Consideration of transition-state (TS) conformer ensembles is required to accurately model a reaction, and thus plays a key role in computational catalyst design. While CREST and GOAT are established methods for TS conformer ensemble generation, the associated computational cost remains a major bottleneck in computational chemistry pipelines, including for the generation of large machine learning data sets for catalyst design. To this end, we present racerTS (RApid Conformer Ensembles with RDKit for Transition States), a method for efficient TS conformer ensemble generation. In this work, we describe the algorithm behind racerTS, which is based on constrained distance geometry. To benchmark the performance of racerTS against CREST and GOAT, we created conformer ensembles for transition states of 20 diverse reactions. To assess the utility of each conformer generator in computational chemistry workflows, we optimize selected low-energy and diverse conformers at the DFT level. We use the generated conformer ensembles and the results of this pipeline to assess conformer generators according to the following metrics: computational cost, exhaustiveness, validity, and accuracy in low-energy regions. Considering the generated ensembles, we find that racerTS covers the conformer space similarly to CREST and slightly less comprehensively than GOAT, while the validity of the DFT-optimized TSs is better and the accuracy in the low-energy region is sufficient for computational chemistry applications (median error of 0.17 kcal/mol). Remarkably, racerTS achieves these results with a significant reduction in required wall-time. Our results demonstrate that racerTS is a highly efficient TS conformer ensemble generator, allowing for rapid TS conformer sampling in computational chemistry pipelines. Additionally, racerTS paves the way to create meaningful TS data sets to advance machine learning methods for the discovery of novel and sustainable catalysts.

Encoding bond and stereochemical changes explicitly into reaction strings can benefit language model property predictions, particularly when training data is scarce.

C–N and C–C cross-coupling reactions are central to medicinal chemistry, but selecting effective conditions remains a major bottleneck because outcomes depend on strongly coupled choices of catalyst/ligand, base, solvent, and additives. In high-throughput experimentation (HTE), these choices must also be translated into concrete plate layouts that obey practical constraints for which components can be combined together. Here we present an end-to-end machine-learning workflow that connects data curation, condition generation, and constraint-aware plate design. We curate and share two industrial HTE datasets for Buchwald–Hartwig amination and Suzuki–Miyaura coupling that include a substantial number of reactions with negative outcomes. Using these data, we develop a conditional variational autoencoder (cVAE) that generates candidate conditions while explicitly modelling both productive and unproductive regions of condition space, and we introduce a Frequency Chain baseline that is competitive under limited sampling budgets. The cVAE outperforms the baselines for low-to-moderate (below 100) sampled conditions, with improvements compared to previous cVAEs in six of eight tested metrics on the Buchwald–Hartwig reaction and one of eight on the Suzuki–Miyaura reaction. In single-transformation case studies, moving from a zero-shot to a few-shot regime improves key component recovery from 33% to 75% (Buchwald–Hartwig) and from 25% to 67% (Suzuki–Miyaura). Finally, an integer linear programming (ILP)-based plate design algorithm converts model predictions into executable HTE plates, matching or outperforming greedy positive-condition coverage while reducing predicted negative wells by 6.1 on average across ten evaluated transformations. Our integrated, data-driven plate design that combines machine learning predictions and lab automation has the potential to enable closed-loop optimisation of reaction conditions for rapid hit-to-lead optimisation.

The discovery of new organic photocatalysts (PCs) for energy transfer (EnT) catalysis remains a significant challenge, largely due to the vast and underexplored chemical space and the delicate balance of the photocatalytic properties. While transition-metal catalysts are effective, their high cost and environmental impact necessitate the development of metal-free alternatives. In this work, we present a hybrid inverse molecular design strategy that combines global exploration with targeted local optimization to discover highly efficient organic PCs. Our approach leverages a generative model, guided by machine learning predictions and semiempirical simulations, to efficiently navigate chemical space and identify promising molecular scaffolds. We demonstrate the utility of this strategy by rediscovering known PCs and, more importantly, exploring uncharted structural regions, leading to the identification of novel candidates with favorable photophysical properties. A subsequent local exploration stage, using quantum mechanical calculations, allows refinement of the properties as well as control of the synthetic complexity. The practical applicability of the approach is demonstrated by performing a local exploration of one of the identified scaffolds and successfully synthesizing four candidate PCs. We showcase their catalytic aptitude in three different EnT-mediated reactions, including a challenging aza-photocycloaddition, where one of our designed PCs achieved 90% yield, a performance comparable to a state-of-the-art iridium-based catalyst. This study highlights the power of a data-driven inverse design framework to bridge computational discovery and experimental validation, accelerating the identification of novel PCs and expanding the scope of EnT catalysis.

General reaction conditions are a long-standing goal in chemical synthesis, as such conditions facilitate library synthesis and a broad substrate scope. However, despite their importance, the generality of reaction conditions is mostly an afterthought when reaction conditions are optimised. Considering multiple substrates during reaction condition optimisation alleviates this problem and enables the identification of conditions that work well for multiple substrates. Inspired by data-driven optimisation techniques for one model substrate, machine learning based strategies have also been proposed to optimise reactions towards general reaction conditions. In this work, we describe recent algorithmic advances in this domain, including our state-of-the-art algorithm. This algorithm is also available as an easy-to-use website to allow experimental chemists to use it without code.

Adapting generative foundation models, in particular diffusion and flow models, to optimize given reward functions (e.g., binding affinity) while satisfying constraints (e.g., molecular synthesizability) is fundamental for their adoption in real-world scientific discovery applications such as molecular design or protein engineering. While recent works have introduced scalable methods for reward-guided fine-tuning of such models via reinforcement learning and control schemes, it remains an open problem how to algorithmically trade-off reward maximization and constraint satisfaction in a reliable and predictable manner. Motivated by this challenge, we first present a rigorous framework for , which brings an optimization viewpoint to the introduced adaptation problem and retrieves the relevant task of constrained generation as a sub-case. Then, we introduce Constrained Flow Optimization (CFO), an algorithm that automatically and provably balances reward maximization and constraint satisfaction by reducing the original problem to progressive fine-tuning via established, scalable methods. We provide convergence guarantees for constrained generative optimization and constrained generation via CFO. Ultimately, we present an experimental evaluation of CFO on both synthetic, yet illustrative, settings, and a molecular design task.

The success of generative molecular design hinges on a model's steerability toward high-reward samples. Because many molecular properties are intrinsically linked to molecular size, accurately capturing the joint distribution of properties and the number of atoms is essential. However, current diffusion and flow-based models fix the number of atoms, which ultimately limits their ability to navigate this complex relationship. To address this, we introduce Morph, a flexible-size generative model for conditional and unconditional 3D molecular design based on geometric graphs. By dynamically adapting size, Morph can seamlessly integrate existing structural priors, like scaffolds, and significantly enhances property steering. We show that Morph matches current fixed-size state-of-the-art models while offering the benefit of unparalleled sampling flexibility. We demonstrate out-of-distribution generation in regimes where previous models fail, paving the way for enhanced generative modeling for molecular design.

Transition metal complexes are central to catalysis, drug design, and materials science, with relevant properties strongly sensitive to their three-dimensional geometry. However, the electronic diversity and unconventional bonding environments of transition metal complexes pose a major challenge for accurate structure generation. In this work, we introduce TMCgen, a manifold diffusion machine learning model that efficiently and accurately generates geometries of transition metal complexes. By formulating the diffusion process over the metal-ligand coordination angles, combined with torsional and rotational diffusion of the ligands, TMCgen focuses on the key geometric degrees of freedom of transition metal complexes. TMCgen shows strong performance in generating accurate coordination environments on a diverse set of experimentally derived bioinorganic and organometallic complexes while requiring only few inference steps, enabling efficient generation. Our results demonstrate the potential of manifold-based generative modeling for data-efficient geometry generation, paving the way for property-conditioned design of transition metal complexes.

Uncertainty quantification for image data is dominated by complex deep learning methods, yet the field lacks an interpretable, mathematically grounded baseline. We propose Bayesian scattering to fill this gap, serving as a first-step baseline akin to the role of Bayesian linear regression for tabular data. Our method couples the wavelet scattering transform-a deep, non-learned feature extractor-with a simple probabilistic head. Because scattering features are derived from geometric principles rather than learned, they avoid overfitting the training distribution. This helps provide sensible uncertainty estimates even under significant distribution shifts. We validate this on diverse tasks, including medical imaging under institution shift, wealth mapping under country-to-country shift, and Bayesian optimization of molecular properties. Our results suggest that Bayesian scattering is a solid baseline for complex uncertainty quantification methods.

Designing safe and sustainable chemicals is critical to combat chemical pollution in our environment. Machine learning (ML) methods have been developed to aid with de novo molecule design. However, data on the environmental impacts of chemical compounds are sparse, resulting in low-fidelity ML oracles and unreliable candidate proposals. Furthermore, generative ML models rely on numerical scoring functions that cannot fully capture the nuanced chemical intuition of expert scientists required for real-world molecular design. We present GEMS-an interactive visual analytics tool that enables domain experts to directly collaborate with a genetic algorithm for molecule design. Users can integrate their expert knowledge to guide the evolutionary process by modifying the scoring function and molecule population without programming knowledge or ML developer support. A usage scenario demonstrates the system's application in designing sustainable antioxidant alternatives. In an interview session with domain scientists, we collected feedback on the usefulness of GEMS.

The current digital revolution has driven a rapid adoption of machine learning (ML) to support decision-making in chemistry. In this context, many studies now incorporate ML models based on quantum-mechanical (QM) descriptors to accelerate the discovery of candidates across a wide range of chemical problems. While density functional theory (DFT) has traditionally served as the primary source of such descriptors, recent advances in cost-effective QM methods offer an opportunity to reduce computational cost and facilitate faster integration of ML into chemical workflows. Herein, we present an automated and user-friendly workflow that generates 39 electronic and steric cost-effective descriptors and their integration into ML models. These models are broadly applicable to fields involving finite molecular systems, such as homogeneous catalysis and drug discovery. We evaluated the acceptance of these cost-effective descriptors in ML-driven catalysis through a blinded survey of 52 participants, which indicated that experts consider the resulting ML models to be as reliable and interpretable as publication-quality DFT-based models. Finally, we introduced a fully automated protocol that generates descriptors and ML predictors from SMILES strings, enabling rapid predictor development accessible to the broader chemistry community. This cost-effective framework holds the potential to accelerate the integration of ML into everyday chemical research and contribute to a more inclusive digital transformation of the field.

Accurate prediction of reaction properties, such as activation energies, enthalpies, rate constants and yields, is key to the design of efficient and sustainable chemical processes. While graph-based models using Condensed Graphs of Reaction (CGR) have shown strong performance, text-based Transformer models trained on reaction SMILES have gained popularity due to their scalability and performance. In this work, we analyze the performance of an existing CGR-based string representation for reaction property prediction, the SMILES/CGR. In contrast to conventional reaction SMILES, this representation explicitly encodes atom and bond changes within a unified SMILES-like sequence. A key limitation of SMILES/CGR is that it does not encode stereochemistry. To overcome this, we introduce the superimposed reaction SMILES (sr-SMILES), a compact CGR-based representation that preserves full stereochemical information. To evaluate both representations, we pretrained a BERT transformer on two million USPTO reactions via masked language modeling, and then finetuned separate models on six benchmark datasets. We compared against models trained on standard reaction SMILES, naïve baselines, and results from the literature. Our results show that sr-SMILES consistently matches or outperforms reaction SMILES, particularly on benchmarks where mechanistic information is critical, as well as in low-data regimes. In contrast, SMILES/CGR tends to underperform, highlighting that the details of CGR encoding matter. While sr-SMILES did not close the performance gap with graph-based CGR methods, it shows that language models can benefit from explicit mechanistic information when appropriately encoded. Moreover, by comparing SMILES/CGR and sr-SMILES models with different atom mappings, we reinforce the notion that CGR-based approaches are highly sensitive to the mapping quality. A plausible explanation for the similar performance observed across string-based representations is that language models implicitly learn the atom mapping during (pre-)training, rendering explicit mapping unnecessary for many prediction tasks.

Artificial intelligence (AI) and machine learning (ML) are rapidly reshaping the landscape of computational chemistry, offering new opportunities for accelerating catalyst discovery and deepening our understanding of chemical reactivity. This perspective highlights emerging methodologies ranging from machine learning potentials and reinforcement learning to generative AI and large language models that are poised to transform computational catalysis. We discuss challenges in developing robust molecular representations for transition-metal complexes, bridging mechanistic understanding with AI-driven predictions, and constructing reliable data sets that capture both successful and failed reactivity outcomes. By drawing on the authors' practical experience across computational, experimental, and AI-driven domains, we emphasize the importance of integrating chemical intuition and methodological expertise with data-driven approaches while remaining open to serendipitous discoveries enabled by automation and self-driving laboratories. Ultimately, the future of computational catalysis lies in balancing human intuition with algorithmic power, leveraging AI not as a replacement but as an accelerator of chemical insight, mechanistic understanding, and catalyst design.

The Suzuki-Miyaura cross-coupling reaction is plagued by protodeboronation, an undesirable side reaction with water that consumes the boronic acid derivatives required for the cross-coupling reaction. Meticulous mechanistic studies have previously established protodeboronation to be highly sensitive to the nature of the boronic reagent and reaction conditions. Particularly, the presence of bases, which are essential for the Suzuki-Miyaura coupling, is known to catalyze protodeboronation. However, protodeboronation catalyzed by palladium-phosphine complexes, the benchmark catalyst system for Suzuki-Miyaura cross-coupling, has been understudied compared to its base-catalyzed counterpart. We demonstrate, using automated high-throughput experimentation, comprehensive computational mechanistic analyses and kinetic modeling, that protodeboronation is accelerated by palladium(II) complexes bound to bulky phosphine ligands. While sterically hindered ligands are typically used to facilitate difficult cross-couplings, these ligands can instead paradoxically impede cross-coupling product formation, requiring careful and judicious consideration when choosing ligands for Suzuki-Miyaura cross-couplings.

The regio- and site-selectivity of organic reactions is one of the most important aspects when it comes to synthesis planning. Due to that, massive research efforts were invested into computational models for regio- and site-selectivity prediction, and the introduction of machine learning to the chemical sciences within the past decade has added a whole new dimension to these endeavors. This review article walks through the currently available predictive tools for regio- and site-selectivity with a particular focus on machine learning models while being organized along the individual reaction classes of organic chemistry. Respective featurization techniques and model architectures are described and compared to each other; applications of the tools to critical real-world examples are highlighted. This paper aims to serve as an overview of the field's status quo for both the intended users of the tools, that is synthetic chemists, as well as for developers to find potential new research avenues.

Volcano plots, stemming from the Sabatier principle, visualize descriptor-performance relationships, allowing rational catalyst design. Manually drawn volcanoes originating from experimental studies are potentially prone to human bias as no guidelines or metrics exist to quantify the goodness of fit. To address this limitation, we introduce a framework called SPOCK (systematic piecewise regression for volcanic kinetics) and validate it using experimental data from heterogeneous, homogeneous, and enzymatic catalysis to fit volcano-like relationships. We then generalize this approach to DFT-derived volcanoes and evaluate the tool's robustness against noisy kinetic data and in identifying false-positive volcanoes, i.e., cases where studies claim a volcano-like relationship exists, but such correlations are not statistically significant. Once the SPOCK's functional features are established, we demonstrate its potential to identify descriptor-performance relationships, exemplified via the ceria-promoted water-gas shift and single-atom-catalyzed electrocatalytic carbon dioxide reduction reactions. In both cases, the model uncovers descriptors previously unreported, revealing insights that are not easily recognized by human experts. Finally, we showcase SPOCK's capabilities to formulate multivariable descriptors, an emerging topic in catalysis research. Our work pioneers an automated and standardized tool for volcano plot construction and validation, and we release the model as an open-source web application for greater accessibility and knowledge generation in catalysis.

Modeling of solid-state material-molecule interfaces in heterogeneous catalysis requires the extensive evaluation of the energy of molecules on surfaces. Obtaining the binding energy of many configurations of large organic molecules requires a vast amount of computational time with density functional theory (DFT). Here, we use a graph neural network (GNN) to evaluate the adsorption energy of molecular species adsorbed on metallic surfaces. The GNN is trained on a set of C1–4 fragments including N, O, S heteroatoms and C6–10 aromatic rings. Compared to DFT, the GNN shows a mean absolute error (MAE) of 0.17 eV on the test set being 6 orders of magnitude faster. When applying the trained model with subsequent hyperparameter optimization to molecules of industrial interest (biomass, plastics and polyurethanes precursors) containing up to 22 carbon atoms, the prediction performance for the adsorption energy yields a MAE of 0.03 eV/(non-H atom). While the error for out-of-distribution molecules is higher, it is still within the acceptable limit for adsorption energies (0.05 eV/atom), confirming the viability of the approach. The proposed framework represents a potential tool for the fast screening of catalytic materials, as well as their inverse design, enabling the multi-scale modeling for systems that cannot be easily simulated by DFT.

Here we describe the design and applications of borindolizine, a novel scaffold with broadly tunable fluorescence and a high Stokes shift. Two classes of emitters were synthesized through rational scaffold modification, resulting in blue-emitting carboxyborindolizines (λ max,em = 431-459 nm) and green-emitting aryl borindolizines (λ max,em = 489-519 nm). Experimental structure-emission trends were used to validate a computational spectral prediction model and to subsequently design a red-emissive borindolizine scaffold. The red-emissive isoquinolyl borinidolizine was prepared, and the experimental emission (λ max,abs = 370 nm, λ max,em = 635 nm) was in excellent agreement with the theoretical emission (λ max,em = 646 nm). These results show how the application of data science can produce fluorophores with desirable spectroscopic properties through the borindolizine scaffold.

General chemical reaction conditions that achieve consistently high performance across multiple substrates are important for practical applications such as library synthesis and high-throughput experimentation. However, identifying such conditions efficiently has been a longstanding challenge, as it requires decision making under uncertainty with respect to both conditions and substrates, while minimizing the number of required experiments. Here, we introduce CurryBO, a high-level framework for generality-oriented optimization. By formalizing the problem as Bayesian optimization over curried functions, CurryBO provides a unified framework that accommodates different generality definitions (e.g., mean yield across substrates), and supports a range of substrate and condition selection strategies. We evaluate this framework on four benchmark tasks in experimental reaction optimization, and systematically analyze key algorithmic components. Our results show that efficient experiment planning can be achieved by emphasizing exploration when selecting reaction conditions, followed by the uncertainty-guided prioritization of substrates in a sequential decison-making scheme. Based on these insights, we design and validate an optimization policy that substantially improves sample efficiency relative to previously reported approaches across all benchmarks. Overall, the flexibility and modularity of CurryBO facilitate the integration of generality-oriented optimization into experimental settings, enabling more efficient identification of solutions that perform robustly across diverse tasks.

This study investigates the impact of incorporating stereochemical information, a crucial aspect of computational drug discovery and materials design, in molecular generative modeling. We present a detailed comparison of stereochemistry-aware and conventionally stereochemistry-unaware string-based generative approaches, utilizing both genetic algorithms and reinforcement learning-based techniques. To evaluate these models, we introduce novel benchmarks specifically designed to assess the importance of stereochemistry-aware generative modeling. Our results demonstrate that stereochemistry-aware models generally perform on par with or surpass conventional algorithms across various stereochemistry-sensitive tasks. However, we also observe that in scenarios where stereochemistry plays a less critical role, stereochemistry-aware models may face challenges due to the increased complexity of the chemical space they must navigate. This work provides insights into the trade-offs involved in incorporating stereochemical information in molecular generative models and offers guidance for selecting appropriate approaches based on specific application requirements.