a more precise means to select synthesis routes with minimal waste and avoid surprises on pilot production lines.

The MAELLE study demonstrates that moving from superficial data generation to deep mechanistic modeling is the most reliable path toward building dependable AI tools in the exact sciences, where interpretability and adherence to physical laws become essential prerequisites for any genuine industrial application.