Li-Cheng Xu

The integration of large language models (LLMs) into molecular synthesis is rapidly transforming the field, offering a paradigm shift beyond traditional data-driven approaches. LLMs bring a unique combination of massive knowledge absorption, sophisticated contextual representation, and multi-step logical inference, holding the potential to intelligently automate the entire synthesis workflow. This review explores how LLMs are revolutionizing molecular synthesis by functioning across four key roles: (1) as interactive domain-specific knowledge bases for instant expert consultation; (2) as powerful tools for structuring unstructured reaction data from literature; (3) as versatile predictors for molecular properties and reaction outcomes; and (4) as the central “brain” in intelligent agent systems for autonomous synthesis planning and execution. By systematically outlining these advances, we aim to provide a comprehensive roadmap for chemists to leverage the distinctive capabilities of LLMs, thereby accelerating the journey toward a more automated and intelligent future for molecular synthesis.

In catalyst discovery, data-driven design and screening are playing an increasingly central role. However, molecular descriptors that encode the decisive steric and electronic properties often lack automation and unification. We present CatEmb, a stereoelectronic-aware catalyst embedding learned through contrastive alignment of 2D and 3D molecular graph representations. Trained on the curated CatCompDB dataset, CatEmb encodes implicit 3D structural and energetic information directly from a 2D molecular graph. We demonstrate its efficacy in clustering chemically diverse ligand classes and enhancing reaction performance prediction in quantitative structure-performance relationship models. Furthermore, we develop a novel, similarity-based iterative strategy using CatEmb. For high-throughput experimental campaigns aimed at exhaustively mapping a large condition space, this strategy efficiently identifies high-performing, substrate-general catalysts, significantly outperforming random and model-based approaches. When applied to large-scale virtual libraries with limited experimental validation, it successfully prioritizes high-performance catalysts tailored for specific reactions. CatEmb provides a versatile and efficient foundation for data-driven catalyst discovery.

Transition states (TS) are pivotal for elucidating reaction mechanisms, but their calculation constitutes a major bottleneck. While generative AI provides a route to automate transition state generation, existing methods remain confined to simplified chemical systems, lacking generalizability to complex reactions like transition metal catalysis. Here, we present a unified framework for general-purpose TS discovery, comprising the UniTS-Lib library of 4,391 high-fidelity structures spanning 42 elements and diverse chemical transformations, coupled with the UniTS-Gen diffusion model that generates 3D TS configurations from 2D reactant graphs using a higher-degree equivariant network. Validation demonstrates UniTS-Gen’s superior structural accuracy and robust generalization to unseen chemical systems. We show that UniTS-Gen provides highly reliable initial guesses and actively accelerates discovery by locating kinetically preferred TS conformation. This work provides a scalable and transferable solution for automating mechanistic studies in organic synthesis.

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 modelling, and training on 13 million reactions via a carefully designed strategy, RXNGraphormer achieves state-of-the-art performance across eight benchmark datasets for reactivity or selectivity prediction and forward-synthesis or retrosynthesis planning, as well as three external realistic datasets for reactivity and selectivity prediction. Notably, the model generates chemically meaningful embeddings that spontaneously cluster reactions by type without explicit supervision. 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.

Enantioselectivity prediction in asymmetric catalysis has been a long-standing challenge in synthetic chemistry because of the high-dimensional nature of the structure–enantioselectivity relationship. A lack of understanding of the synthetic space results in laborious and time-consuming efforts in the discovery of asymmetric reactions, even if the same transformation has already been optimized on model substrates. Here we present a data-driven workflow to achieve a holistic enantioselectivity prediction of asymmetric pallada-electrocatalysed C–H activation by implementing transition state knowledge in machine learning. The vectorization of transition state knowledge allowed for an excellent description and extrapolation of the machine learning model, and enabled the quantitative evaluation of 846,720 possibilities. Model interpretation revealed the non-intuitive olefin effect on the enantioselectivity determination. Subsequent density functional theory calculations unravelled mechanistic knowledge that the rate-determining step depends on the olefin reactivity in the insertion step. Therefore, the olefin insertion step can be involved in the overall enantioselectivity determination. These results highlight the complementary features of knowledge-based machine learning with an interpretation-driven mechanistic study, which provides the opportunity to harness widely existing catalysis screening data and transition state models in molecular synthesis.

Asymmetric hydrogenation of olefins is one of the most powerful asymmetric transformations in molecular synthesis. Although several privileged catalyst scaffolds are available, the catalyst development for asymmetric hydrogenation is still a time- and resource-consuming process due to the lack of predictive catalyst design strategy. Targeting the data-driven design of asymmetric catalysis, we herein report the development of a standardized database that contains the detailed information of over 12000 literature asymmetric hydrogenations of olefins. This database provides a valuable platform for the machine learning applications in asymmetric catalysis. Based on this database, we developed a hierarchical learning approach to achieve predictive machine leaning model using only dozens of enantioselectivity data with the target olefin, which offers a useful solution for the few-shot learning problem and will facilitate the reaction optimization with new olefin substrate in catalysis screening.

Description of molecular stereostructure is critical for the machine learning prediction of asymmetric catalysis. Herein we report a spherical projection descriptor of molecular stereostructure (SPMS), which allows precise representation of the molecular van der Waals (vdW) surface. The key features of SPMS descriptor are presented using the examples of chiral phosphoric acid, and the machine learning application is demonstrated in Denmark’s dataset of asymmetric thiol addition to N-acylimines. In addition, SPMS descriptor also offers a color-coded diagram that provides straightforward chemical interpretation of the steric environment.