Machine learning interatomic potentials (MLIPs) promise quantum‐level accuracy at classical force field speeds, but their performance hinges on the quality and diversity of training data. An efficient and fully automated approach to sample chemical reaction space without relying on human intuition, addressing a critical gap in MLIP development is presented. The method combines the speed of tight‐binding calculations with selective high‐level refinement, generating diverse datasets that capture both equilibrium and reactive regions of potential energy surfaces. By employing single‐ended growing string and nudged elastic band methods, reaction pathways previously underrepresented in MLIP training sets, particularly near transition states are systematically explored. This approach yields datasets with rich structural and chemical diversity, essential for robust MLIP development. Open‐source code is provided for the entire workflow, facilitating the integration of the approach into existing MLIP development pipelines.
Artificial intelligence (AI) is reshaping biomedical research by providing scalable computational frameworks suited to the complexity of biological systems. Central to this revolution are bio/chemical language models, including large language models, which are reconceptualizing molecular structures as a form of ‘language’ amenable to advanced computational techniques. Here we critically examine the role of these models in biology and chemistry, tracing their evolution from molecular representation to molecular generation and optimization. This review covers key molecular representation strategies for both biological macromolecules and small organic compounds—ranging from protein and nucleotide sequences to single-cell data, string-based chemical formats, graph-based encodings and three-dimensional point clouds—highlighting their respective advantages and inherent limitations in AI applications. The discussion further explores core model architectures, such as bidirectional encoder representations from transformers-like encoders, generative pretrained transformer-like decoders and encoder–decoder transformers, alongside their sophisticated pretraining strategies such as self-supervised learning, multitask learning and retrieval-augmented generation. Key biomedical applications, spanning protein structure and function prediction, de novo protein design, genomic analysis, molecular property prediction, de novo molecular design, reaction prediction and retrosynthesis, are explored through representative studies and emerging trends. Finally, the review considers the emerging landscape of agentic and interactive AI systems, showcasing briefly their potential to automate and accelerate scientific discovery while addressing critical technical, ethical and regulatory considerations that will shape the future trajectory of AI in biomedicine.
Machine learning interatomic potentials (MLIPs) promise to revolutionize computational chemistry; however, their performance depends critically on the quality and diversity of the training data. Existing quantum chemical datasets predominantly focus on equilibrium structures and exhibit limited halogen coverage, despite halogens being present in approximately 25% of pharmaceuticals and numerous materials. We present Halo8, a comprehensive dataset that addresses this gap by systematically incorporating fluorine, chlorine, and bromine chemistry into reaction pathway sampling. Using our efficient multi-level computational workflow, which achieves a 110-fold speedup over pure DFT approaches, Halo8 comprises approximately 20 million quantum chemical calculations from 19,000 unique reaction pathways. The dataset combines recalculated Transition1x reactions with new halogen-containing molecules from GDB-13, employing systematic halogen substitution to maximize chemical diversity. All calculations were performed at the ω B97X-3c level, providing accurate energies, forces, dipole moments, and partial charges. Validation demonstrates that Halo8 captures diverse structural distortions and chemical environments essential for reactive systems, serving as a valuable resource for training MLIPs applicable to pharmaceutical discovery, materials design, and catalysis.
We present a fluorescence-oriented molecular design strategy that integrates multiple generative models within a chemically constrained and photophysically relevant subspace. Rather than relying on large-scale data sets or model-specific tuning, our approach builds on a curated training set constructed exclusively from atom-in-SMILES (AIS) fragments spanning chromophore-like molecules. The pipeline accommodates reinforcement learning–based ReLeaSE, which employs AIS-based tokenization, MolDQN, which utilizes a graph-based molecular representation, and the combinatorial optimization–based MolFinder, which operates on SMILES strings, each guided by a semiheuristic objective function targeting excitation energy, oscillator strength, and fluorophore similarity. Despite the compact training corpus, the pipeline yields structurally novel and optically active molecules, validated via QM calculations at the sTDA/CAM-B3LYP level. Our findings demonstrate that strategic data curation and property-driven integration can enable general-purpose models to succeed in specialized discovery tasks such as de novo fluorophore design.
SMILES is the most dominant molecular representation used in AI-based chemical applications, but it has innate limitations associated with its internal structure. Here, we exploit the idea that a set of structural fingerprints can be used as efficient alternatives to unique molecular representations. For this purpose, we trained neural-machine-translation based models that translate a set of various structural fingerprints to conventional text-based molecular representations, i.e., SMILES and SELFIES. The assessment of their conversion efficiency showed that our models successfully reconstructed molecules and achieved a high level of accuracy. Therefore, our approach brings structural fingerprints into play as strong representational tools in chemical natural language processing applications by restoring the connectivity information that is lost during fingerprint transformation. This comprehensive study addressed the major limitation of structural fingerprints, which precludes their implementation in NLP models. Our findings would facilitate the development of text or fingerprint-based chemoinformatic models for generative and translational tasks.
The inherent complexity of mass spectra and the lack of direct correlation between spectral and structural similarities retards structure elucidation and accurate peak annotation. Our methodology employs modified atomic environments from topological radii zero to represent collections of annotated spectral peaks. Rather than aiming for de novo structure prediction from spectra, our objective is to refine and re-rank candidate structures retrieved by existing library search methods using predicted atom environments. We conducted a multi-step complexity reduction to mass-to-fragment mappings and trained the Transformer model to predict the atomic environments of compounds directly from mass and intensity data, achieving a peak precision of 86.1% and a recall rate of 78.4% on the test set. This novel framework not only aids in interpreting EI-MS data by providing insights into structural contents but also refines cosine similarity rankings by suggesting the inclusion or exclusion of certain atomic environments. Our findings over the NIST database suggest that our approach complements conventional methods by improving spectra matching through an in-depth atomic-level analysis.