A graphical abstract is available for this content

The effectiveness of antibiotics is greatly enhanced by their ability to target invasive organisms involved in the ancient evolutionary battle between hosts and pathogens. Conventional antibiotics no longer offer adequate protection due to the evolution of strategies to evade them. As a result, efforts are needed to design novel replacement antibiotics, making them unique from most other forms of drug development. As drug discovery costs have steadily increased along with the need for novel antibiotics, the interest in antimicrobial peptides (AMPs) as alternative antimicrobial treatments has grown in recent years. As a complement to experimental high-throughput screening, computational methods have become essential in hit and lead discovery in pharmaceutical research. It may be possible to access unexplored chemical space with customized virtual compound libraries. It has been questioned whether screening billions of molecules virtually with the risk of false positives is practical despite their unlimited potential size. In terms of finding novel chemical compounds capable of solving many global problems, machine learning, deep learning, and generative models hold significant promise. It is anticipated that the current challenges and limitations about the applicability of the stated approaches will be overcome in the coming years. However, plenty of advances are still required to achieve their full potential. In this perspective, we review the previous and ongoing work based on the latest scientific breakthroughs and technologies that could offer new opportunities and alternative strategies for developing novel AMPs.

We present a machine-learning approach toward predicting spectroscopic constants based on atomic properties. After collecting spectroscopic information on diatomics and generating an extensive database, we employ Gaussian process regression to identify the most efficient characterization of molecules to predict the equilibrium distance, vibrational harmonic frequency, and dissociation energy. As a result, we show that it is possible to predict the equilibrium distance with an absolute error of 0.04 Å and vibrational harmonic frequency with an absolute error of 36 cm⁻¹, including only atomic properties. These results can be improved by including prior information on molecular properties leading to an absolute error of 0.02 Å and 28 cm⁻¹ for the equilibrium distance and vibrational harmonic frequency, respectively. In contrast, the dissociation energy is predicted with an absolute error ≲0.4 eV. Alongside these results, we prove that it is possible to predict spectroscopic constants of homonuclear molecules from the atomic and molecular properties of heteronuclears. Finally, based on our results, we present a new way to classify diatomic molecules beyond chemical bond properties.

The BigSMILES notation, a concise tool for polymer ensemble representation, is augmented here by introducing an enhanced version called generative BigSMILES. G-BigSMILES is designed for generative workflows, and is complemented by tailored software tools for ease of use. This extension integrates additional data, including reactivity ratios (or connection probabilities among repeat units), molecular weight distributions, and ensemble size. An algorithm, interpretable as a generative graph is devised that utilizes these data, enabling molecule generation from defined polymer ensembles. Consequently, the G-BigSMILES notation allows for efficient specification of complex molecular ensembles via a streamlined line notation, thereby providing a foundational tool for automated polymeric materials design. In addition, the graph interpretation of the G-BigSMILES notation sets the stage for robust machine learning methods capable of encapsulating intricate polymeric ensembles. The combination of G-BigSMILES with advanced machine learning techniques will facilitate straightforward property determination and in silico polymeric material synthesis automation. This integration has the potential to significantly accelerate materials design processes and advance the field of polymer science.

A graphical abstract is available for this content

In light of the pressing need for practical materials and molecular solutions to renewable energy and health problems, to name just two examples, one wonders how to accelerate research and development in the chemical sciences, so as to address the time it takes to bring materials from initial discovery to commercialization. Artificial intelligence (AI)-based techniques, in particular, are having a transformative and accelerating impact on many if not most, technological domains. To shed light on these questions, the authors and participants gathered in person for the ASLLA Symposium on the theme of ‘Accelerated Chemical Science with AI’ at Gangneung, Republic of Korea. We present the findings, ideas, comments, and often contentious opinions expressed during four panel discussions related to the respective general topics: ‘Data’, ‘New applications’, ‘Machine learning algorithms’, and ‘Education’. All discussions were recorded, transcribed into text using Open AI's Whisper, and summarized using LG AI Research's EXAONE LLM, followed by revision by all authors. For the broader benefit of current researchers, educators in higher education, and academic bodies such as associations, publishers, librarians, and companies, we provide chemistry-specific recommendations and summarize the resulting conclusions.

In committee of experts strategies, small datasets are extracted from a larger one and utilised for the training of multiple models. These models' predictions are then carefully weighted so as to obtain estimates which are dominated by the model(s) that are most informed in each domain of the data manifold. Here, we show how this divide-and-conquer philosophy provides an avenue in the making of machine learning potentials for atomistic systems, which is general across systems of different natures and efficiently scalable by construction. We benchmark this approach on various datasets and demonstrate that divide-and-conquer linear potentials are more accurate than their single model counterparts, while incurring little to no extra computational cost.

Artificial intelligence (AI) contributes new methods for designing compounds in drug discovery, ranging from de novo design models suggesting new molecular structures or optimizing existing leads to predictive models evaluating their toxicological properties. However, a limiting factor for the effectiveness of AI methods in drug discovery is the lack of access to high-quality data sets leading to a focus on approaches optimizing data generation. Combinatorial library design is a popular approach for bioactivity testing as a large number of molecules can be synthesized from a limited number of building blocks. We propose a framework for designing combinatorial libraries using a molecular generative model to generate building blocks de novo, followed by using k-determinantal point processes and Gibbs sampling to optimize a selection from the generated blocks. We explore optimization of biological activity, Quantitative Estimate of Drug-likeness (QED) and diversity and the trade-offs between them, both in single-objective and in multi-objective library design settings. Using retrosynthesis models to estimate building block availability, the proposed framework is able to explore the prospective benefit from expanding a stock of available building blocks by synthesis or by purchasing the preferred building blocks before designing a library. In simulation experiments with building block collections from all available commercial vendors near-optimal libraries could be found without synthesis of additional building blocks; in other simulation experiments we showed that even one synthesis step to increase the number of available building blocks could improve library designs when starting with an in-house building block collection of reasonable size.

Despite its fundamental importance and widespread use for assessing reaction success in organic chemistry, deducing chemical structures from nuclear magnetic resonance (NMR) measurements has remained largely manual and time consuming. To keep up with the accelerated pace of automated synthesis in self driving laboratory settings, robust computational algorithms are needed to rapidly perform structure elucidations. We analyse the effectiveness of solving the NMR spectra matching task encountered in this inverse structure elucidation problem by systematically constraining the chemical search space, and correspondingly reducing the ambiguity of the matching task. Numerical evidence collected for the twenty most common stoichiometries in the QM9-NMR database indicate systematic trends of more permissible machine learning prediction errors in constrained search spaces. Results suggest that compounds with multiple heteroatoms are harder to characterize than others. Extending QM9 by ∼10 times more constitutional isomers with 3D structures generated by Surge, ETKDG and CREST, we used ML models of chemical shifts trained on the QM9-NMR data to test the spectra matching algorithms. Combining both ¹³C and ¹H shifts in the matching process suggests twice as permissible machine learning prediction errors than for matching based on ¹³C shifts alone. Performance curves demonstrate that reducing ambiguity and search space can decrease machine learning training data needs by orders of magnitude.

Hansen solubility parameters (HSPs) have three components, δ_d, δ_p and δ_h, accounting for dispersion forces, polar forces, and hydrogen bonding of a molecule, which were designed to better understand how molecular structure affects miscibility/solubility. HSP is widely used throughout the pipeline of pharmaceutical research and yet has not been as well studied computationally as the aqueous solubility. In the current study, we predicted HSPs using only the SMILES of molecules and utilise the molecular embedding approach inspired by Natural Language Processing (NLP). Two pre-trained deep learning models – Mol2Vec and ChemBERTa have been used to derive the embeddings. A dataset of ∼1200 organic molecules with experimentally determined HSPs was used as the labelled dataset. Upon finetuning, the ChemBERTa model “learned” relevant molecular features and shifted attention to functional groups that give rise to the relevant HSPs. The finetuned ChemBERTa model outperforms both the Mol2Vec model and the baseline Morgan fingerprint method albeit not to a significant extent. Interestingly, the embedding models can predict δ_d significantly better than δ_h and δ_p and overall, the accuracy of predicted HSPs is lower than the well-benchmarked ESOL aqueous solubility. Our study indicates that the extent of transfer learning leveraged from the pre-trained models is related to the labelled molecular properties. It also highlights how δ_p and δ_h may have large intrinsic errors in the way they are defined and therefore introduces inherent limitations to their accurate prediction using machine learning models. Our work reveals several interesting findings that will help explore the potential of BERT-based models for molecular property prediction. It may also guide the possible refinement of the Hansen solubility framework, which will generate a wide impact across the pharmaceutical industry and research.

Availability of material datasets through high performance computing has enabled the use of machine learning to not only discover correlations and employ materials informatics to perform screening, but also to take the first steps towards materials by design. Computational materials databases are well-labelled and provide a fertile ground for predicting both ground-state and functional properties of materials. However, a clear design approach that allows prediction of materials with the desired functional performance does not yet exist. In this work, we train various machine learning models on a dataset curated from a combination of Materials Project as well as computationally calculated thermoelectric electronic power factor using a constant relaxation time Boltzmann transport equation (BoltzTrap). We show that simple random forest-based machine learning models outperform more complex neural network-based approaches on the moderately sized dataset and also allow for interpretability. In addition, when trained on only cubic material systems, the best performing machine learning model employs a perturbative scanning approach to find new candidates in Materials Project that it has never seen before, and automatically converges upon half-Heusler alloys as promising thermoelectric materials. We validate this prediction by performing density functional theory and BoltzTrap calculations to reveal accurate matching. One of those predicted to be a good material, NbFeSb, has been studied recently by the thermoelectric community; from this study, we propose four new half-Heusler compounds as promising thermoelectric materials – TiGePt, ZrInAu, ZrSiPd and ZrSiPt. Our approach is generalizable to extrapolate into previously unexplored material spaces and establishes an automated pipeline for the development of high-throughput functional materials.

Metabolomics data analysis for phenotype identification commonly reveals only a small set of biochemical markers, often containing overlapping metabolites for individual phenotypes. Differentiation between distinctive sample groups requires understanding the underlying causes of metabolic changes. However, combining biomarker data with knowledge of metabolic conversions from pathway databases is still a time-consuming process due to their scattered availability. Here, we integrate several resources through ontological linking into one unweighted, directed, labeled bipartite property graph database for human metabolic reactions: the Directed Small Moleicules Network (DSMN). This approach resolves several issues currently experienced in metabolic graph modeling and data visualization for metabolomics data, by generating (sub)networks of explainable biochemical relationships. Three datasets measuring human biomarkers for healthy aging were used to validate the results from shortest path calculations on the biochemical reactions captured in the DSMN. The DSMN is a fast solution to find and visualize biological pathways relevant to sparse metabolomics datasets. The generic nature of this approach opens up the possibility to integrate other omics data, such as proteomics and transcriptomics.

The quest for accurate and efficient Machine Learning (ML) models to predict complex molecular properties has driven the development of new quantum-inspired representations (QIR). This study introduces MODA (Molecular Orbital Decomposition and Aggregation), a novel QIR-class descriptor with enhanced predictive capabilities. By incorporating wave-function information, MODA is able to capture electronic structure intricacies, providing deeper chemical insight and improving performance in unsupervised and supervised learning tasks. Specially designed to be separable, the multi-moiety regularization technique unlocks the predictive power of MODA for both intra- and intermolecular properties, making it the first QIR-class descriptor capable of such distinction. We demonstrate that MODA shows the best performance for intermolecular magnetic exchange coupling (J_AB) predictions among the descriptors tested herein. By offering a versatile solution to address both intra- and intermolecular properties, MODA showcases the potential of quantum-inspired descriptors to improve the predictive capabilities of ML-based methods in computational chemistry and materials discovery.

Reproducible data and results underpin the credibility and integrity of research findings across the sciences. However, experiments and measurements conducted across laboratories, or by different researchers, are often hindered by incomplete or inaccessible procedural data. Additionally, the time and resources needed to manually perform repeat experiments and analyses limit the scale at which experiments can be reproduced. Both improved methods for recording and sharing experimental procedures in machine-readable formats and efforts towards automation can be beneficial to circumvent these issues. Here we report the development of ExpFlow, a data collection, sharing, and reporting software currently customized for electrochemical experiments. The ExpFlow software allows researchers to systematically encode laboratory procedures through a graphical user interface that operates like a fill-in-the-blank laboratory notebook. Built-in calculators automatically derive properties such as diffusion coefficient and charge-transfer rate constant from uploaded data. Further, we deploy ExpFlow procedures with robotic hardware and software to perform cyclic voltammetry (CV) experiments in triplicate for eight well-known electroactive systems. The resulting oxidation potentials and diffusion coefficients are consistent with literature-reported values, validating our approach and demonstrating the utility of robotic experimentation in promoting reproducibility. Ultimately, these tools enable automated and (semi)autonomous cyclic voltammetry experiments and measurements that will facilitate high-throughput experimentation, reproducibility, and eventually data-driven electrochemical discovery.

The identification of a compound's chemical structure remains one of the most crucial everyday tasks in chemistry. Among the vast range of existing analytical techniques NMR spectroscopy remains one of the most powerful tools. As a step towards structure prediction from experimental NMR spectra, this article introduces a novel machine-learning (ML) Structure Seer model that is designed to provide a quantitative probabilistic prediction on the connectivity of the atoms based on the information on the elemental composition of the molecule along with a list of atom-attributed isotropic shielding constants, obtained via quantum chemical methods based on a Hartree–Fock calculation. The utilization of shielding constants in the approach instead of NMR chemical shifts helps overcome challenges linked to the relatively limited sizes of datasets comprising reliably measured spectra. Additionally, our approach holds significant potential for scalability, as it can harness vast amounts of information on known chemical structures for the model's learning process. A comprehensive evaluation of the model trained on the QM9 and custom dataset derived from the PubChem database was conducted. The trained model was demonstrated to have the capability of accurately predicting up to 100% of the bonds for selected compounds from the QM9 dataset, achieving an impressive average accuracy rate of 37.5% for predicted bonds in the test fold. The application of the model to the tasks of NMR peak attribution, structure prediction and identification is discussed, along with prospective strategies of prediction interpretation, such as similarity searches and ranking of isomeric structures.

The extraction of compounds from natural sources is essential to organic chemistry, from identifying bioactive molecules for potential therapeutics to obtaining complex, chiral molecule building blocks. One industry that is currently leading in innovation of new botanical extraction methods and products is the cannabis industry, although it is still hampered by a lack of efficiency. Similar to chemical syntheses, anticipating the extraction conditions (flow rate, time, pressure, etc.) that will lead to the highest purity or recovery of a target molecule, like cannabinoids, is difficult. Machine learning algorithms have been demonstrated to streamline reaction optimization processes by constraining the parameter space to be physically tested to predicted regions of high performance; however, it is not altogether clear if these techniques extend to the optimization of extractions where the process conditions are even more expensive to evaluate, limiting the data available for assessment. Combining information from several sources could provide access to the requisite data necessary for implementing a data-driven approach to optimization, but little data has been made publicly available. To address this challenge and to evaluate the capabilities of machine learning for optimizing extraction processes, we built a dataset on the carbon dioxide supercritical fluid extraction (CO₂ SFE) of cannabis by harmonizing data from various companies. Using this combinatorial dataset and new techniques for maximizing the information obtained from a single large scale experiment, we built robust machine learning models to accurately predict extraction yields. The resulting machine learning models also allow for the prediction of out-of-sample biomass variations, process conditions, and scales.

The open-source Reaction Mechanism Generator (RMG) has been enhanced with new features to handle multidentate adsorbates. New reaction families have been added based upon ab initio data from 26 reactions involving C_xO_yH_z bidentate adsorbates with two heavy atoms on Pt(111). Additionally, the estimation routines for thermophysical properties were improved and extended towards bidentate species. Non-oxidative dehydrogenation of ethane over Pt(111) is used as a case study to demonstrate the effectiveness of these new features. RMG not only discovered the pathways from prior literature but also uncovered new elementary steps involving abstraction reactions. Various mono- and bimetallic catalysts for this process were screened using linear scaling relations within RMG, where a unique mechanism is generated for each catalyst. These results are consistent with prior literature trends, but they add additional insight into the rate-determining steps across the periodic table. With these additions, RMG can now explore more intricate reaction mechanisms of heterogeneously catalyzed processes for the conversion of larger molecules, which will be particularly important in fuel synthesis.

We investigate feature selection algorithms to reduce experimental time of nanoscale imaging via X-ray Absorption Fine Structure spectroscopy (nano-XANES imaging). Our approach is to decrease the required number of measurements in energy while retaining enough information to, for example, identify spatial domains and the corresponding crystallographic or chemical phase of each domain. We find sufficient accuracy in inferences when comparing predictions using the full energy point spectra to the reduced energy point subspectra recommended by feature selection. As a representative test case in the hard X-ray regime, we find that the total experimental time of nano-XANES imaging can be reduced by ∼80% for a study of Fe-bearing mineral phases. These improvements capitalize on using the most common analysis procedure – linear combination fitting onto a reference library – to train the feature selection algorithm and thus learn the optimal measurements within this analysis context. We compare various feature selection algorithms such as recursive feature elimination (RFE), random forest, and decision tree, and we find that RFE produces moderately better recommendations. We further explore practices to maintain reliable feature selection results, especially when there is large uncertainty in the system, thus requiring a more expansive reference library that results in high linear mutual dependence within the reference set. More generally, the class of spectroscopic imaging experiments that scan energy by energy (rather than collecting an entire spectrum at once) is well-addressed by feature selection, and our approach is equally applicable to the soft X-ray regime via Scanning Transmission X-ray Microscopy (STXM) experiments.

The idea of materials discovery has excited and perplexed research scientists for centuries. Several different methods have been employed to find new types of materials, ranging from the arbitrary replacement of atoms in a crystal structure to advanced machine learning methods for predicting entirely new crystal structures. In this work, we pursue three primary objectives. (I) Introduce CrysTens, a crystal encoding that can be used in a wide variety of deep learning generative models. (II) Investigate and analyze the relative performance of Generative Adversarial Networks (GANs) and Diffusion Models to find an innovative and effective way of generating theoretical crystal structures that are synthesizable and stable. (III) Show that the models that have a better “understanding” of the structure of CrysTens produce more symmetrical and realistic crystals and exhibit a better apprehension of the dataset as a whole. We accomplish these objectives using over fifty thousand Crystallographic Information Files (CIFs) from Pearson's Crystal Database.