In this work, we present a novel approach by fine-tuning Meta’s Llama3 Large Language Models (LLMs) to create SynLlama, which generates full synthetic pathways made of commonly accessible building blocks and robust organic reaction templates. SynLlama explores a large synthesizable space using significantly less data and offers strong performance in both forward and bottom-up synthesis planning compared to other state-of-the-art methods.

We find that SynLlama, even without training on external building blocks, can effectively generalize to unseen yet purchasable building blocks, meaning that its reconstruction capabilities extend to a broader synthesizable chemical space than those of the training data.

We also demonstrate the use of SynLlama in a pharmaceutical context for synthesis planning of analog molecules and hit expansion leads for proposed inhibitors of target proteins, offering medicinal chemists a valuable tool for discovery.

In this review, we examine the prominent ML approaches used to tackle the imbalanced data challenge in different areas of chemistry, including resampling techniques, data augmentation techniques, algorithmic approaches, and feature engineering strategies.

Each of these methods is evaluated in the context of its application across various aspects of chemistry, such as drug discovery, materials science, cheminformatics, and catalysis. We also explore future directions for overcoming the imbalanced data challenge and emphasize data augmentation via physical models, large language models (LLMs), and advanced mathematics.

The benefit of balanced data in new material design and production and the persistent challenges are discussed. Overall, this review aims to elucidate the prevalent ML techniques applied to mitigate the impacts of imbalanced data within the field of chemistry and offer insights into future directions for research and application.

Predicting the stereoselectivity of glycosylations is a major challenge in carbohydrate chemistry. Herein we show that it is possible to build machine learning models that can predict the major anomer of a glycosylation, whether the other anomer is observed as the minor product, and the anomeric ratio of the two anomers. The three models are integrated into a publicly available tool, GlycoPredictor.

From a statistical analysis of literature data, we analyze glycosylation trends and compare them to known trends in the field of carbohydrate chemistry, making it possible to elucidate a hierarchy of rules governing the stereoselectivity of glycosylations and discover promising new trends that complement expert intuition, which are tested in novel glycosylation methods.

In this work, we introduce SubGrapher, a method for the visual fingerprinting of molecule and Markush structure images. Unlike conventional Optical Chemical Structure Recognition (OCSR) models that attempt to reconstruct full molecular graphs, SubGrapher focuses on extracting fingerprints directly from images.

Using learning-based instance segmentation, SubGrapher identifies functional groups and carbon backbones, constructing a substructure-based fingerprint that enables the retrieval of molecules and Markush structures.

Our approach is evaluated against state-of-the-art OCSR and fingerprinting methods, demonstrating superior retrieval performance and robustness across diverse molecule and Markush structure depictions. The benchmark datasets, models, and inference code are publicly available.

We propose a novel framework that leverages large language models (LLMs) to transform synthetically infeasible inorganic crystal structures into synthetically feasible ones. Unlike previous studies on synthesis predictions, which focus primarily on estimating synthesizability, our method provides actionable solutions for redesigning unsynthesizable materials into synthesizable ones. By integrating an invertible structural representation and an iterative fine-tuning strategy, our framework not only predicts synthetic feasibility but also modifies unsynthesizable materials into viable candidates.

As a result, we demonstrate that LLMs can effectively modify materials of various types, enhancing their synthesizability and increasing the likelihood of successful synthesis. As an indirect experimental validation, we demonstrate that 34 materials among the top 100 redesigned (but originally unsynthesizable) structures have indeed been experimentally reported in the literature. This approach addresses a critical gap between design and synthesis in materials science, and enables the discovery of experimentally realizable compounds by employing the “learn-and-regenerate” strategy in LLMs.

This study introduces a data-driven framework that combines DFT calculations with machine learning to facilitate accurate and scalable predictions of ground- and excited-state redox potentials for iridium(III) photocatalysts.

We first constructed independent models to identify key geometric and electronic descriptors governing redox behavior. Shapley additive explanations-based analyses revealed clear structure–activity relationships, offering mechanistic insights and rational guidance for tuning redox potentials. Based on these insights, we developed unified multi-output models—Model G for ground-state and Model E for excited-state redox potentials—to enable rapid, cost-effective, and high-throughput predictions. By modeling oxidation and reduction processes within a shared descriptor space, we can reduce computational overhead while maintaining high predictive accuracy.

To assess cross-metal generalizability, residual transfer learning was applied to osmium (Os) photocatalysts. Using feature-similar complexes, the resulting transfer models (G-T, E-T) achieved performance comparable to Os-only baselines, demonstrating efficient few-shot cross-metal transfer. Collectively, this study establishes an interpretable and transferable machine-learning framework for photocatalyst discovery. This framework provides a foundation for large-scale screening and rational design across diverse transition-metal platforms, accelerating advancements in photoredox catalysis, solar fuel production, and broader sustainable energy technologies.

Boryl radicals have become indispensable in organic synthesis, yet, translating their complex steric and electronic properties into actionable reactivity insights remains challenging. Herein, we present a comprehensive classification of boryl radicals, including a publicly accessible database of 141 neutral 7e-4c boryl radicals, each parametrized by a set of electronic and steric features derived from DFT calculations.

Unsupervised machine learning (k-means clustering) and dimensionality reduction (PCA/UMAP) condense this high dimensional descriptor space into the “B-rad map”, capturing trends in sterics and electronics among the resulting five clusters. Global electrophilicity (ω) and nucleophilicity (N) indices are overlaid to create a polarity‑annotated guide, while DFT‑computed activation free energies for six benchmark reactions (HAT, radical addition, and XAT for two different substrates) yield the React‑B‑rad maps that directly link intrinsic properties to specific reaction performance. To demonstrate predictive power, supervised machine learning models (random forest) are trained on the descriptors and successfully predict radical reactivity regimes across all reaction types.

Overall, this integrated, machine-learning-driven platform can serve as both a practical guide for experimental decision-making and a foundation for data-driven discovery, paving the way towards rational design and virtual screening of boryl-radical reagents for diverse synthetic applications.

Here we propose a latent variable approach that bridges the gap between ligand-only data and protein–ligand complexes, enabling target-aware generative models to explore a broader chemical space, thereby enhancing the quality of molecular generation. Inspired by quantum molecular simulations, we introduce ECloudGen, a generative model that leverages electron clouds as meaningful latent variables.

ECloudGen incorporates techniques such as latent diffusion models, Llama architectures and a contrastive learning task, which organizes the chemical space into a structured and highly interpretable latent representation.

Benchmark studies demonstrate that ECloudGen outperforms state-of-the-art methods by generating more potent binders with superior physiochemical properties and by covering a broader chemical space.

The increasing complexity in designing nanostructured materials for electronics, biomedicine, and energy applications requires advanced computational methods to enhance research efficiency and minimize experimental costs. This study proposes an innovative agent-based retrieval-augmented generation (RAG) system integrated with large language models (LLMs) to automate the extraction and analysis of scientific information from extensive literature databases, specifically targeting nanostructured materials developed via two-photon polymerization (2PP). In addition to extracting and analyzing scientific data, our approach emphasizes understanding how these nanostructured materials interact with cells, which is crucial for controlling their application in biomedicine.

The developed platform demonstrates robust semantic accuracy (cosine similarity: 0.82) and high overall task precision (0.81), significantly reducing the likelihood of misinformation by incorporating dynamic query refinement mechanisms. The intuitive, user-friendly interface facilitates quick access to relevant scientific data, thereby improving researchers’ productivity and enabling more accurate experimental planning. Although the system exhibits certain limitations regarding domain-specific terminology coverage, further fine-tuning and specialized training are anticipated to enhance its performance and reliability for advanced scientific applications.

Polar organic cocrystals hold significant promise for various advanced technological applications. However, their relatively low occurrence emphasizes the difficulties in achieving the desired polar packing arrangements, making their discovery complex and challenging.

Here, we introduce a data-driven method that combines machine learning (ML) with high-throughput (HT) automation to speed up the discovery of polar organic cocrystals. Using ML techniques, we identified key factors that influence polar cocrystal formation, allowing for targeted selection of molecular candidates. We examined 13 cocrystal combinations with chloranilic acid (CA), screening 20 solvent systems for each, which enabled a highly efficient search across a broad chemical space. HT automation further enhanced the synthesis and characterization by enabling rapid screening and precise structural validation, while thoroughly exploring the chemical landscape. Experimental results confirmed 13 pairs of CA cocrystals, with 6 crystallizing in polar space groups, resulting in a polar discovery rate of 46%-nearly three times higher than the average in the Cambridge Structural Database (CSD) (∼13.2%). This integrated approach offers a new strategy in polar organic cocrystal research. The findings demonstrate the potential of this method to advance functional molecular materials and pave the way for next-generation applications using polar organic cocrystals.

Artificial intelligence (AI) is transforming research in metal–organic frameworks (MOFs), where models trained on structured computational data routinely predict new materials and optimize their properties. This raises a central question: What if we could leverage the full breadth of MOF knowledge, not just structured data sets, but also the scientific literature? For researchers, the literature remains the primary source of knowledge, yet much of its content, including experimental data and expert insight, remains underutilized by AI systems.

We introduce MOF-ChemUnity, a structured, extensible, and scalable knowledge graph that unifies MOF data by linking literature-derived insights to crystal structures and computational data sets. By disambiguating MOF names in the literature and connecting them to crystal structures in the Cambridge Structural Database, MOF-ChemUnity unifies experimental and computational sources and enables cross-document knowledge extraction and linking. We showcase how this enables multiproperty machine learning across simulated and experimental data, compilation of complete synthesis records for individual compounds by aggregating information across multiple publications, and expert-guided materials recommendations via structure-based machine learning descriptors for pore geometry and chemistry. When used as a knowledge source to augment large language models (LLMs), MOF-ChemUnity enables a literature-informed AI assistant that operates over the full scope of MOF knowledge. Expert evaluations show improved accuracy, interpretability, and trustworthiness across tasks such as retrieval, inference of structure–property relationships, and materials recommendation, outperforming standard LLMs. This work lays the foundation for literature-informed materials discovery, enabling both scientists and AI systems to reason over the full existing knowledge in a new way.

As a major burden on global healthcare systems, adverse drug events (ADEs) result in significant morbidity, mortality, and healthcare resource consumption. With the rapid advances in precision medicine, personalized ADEs and their molecular mechanisms are important components of drug repurposing and drug safety improvement. Thus, extensive studies have been conducted to collect valuable information on personalized ADEs, but no database has yet been available to provide such data.

In this work, PersADE, a database aiming to provide personalized drug adverse events and their molecular mechanisms, was constructed. It integrated 4 061 772 personalized drug-ADE associations, 31 756 protein-ADE associations, and 108 677 drug-protein interactions, with a particular emphasis on off-target effects.

The uniqueness of these data lies in (a) providing demographic characteristics, disease context and drug administration parameters associated with ADEs, enabling stratification of drug-ADE associations; (b) systematically integrating interactions among drugs, human proteins and ADEs, describing the mechanistic insights. Given the growing global focus on precision medicine, PersADE is highly anticipated to significantly impact studies on personalized ADEs and mechanistic explorations by providing researchers and clinicians with evidence-based tools. It is now freely accessible at: https://idrblab.org/PersADE

A domain-specific large language model, MOFs-LLM, is developed to accelerate the inverse design and synthesis of metal—organic frameworks (MOFs) for hydrogen storage. Trained on 210 million tokens derived from over 6 000 MOF-related publications and 15 000 crystal structures, the model integrates chemical knowledge with structural features to improve structure–property reasoning. Compared to baseline methods, MOFs-LLM achieves a 46.7% enhancement in capturing structure–property relationships. It enables the inverse design of 60 candidate frameworks optimized for both hydrogen storage performance and synthetic accessibility.

Guided by the model, a novel MOF (Cu-LLMs-1) was synthesized in three experimental iterations, exhibiting a hydrogen uptake of 1.33 wt% at room temperature, ranking among the top five pure MOFs under comparable conditions. These findings highlight the potential of domain-trained language models to bridge virtual screening and experimental realization in materials discovery.

We developed MGDB, a specialized open-access repository integrating rigorously curated multidimensional data for MGs. MGDB contains 7396 curated MGs being sourced from 162 peer-reviewed publications and 156 patents. It consolidates structural data, 9728 experimental bioactivity data points (covering degradation efficiency, binding affinity, cellular/animal activity) across 201 targets and 108 effectors, 115 296 computed physicochemical properties, and 270 785 ADMET profiles.

The database supports text-based and chemical structure-based queries and interoperability with external resources (e.g. PubChem, ChEMBL, DrugBank, UniProt, and WIPO) via hyperlinks.

By centralizing and standardizing specialized MG information, MGDB empowers researchers to rapidly explore MG research landscapes and provides high-quality datasets for artificial intelligence-driven rational therapeutic design. MGDB is freely available at http://mgdb.idruglab.cn/.

We provide an overview of core molSimplify functionality and recent updates that enhance its capabilities for automated molecular and materials modeling. We describe the mol3D and atom3D classes, which store atomic and bonding information for a wide range of functions, including reading, modifying, and characterizing molecular geometries from common file formats. Enhancements to decoration and substructure addition functions enable systematic derivatization of template molecules.

We introduce a new mol2D class that enables graph-based uniqueness checks and substructure identification. Most importantly, we introduce improvements to transition metal complex (TMC) generation that eliminate steric clashes and enable structure building with ligands of higher denticity. Integration with machine learning models that predict coordinating atom identities enables truly high-throughput, de novo TMC generation.

We describe applications of molSimplify outside of isolated TMCs, including extensions to periodic systems (i.e., particularly metal–organic frameworks) and to metalloenzymes through the protein3D class. We demonstrate our improved combined structure prediction and generation workflow by generating structures of a database of experimentally characterized Ir complexes from only the SMILES strings of their respective ligands.

We envision that recent enhancements will make the code easily extendible to other periodic materials such as covalent organic frameworks and zeolites or to multimetallic transition metal complexes.

Validating chemical synthesis success requires confirming the desired product using various analytical techniques. While spectroscopic data collection is increasingly automated, interpreting results remains a major bottleneck, often requiring expert input. With advances in laboratory automation and high-throughput synthesis, this challenge is expected to intensify.

We introduce the MultiModalSpectralTransformer (MMST), a machine learning method that predicts chemical structures directly from diverse spectral data (NMR, IR, and MS). Trained on 4 million simulated compounds, MMST achieves 72% and 80% as top-1 and top-3 accuracy, respectively. To address out-of-distribution challenges, we implemented an active learning improvement cycle that generates molecules in similar chemical spaces, enabling the model to adapt to chemical structures beyond its original training data. We demonstrate MMST's capabilities through comprehensive benchmarking across diverse molecular weight ranges and chemical spaces. Notably, despite training solely on simulated data, MMST demonstrates good performance with experimental spectra. This research represents a significant advancement in automated structure elucidation, offering a powerful and adaptable tool that bridges the gap between simulated and real-world data.

The disconnect between AI-generated molecules with desirable properties and their synthetic feasibility remains a critical bottleneck in computational discovery of drugs and materials. While generative AI has accelerated the proposal of candidate molecules, many of these structures prove challenging or impossible to synthesize using established chemical reactions.

Here, we introduce SynTwins, a novel retrosynthesis-guided molecule design framework that finds synthetically accessible molecular analogs by emulating expert chemists' strategies in three steps: retrosynthesis, searching similar building blocks, and virtual synthesis. Using a search algorithm instead of a stochastic data-driven generator, SynTwins outperforms state-of-the-art machine learning models at exploring synthetically accessible analogs while maintaining high structural similarity to original target molecules. Furthermore, when integrated into existing molecular property-optimization frameworks, our hybrid approach produces synthetically feasible analogs with minimal loss in property scores.

Our comprehensive benchmarking across diverse molecular datasets demonstrates that SynTwins effectively bridges the gap between computational design and experimental synthesis, providing a practical solution for accelerating the discovery of synthesizable molecules with desired properties for a wide range of applications.

The application of machine learning approaches to meaningful problems in chemistry and materials science is still challenged by the limited availability of data. In order to close this gap, we report the tmQMg* data set, which provides excited state properties for 74k mononuclear transition metal complexes extracted from the Cambridge Structural Database. All properties were computed at the TD-DFT ωB97xd/def2SVP level of theory. The strongest electron excitations in the ultraviolet, visible, and near-infrared ranges are included, together with the wavelengths and intensities of the first 30 excited states.

Further, natural transition orbitals were computed for the strongest excitations in the visible range to determine the nature of the associated charge transfers. By computing the TD-DFT spectra in both gas phase and acetone, we quantified solvatochromic effects, which are also provided with the data set, in terms of both wavelength shifts and intensity changes.

The tmQMg* data set will enable the development of discriminative and generative artificial intelligence models with respect to absorption spectra, charge transfer character, and solvatochromism, enabling novel advances in the field of transition metal photochemistry.

In this paper, we present the results of training machine learning (ML) models for accurate prediction of several key photophysical characteristics (absorption maximum wavelength, molar absorption coefficient, emission maximum wavelength, fluorescence quantum yield and lifetime, singlet oxygen generation quantum yield) for BODIPYs. ML models were trained using experimental data comprising more than 35,000 records for the predicted parameters. Particular emphasis was placed on model interpretability and on accounting for the solvent nature effect on the predicted photophysical parameters. To ensure open data access and a user-friendly interface, all developed models were integrated into the created web tool SpecML (http://specml.isc-ras.ru/).

SpecML allows prediction of photophysical parameters for individual BODIPY molecules and the screening of entire series of BODIPYs. We believe that our created SpecML web tool will become an effective resource for accelerating the rational design of BODIPYs with desired photophysical properties and will be useful for a wide range of researchers in the fields of photonics, organic electronics, and molecular design.