Clémence Corminboeuf

Chemical datasets used in computational chemistry are often distributed across separate repositories, stored in dataset-specific formats, and accompanied by heterogeneous metadata, property definitions, and structure representations. This fragmentation increases the manual preprocessing required to reuse datasets for molecular property prediction, reaction modeling, benchmarking, and machine-learning workflows. Here, we present LCMD-db, a chemical data hosting and analysis platform for organizing molecular, fragment, and reaction datasets within a common framework. LCMD-db represents data as datasets and self-consistent subsets, with explicit metadata, property definitions, structure mappings, and standardized ingestion pipelines. The platform supports molecular structures, fragments, equilibrium geometries, transition states, reaction participants, and computed properties linked to SMILES strings, molecular graphs, and three-dimensional structure files. For reaction datasets, LCMD-db stores molecular participants together with their chemical roles and reaction-coordinate relationships, allowing intermediates, transition states, reactants, products, catalysts, and other species to be treated as connected chemical entities rather than isolated files. The platform provides interactive visualization, property inspection, filtering, export of selected entries, and Python-native access to structured molecular and reaction objects. We demonstrate LCMD-db using in-house and external datasets spanning small organic molecules, transition-metal complexes, molecular fragments, and reaction datasets with thermodynamic or kinetic annotations. LCMD-db connects data hosting with analysis and modeling workflows by providing structured, searchable, and directly loadable chemical datasets.

Here, we present AutomaticMolCraft, a locally deployed browser platform for running and organising 3D molecular discovery campaigns. It is designed to complement MolCraftDiffusion, a 3D molecular-generation platform that provides the computational layer for model training, 3D molecular generation, and analysis, while AutomaticMolCraft serves as the interactive interface for visualising molecular data and generated results. It also supports molecular data management while keeping data, computations, and results connected and traceable. The three applications are selected to showcase distinct capabilities of the interface across data-driven molecular discovery tasks. First, we demonstrate the data management and visualisation components by curating three organocatalyst datasets and comparing their molecular structures and properties to support the selection. Second, we show how these data components connect with the computational layer of MolCraftDiffusion by running and comparing property-directed generation workflows for producing 3D molecules with specified scalar properties. Third, we extend this integration to steric-property-directed generation, where AutomaticMolCraft is used to identify underrepresented regions of property space, generate molecules toward those regions, and incorporate the resulting candidates back into the molecular data for subsequent analysis. The codebases are available at: https://github.com/lcmd-epfl/MolCraftDiffusion and https://github.com/lcmd-epfl/AutomaticMolCraft.

Applying generative molecular design requires both chemical knowledge to turn a scientific aim into a computable objective and technical skill to configure and run the workflow that pursues it. Here, we present Solitarius , an architecture for automated computer-aided molecular discovery. It leverages agentic large language models (LLMs) for planning and configuring molecular generation workflows while delegating the underlying operations (e.g., model training, generating chemical structures, and optimisation) to a dedicated molecular generative AI platform. Solitarius further guides the LLM agents to use this platform as a human expert would, selecting appropriate procedures for a given objective and executing them accordingly. We first test whether LLM agents alone, without this guidance, are capable of reliably running the generative chemistry workflow, which exposes a gap in their ability to manage multi-step procedures. Equipped with both key generative algorithms and guidance, first, we apply Solitarius to the design of singlet-fission chromophores that satisfy the desired excited-state energetic criteria. We then use Solitarius to expand an organocatalyst dataset into sparsely populated regions of its property space. Across these examples, we demonstrate how Solitarius interprets scientific intent from plain-language goals, formulates scoring logic that reflects the desired objective, and iteratively executes molecular-generation workflows until the desired objectives are met. In the singlet-fission design example, Solitarius revises the planned design workflow as the search strategy evolves with the new results, shifting from open-ended de novo generation to structure-guided exploration using filters to avoid rediscovery and focus on the vicinity of the input molecules. For dataset expansion, Solitarius independently analyses the dataset to identify sparse regions. It then formulates an approach to populate them while preserving key catalytic structural motifs.

Applying generative molecular design requires both chemical knowledge to turn a scientific aim into a computable objective and technical skill to configure and run the workflow that pursues it. Here, we present Solitarius , an architecture for automated computer-aided molecular discovery. It leverages agentic large language models (LLMs) for planning and configuring molecular generation workflows while delegating the underlying operations (e.g., model training, generating chemical structures, and optimisation) to a dedicated molecular generative AI platform. Solitarius further guides the LLM agents to use this platform as a human expert would, selecting appropriate procedures for a given objective and executing them accordingly. We first test whether LLM agents alone, without this guidance, are capable of reliably running the generative chemistry workflow, which exposes a gap in their ability to manage multi-step procedures. Equipped with both key generative algorithms and guidance, first, we apply Solitarius to the design of singlet-fission chromophores that satisfy the desired excited-state energetic criteria. We then use Solitarius to expand an organocatalyst dataset into sparsely populated regions of its property space. Across these examples, we demonstrate how Solitarius interprets scientific intent from plain-language goals, formulates scoring logic that reflects the desired objective, and iteratively executes molecular-generation workflows until the desired objectives are met. In the singlet-fission design example, Solitarius revises the planned design workflow as the search strategy evolves with the new results, shifting from open-ended de novo generation to structure-guided exploration using filters to avoid rediscovery and focus on the vicinity of the input molecules. For dataset expansion, Solitarius independently analyses the dataset to identify sparse regions. It then formulates an approach to populate them while preserving key catalytic structural motifs.

By operating on full Cartesian geometries, three-dimensional (3D) molecular generative models offer direct control over the spatial arrangements dictating molecular properties. Yet generating structures that are both geometrically valid (reasonable bond lengths and angles, valid valencies, and minimal steric clashes) and practically useful, including charge neutrality and synthetic accessibility, remains challenging. Most existing models primarily target geometric correctness. Here, we present a diffusion-based framework for organic molecule generation that jointly addresses these criteria. The framework combines three complementary components: (i) a dataset construction strategy that includes positive examples, defined as geometrically valid, charge-neutral, and synthetically accessible molecules, together with negative examples containing geometric distortions, disconnected fragments, or non-zero net charges, annotated by scalar scores for distortion, fragment count, net charge, and synthetic accessibility; (ii) conditional training schemes that enable the model to learn both molecular structures and these associated attributes; and (iii) guidance mechanisms that use contrasting attributes during sampling to promote favorable features while suppressing undesirable ones. Across multiple 3D molecular datasets, the framework consistently improves geometric validity, charge neutrality, and synthetic accessibility, advancing 3D molecular generation toward more reliable use in practical organic molecular design.

Three-dimensional molecular generative models construct molecules by assigning atoms explicit Cartesian coordinates. This formulation allows the molecular generation process to be guided by geometric (e.g., steric effects, molecular shape) and physicochemical constraints toward desired objectives in a given chemical problem. The practical adoption of these models, however, remains limited by two key challenges: the high computational cost associated with model training as well as the lack of standardized platforms that allow models to be developed, evaluated, and compared. The landscape of 3D generative models for molecular design is fragmented, with most methods developed for specific molecular generation objectives. The implementations of these models are scattered across different repositories, software environments, and evaluation protocols, which impedes reproducibility and controlled comparison of different approaches and, generally, limits broader adoption. Here, we introduce MolCraftDiffusion, a modular and extensible platform for building, deploying, and evaluating 3D molecular diffusion models in computational chemistry. Its layered architecture decouples core training logic from model definitions and task implementations, enabling new generative architectures, guidance strategies, and evaluation metrics to be integrated. As part of the package, we implement curriculum learning, a progressive chemical complexity training approach used to construct a pretrained diffusion model on 3D molecular data sets compiled from multiple sources, circumventing the cost of full training in downstream applications. The framework provides a modular set of guidance mechanisms for directing generation toward chemically relevant objectives: structure control via molecular inpainting (systematic exploration of structural variants around a reference) and outpainting (extending molecules with new chemical groups), and property-conditioned generation via gradient-based and classifier-free approaches. We illustrate these capabilities across computational chemistry tasks, including virtual library construction and inverse molecular design. The extensibility of the platform is further demonstrated by porting three architecturally distinct models from the literature: TABASCO, ADiT, and ShEPhERD, each integrated without modifications to the core codebase. The codebase, pretrained models, and examples are available at: https://github.com/lcmd-epfl/MolCraftDiffusion and https://huggingface.co/pregH/MolecularDiffusion.

We introduce NaviCat ReaFS (reaction feature selection), a framework for testing and supporting multivariate linear regression models for reaction optimization in small, heterogeneous datasets, where overfitting risk is high and additional experiments are costly. ReaFS combines systematic benchmarking of feature selection methods with targeted diagnostics spanning data, feature-subsets, and predictions. We also introduce a composite score that captures complementary aspects of extrapolative performance, namely candidate recovery, predictive accuracy, and ranking quality, enabling direct comparison of models and feature selection strategies across homogeneous catalysis datasets. ReaFS further integrates cross-validation schemes, data and feature diagnostics, and robustness checks for prospective model assessment. Together, these components provide a practical route to more informed, low-data decisions in reaction optimization.

Physics-inspired machine learning (ML) models can be categorized into two classes: those relying solely on three-dimensional structure and those incorporating electronic information. In this work, we benchmark both classes for predicting quantum-chemical properties of transition metal complexes with diverse charge and spin states, using three complementary datasets. The evaluated methods include molecular representations (SLATM, FCHL, SOAP, and SPA^HM family) combined with kernel ridge regression, as well as geometric deep learning models (MACE and 3DMol). We examine how the inclusion of electronic information affects predictive accuracy across datasets and target properties. Models that incorporate electronic information consistently outperform purely structurebased models for properties whose distributions are strongly governed by electronic characters, such as spin-splitting energies and frontier orbital energies. In contrast, structure-only models perform well for predicting the HOMO-LUMO gap and dipole moment magnitude, whose distributions are relatively insensitive to electronic characteristics. Geometric deep learning models with charge and spin embeddings (MACE-QS and 3DMol-QS) show the highest overall accuracy, with 3DMol offering the best computational efficiency among the tested models. These results clarify when geometric information is sufficient and when incorporating electronic information becomes essential, providing practical guidance for selecting effective physics-based ML models for transition metal complexes.

Physics-inspired machine learning (ML) models can be categorized into two classes: those relying solely on three-dimensional structure and those incorporating electronic information. In this work, we benchmark both classes for predicting quantum-chemical properties of transition metal complexes with diverse charge and spin states, using three complementary datasets. The evaluated methods include molecular representations (SLATM, FCHL, SOAP, and SPA ^H M family) combined with kernel ridge regression, as well as geometric deep learning models (MACE and 3DMol). We examine how the inclusion of electronic information affects predictive accuracy across datasets and target properties. Models that incorporate electronic information consistently outperform purely structure-based models for properties whose distributions are strongly governed by electronic characters, such as spin-splitting energies and frontier orbital energies. In contrast, structure-only models perform well for predicting the HOMO–LUMO gap and dipole moment magnitude, whose distributions are relatively insensitive to electronic characteristics. Geometric deep learning models with charge and spin embeddings (MACE-QS and 3DMol-QS) show the highest overall accuracy, with 3DMol offering the best computational efficiency among the tested models. These results clarify when geometric information is sufficient and when incorporating electronic information becomes essential, providing practical guidance for selecting effective physics-based ML models for transition metal complexes.