Abstract
Artificial intelligence has transformed the field of precise organic synthesis. Data-driven methods, including machine learning and deep learning, have shown great promise in predicting reaction performance and synthesis planning. However, the inherent methodological divergence between numerical regression-driven reaction performance prediction and sequence generation-based synthesis planning creates formidable challenges in constructing a unified deep learning architecture. Here we present RXNGraphormer, a framework to jointly address these tasks through a unified pre-training approach. By synergizing graph neural networks for intramolecular pattern recognition with Transformer-based models for intermolecular interaction modeling, and training on 13 million reactions via a carefully designed strategy, RXNGraphormer achieves state-of-the-art performance across eight benchmark datasets for reactivity/selectivity prediction and forward-/retro-synthesis planning, as well as three external realistic datasets for reactivity and selectivity prediction. Notably, the model generates chemically meaningful embeddings that: (1) spontaneously cluster reactions by type without explicit supervision, and (2) reveal structure-performance relationships through post-hoc interpretation. This work bridges the critical gap between performance prediction and synthesis planning tasks in chemical AI, offering a versatile tool for accurate reaction prediction and synthesis design.
Introduction
In the field of chemical AI, researchers typically face a "great divide." On one side is reaction performance prediction—numerical regression tasks that ask, "What will the yield be?" or "How selective is this catalyst?" On the other side is synthesis planning—sequence generation tasks that ask, "What are the starting materials for this product?" (retrosynthesis) or "What will these reactants produce?" (forward synthesis).
Because these tasks require different outputs (numbers vs. molecular strings), they have historically relied on fragmented architectures. Regression models often lack mechanistic context, while generative models rarely account for the actual performance or "reactivity" of the proposed steps. The authors of this paper sought to create a single, unified architecture capable of mastering both.
Methods
The authors introduce RXNGraphormer, which synergizes two powerful neural architectures: Graph Neural Networks (GNNs) to encode individual molecules and capture internal structural properties, and Transformers to model complex intermolecular interactions between different reaction components (reactants, reagents, and products).
A core innovation is the delta-link method. Instead of just looking at the "before" and "after" molecules, the model generates "delta-mol" graphs—intermediate representations that highlight exactly which bonds are broken or formed during a reaction. This provides the model with mechanistic insights similar to how a human chemist thinks, without requiring expensive quantum chemistry (DFT) calculations.
The model was pre-trained on a massive scale of 13 million reactions, including 6.8 million real reactions and 6.8 million "fictitious" reactions created via a fragment-exchange algorithm. The pre-training task forced the model to distinguish real chemical transformations from fake ones, learning the fundamental rules of molecular framework conservation.
Key Insights and Results
The study utilized a wide array of high-quality datasets, including Buchwald–Hartwig and Suzuki–Miyaura for yields, and radical C–H functionalization and asymmetric thiol addition for selectivity. Synthesis planning was benchmarked on standard USPTO datasets.
RXNGraphormer achieved state-of-the-art (SOTA) results across the board. In yield prediction, it reached an $R^2$ of 0.971 on the Buchwald–Hartwig dataset. For selectivity, it achieved an $R^2$ of 0.992 for regioselectivity and 0.915 for enantioselectivity. On the massive USPTO-full dataset, it set new SOTA records for retrosynthesis, improving Top-1 accuracy to 47.4%.
Beyond raw metrics, the model demonstrated several key capabilities:
- Spontaneous Clustering: The pre-trained encoder naturally organized reactions by type in its latent space without explicit supervision.
- Smoothing "Reactivity Cliffs": Fine-tuning transformed messy, non-linear relationships into smooth, linear progressions in the model's internal representation, improving generalization.
- Interpretability: Visualization of embeddings revealed that the model successfully captured implicit Structure-Performance Relationships (SPRs), making the AI's "thought process" more interpretable for chemists.
Limitations and Future Work
While RXNGraphormer is a major step forward, performance on literature-derived datasets (like NiCOlit) is lower than on curated benchmarks, highlighting the challenge of "noisy" experimental data. Future work could involve more explicit modeling of physical conditions (temperature, pressure) and incorporating full 3D transition-state geometries to better address complex stereochemical problems.
Summary updated