Generative Models as an Emerging Paradigm in the Chemical Sciences

Generative discovery of safer chemical alternatives using diffusion modeling: A case study in green solvent design for cyclohexane/benzene extractive distillation

Over the past century, advancements in chemistry have significantly propelled human innovation, enhancing both industrial and consumer products. However, this rapid progression has resulted in chemical pollution increasingly surpassing planetary boundaries, as production and release rates have outpaced our monitoring capabilities. To catalyze more impactful efforts, this study transitions from traditional chemical assessment to inverse chemical design, introducing a generative graph latent diffusion model aimed at discovering safer alternatives. In a case study on the design of green solvents for cyclohexane/benzene extraction distillation, we constructed a design database encompassing functional, environmental hazards, and process constraints. Virtual screening of previous design dataset revealed distinct trade-off trends between these design requirements. Based on the screening outcomes, an unconstrained generative model was developed, which covered a broader chemical space and demonstrated superior capabilities for structural interpolation and extrapolation. To further optimize molecular generation towards desired properties, a multi-objective latent diffusion method was applied, yielding 19 candidate molecules. Of these, 7 were identified in PubChem as the most viable green solvent candidates, while the remaining 12 as potential novel candidates. Overall, this study effectively designed green solvent candidates for safer and more sustainable industrial production, setting a promising precedent for the development of environmentally friendly alternatives in other areas of chemical research.

Transforming molecular cores, substituents, and combinations into structurally diverse compounds using chemical language models

Transformer-based chemical language models (CLMs) were derived to generate structurally and topologically diverse embeddings of core structure fragments, substituents, or core/substituent combinations in chemically proper compounds, representing a design task that is difficult to address using conventional structure generation methods. To this end, CLM variants were challenged to learn different fragment-to-compound mappings in the absence of structural rules or any other fragment linking or synthetic information. The resulting alternative models were found to have high syntactic fidelity, but displayed notable differences in their ability to generate valid candidate compounds containing test fragments, with a clear preference for a model variant processing core/substituent combinations. However, the majority of valid candidate compounds generated with all models were distinct from training data and structurally novel. In addition, the CLMs exhibited high chemical diversification capacity and often generated structures with new topologies not encountered during training. Furthermore, all models produced large numbers of close structural analogues of known bioactive compounds covering a large target space, thus indicating the relevance of newly generated candidates for pharmaceutical research. As a part of our study, the new methodology and all data are made publicly available.

Asymmetric heterogeneous catalysis using crystalline porous materials

Asymmetric catalysis has emerged as a pivotal strategy in the synthesis of chiral compounds, offering significant advantages in selectivity and efficiency. In recent years, heterogeneous catalysis has become a focal point in the fields of organic synthesis and materials science due to continuous advancements in science and technology, especially the use of crystalline porous materials (CPMs) as catalysts. This review summarizes recent advances in using CPMs, such as metal-organic frameworks (MOFs), covalent organic frameworks (COFs) and zeolites, as promising supports for asymmetric catalysts. These materials provide high surface areas, tunable porosity, and the ability to host active catalytic sites, which enhance reaction rates and selectivity. In this review, we summarize the stereostructural properties of chiral CPMs to guide the future design of asymmetric heterogeneous catalysts and the study of catalytic mechanisms. Moreover, we discuss various strategies for incorporating catalytic moieties into these frameworks, including direct synthesis, post-synthesis modification and induced synthesis methods. Additionally, we highlight recent examples where CPMs have been successfully applied in asymmetric transformations, examining their mechanistic insights and the role of substrate diffusion in achieving high enantioselectivity. This review concludes with a perspective on the challenges and future directions in this rapidly evolving field, emphasizing the need for further integration of advanced artificial intelligence techniques and design principles to optimize the synthesis and catalytic performance of chiral CPMs.

Machine Learning for Toxicity Prediction Using Chemical Structures: Pillars for Success in the Real World

Machine learning (ML) is increasingly valuable for predicting molecular properties and toxicity in drug discovery. However, toxicity-related end points have always been challenging to evaluate experimentally with respect to in vivo translation due to the required resources for human and animal studies; this has impacted data availability in the field. ML can augment or even potentially replace traditional experimental processes depending on the project phase and specific goals of the prediction. For instance, models can be used to select promising compounds for on-target effects or to deselect those with undesirable characteristics (e.g., off-target or ineffective due to unfavorable pharmacokinetics). However, reliance on ML is not without risks, due to biases stemming from nonrepresentative training data, incompatible choice of algorithm to represent the underlying data, or poor model building and validation approaches. This might lead to inaccurate predictions, misinterpretation of the confidence in ML predictions, and ultimately suboptimal decision-making. Hence, understanding the predictive validity of ML models is of utmost importance to enable faster drug development timelines while improving the quality of decisions. This perspective emphasizes the need to enhance the understanding and application of machine learning models in drug discovery, focusing on well-defined data sets for toxicity prediction based on small molecule structures. We focus on five crucial pillars for success with ML-driven molecular property and toxicity prediction: (1) data set selection, (2) structural representations, (3) model algorithm, (4) model validation, and (5) translation of predictions to decision-making. Understanding these key pillars will foster collaboration and coordination between ML researchers and toxicologists, which will help to advance drug discovery and development.

Evaluating Molecular Similarity Measures: Do Similarity Measures Reflect Electronic Structure Properties?

The rapid adoption of big data, machine learning (ML), and generative artificial intelligence (AI) in chemical discovery has heightened the importance of quantifying molecular similarity. Molecular similarity, commonly assessed as the distance between molecular fingerprints, is integral to applications such as database curation, diversity analysis, and property prediction. AI tools frequently rely on these similarity measures to cluster molecules under the assumption that structurally similar molecules exhibit similar properties. However, this assumption is not universally valid, particularly for continuous properties like electronic structure properties. Despite the prevalence of fingerprint-based similarity measures, their evaluation has largely depended on biological activity data sets and qualitative metrics, limiting their relevance for nonbiological domains. To address this gap, we propose a framework to evaluate the correlation between molecular similarity measures and molecular properties. Our approach builds on the concept of neighborhood behavior and incorporates kernel density estimation (KDE) analysis to quantify how well similarity measures capture property relationships. Using a data set of over 350 million molecule pairs with electronic structure, redox, and optical properties, we systematically evaluate the correlation between several molecular fingerprint generators, distance functions, and these properties. Both the curated data set and the evaluation framework are publicly available.

Leveraging generative models with periodicity-aware, invertible and invariant representations for crystalline materials design

Designing periodicity-aware, invariant and invertible representations provides an opportunity for the inverse design of crystalline materials with desired properties by generative models. This objective requires optimizing representations and refining the architecture of generative models, yet its feasibility remains uncertain, given current progress in molecular inverse generation. In this Perspective, we highlight the progress of various methods for designing representations and generative schemes for crystalline materials, discuss the challenges in the field and propose a roadmap for future developments.

Targeted molecular generation with latent reinforcement learning

Computational methods for generating molecules with specific physiochemical properties or biological activity can greatly assist drug discovery efforts. Deep learning generative models constitute a significant step towards that direction. We introduce a novel approach that utilizes a Reinforcement Learning paradigm, called proximal policy optimization, for optimizing molecules in the latent space of a pretrained generative model. Working in the latent space of a generative model lets us bypass the need for explicitly defining chemical rules when computationally designing molecules. The generation of molecules is achieved through navigating the latent space for identifying regions that correspond to molecules with desired properties. Proximal policy optimization is a state-of-the-art policy gradient algorithm capable of operating in continuous high-dimensional spaces in a sample-efficient manner. We have paired our optimization framework with the latent spaces of two different architectures of autoencoder models showing that the method is agnostic to the underlying architecture. We present results on commonly used benchmarks for molecule optimization that demonstrate that our method has comparable or even superior performance to state-of-the-art approaches. We additionally show how our method can generate molecules that contain a pre-specified substructure while simultaneously optimizing for molecular properties, a task highly relevant to real drug discovery scenarios.

A Perspective on Foundation Models in Chemistry

Foundation models are an emerging paradigm in artificial intelligence (AI), with successful examples like ChatGPT transforming daily workflows. Generally, foundation models are large-scale, pretrained models capable of adapting to various downstream tasks by leveraging extensive data and model scaling. Their success has inspired researchers to develop foundation models for a wide range of chemical challenges, from materials discovery to understanding structure-property relationships, areas where conventional machine learning (ML) models often face limitations. In addition, foundation models hold promise for addressing persistent ML challenges in chemistry, such as data scarcity and poor generalization. In this perspective, we review recent progress in the development of foundation models in chemistry across applications of varying scope. We also discuss emerging trends and provide an outlook on promising approaches for advancing foundation models in chemistry.

Comparing Molecules Generated by MMPDB and REINVENT4 with Ideas from Drug Discovery Design Teams

This study compares molecules designed by drug discovery project teams from the Sanders Tri-Institutional Therapeutics Discovery Institute with molecules generated by two computational tools: MMPDB and REINVENT4. Seven different test cases with diverse chemotypes are studied in order to explore the potential of these computational tools in complementing human expertise in the early stages of drug discovery. By comparing the molecular structures and properties generated by MMPDB and REINVENT4 to those designed by project design teams, we aim to assess the value of such tools. The results indicate that MMPDB and REINVENT4 cover regions of chemical space larger than those covered by ideas from the drug discovery project teams. However, the chemical spaces covered by the two methods are quite different, and neither method completely covers the chemical space identified by the drug discovery project teams. Thus, the computational methods are complementary to one another and to drug discovery project team ideation. Effective application of generative molecule design tools has the potential to accelerate the identification of novel therapeutic candidates by expanding the chemical space explored during drug discovery and enabling optimal exploration.

Cross-disciplinary perspectives on the potential for artificial intelligence across chemistry

From accelerating simulations and exploring chemical space, to experimental planning and integrating automation within experimental labs, artificial intelligence (AI) is changing the landscape of chemistry. We are seeing a significant increase in the number of publications leveraging these powerful data-driven insights and models to accelerate all aspects of chemical research. For example, how we represent molecules and materials to computer algorithms for predictive and generative models, as well as the physical mechanisms by which we perform experiments in the lab for automation. Here, we present ten diverse perspectives on the impact of AI coming from those with a range of backgrounds from experimental chemistry, computational chemistry, computer science, engineering and across different areas of chemistry, including drug discovery, catalysis, chemical automation, chemical physics, materials chemistry. The ten perspectives presented here cover a range of themes, including AI for computation, facilitating discovery, supporting experiments, and enabling technologies for transformation. We highlight and discuss imminent challenges and ways in which we are redefining problems to accelerate the impact of chemical research via AI.

Solvent Screening for Separation Processes Using Machine Learning and High-Throughput Technologies

As the chemical industry shifts toward sustainable practices, there is a growing initiative to replace conventional fossil-derived solvents with environmentally friendly alternatives such as ionic liquids (ILs) and deep eutectic solvents (DESs). Artificial intelligence (AI) plays a key role in the discovery and design of novel solvents and the development of green processes. This review explores the latest advancements in AI-assisted solvent screening with a specific focus on machine learning (ML) models for physicochemical property prediction and separation process design. Additionally, this paper highlights recent progress in the development of automated high-throughput (HT) platforms for solvent screening. Finally, this paper discusses the challenges and prospects of ML-driven HT strategies for green solvent design and optimization. To this end, this review provides key insights to advance solvent screening strategies for future chemical and separation processes.

Use of AI-methods over MD simulations in the sampling of conformational ensembles in IDPs

Intrinsically Disordered Proteins (IDPs) challenge traditional structure-function paradigms by existing as dynamic ensembles rather than stable tertiary structures. Capturing these ensembles is critical to understanding their biological roles, yet Molecular Dynamics (MD) simulations, though accurate and widely used, are computationally expensive and struggle to sample rare, transient states. Artificial intelligence (AI) offers a transformative alternative, with deep learning (DL) enabling efficient and scalable conformational sampling. They leverage large-scale datasets to learn complex, non-linear, sequence-to-structure relationships, allowing for the modeling of conformational ensembles in IDPs without the constraints of traditional physics-based approaches. Such DL approaches have been shown to outperform MD in generating diverse ensembles with comparable accuracy. Most models rely primarily on simulated data for training and experimental data serves a critical role in validation, aligning the generated conformational ensembles with observable physical and biochemical properties. However, challenges remain, including dependence on data quality, limited interpretability, and scalability for larger proteins. Hybrid approaches combining AI and MD can bridge the gaps by integrating statistical learning with thermodynamic feasibility. Future directions include incorporating physics-based constraints and learning experimental observables into DL frameworks to refine predictions and enhance applicability. AI-driven methods hold significant promise in IDP research, offering novel insights into protein dynamics and therapeutic targeting while overcoming the limitations of traditional MD simulations.

Uncertainty quantification with graph neural networks for efficient molecular design

Optimizing molecular design across expansive chemical spaces presents unique challenges, especially in maintaining predictive accuracy under domain shifts. This study integrates uncertainty quantification (UQ), directed message passing neural networks (D-MPNNs), and genetic algorithms (GAs) to address these challenges. We systematically evaluate whether UQ-enhanced D-MPNNs can effectively optimize broad, open-ended chemical spaces and identify the most effective implementation strategies. Using benchmarks from the Tartarus and GuacaMol platforms, our results show that UQ integration via probabilistic improvement optimization (PIO) enhances optimization success in most cases, supporting more reliable exploration of chemically diverse regions. In multi-objective tasks, PIO proves especially advantageous, balancing competing objectives and outperforming uncertainty-agnostic approaches. This work provides practical guidelines for integrating UQ in computational-aided molecular design (CAMD).

Generate what you can make: achieving in-house synthesizability with readily available resources in de novo drug design

Computer-Aided Synthesis Planning (CASP) and CASP-based approximated synthesizability scores have rarely been used as generation objectives in Computer-Aided Drug Design despite facilitating the in-silico generation of synthesizable molecules. However, these synthesizability approaches are disconnected from the reality of small laboratory drug design, where building block resources are limited, thus making the notion of in-house synthesizability with already available resources highly desirable. In this work, we show a successful in-house de novo drug design workflow generating active and in-house synthesizable ligands of monoglyceride lipase (MGLL). First, we demonstrate the successful transfer of CASP from 17.4 million commercial building blocks to a small laboratory setting of roughly 6000 building blocks with only a decrease of -12% in CASP success when accepting two reaction-steps longer synthesis routes on average. Next, we present a rapidly retrainable in-house synthesizability score, successfully capturing our in-house synthesizability without relying on external building block resources. We show that including our in-house synthesizability score in a multi-objective de novo drug design workflow, alongside a simple QSAR model, provides thousands of potentially active and easily in-house synthesizable molecules. Finally, we experimentally evaluate the synthesis and biochemical activity of three de novo candidates using their CASP-suggested synthesis routes employing only in-house building blocks. We find one candidate with evident activity, suggesting potential new ligand ideas for MGLL inhibitors while showcasing the usefulness of our in-house synthesizability score for de novo drug design.Scientific contribution Our core scientific contribution is the introduction of in-house de novo drug design, which enables the practical application of generative methods in small laboratories by utilizing a limited stock of available building blocks. Our fast-to-adapt workflow for in-house synthesizability scoring requires minimal computational retraining costs while supporting a high diversity of generated structures. We highlight the practicality of our approach through a comprehensive in-vitro case study that relies entirely on in-house resources, including in-silico generation, synthesis planning, and activity evaluation.

Molecular Denoising Using Diffusion Models with Physics-Informed Priors

Denoising Diffusion Probabilistic Models (DDPMs) are powerful generative models that have demonstrated superior performance in a variety of tasks and applications in material science and molecular graph modeling. Inspired by nonequilibrium statistical mechanics, these models iteratively degrade data through a forward diffusion process and then restore it by learning the time-reversal of the forward process. Despite their success, a significant drawback of DDPMs is their reliance on numerous iterations to generate high-quality samples, resulting in slow sampling. In this Letter, we introduce a strategy to improve DDPMs for atomistic systems by leveraging the thermodynamics of the data by deriving physics-informed priors. Drawing on principles from statistical mechanics, we derive physics-informed parameters for the prior distribution to initialize the Markov chain closer to the true data distribution. This strategy shortens the Markov chain, thereby improving the model's training efficiency and accelerating the sampling process. We demonstrate the effectiveness of our method in denoising noisy radial distribution functions obtained from a single atomic configuration of diverse Lennard-Jones and multiatomic liquids.

Increasing the Accuracy and Robustness of the CHARMM General Force Field with an Expanded Training Set

Small molecule empirical force fields (FFs), including the CHARMM General Force Field (CGenFF), are designed to have wide coverage of organic molecules and to rapidly assign parameters to molecules not explicitly included in the FF. Assignment of parameters to new molecules in CGenFF is based on a trained bond-angle-dihedral charge increment linear interpolation scheme for the partial atomic charges along with bonded parameters assigned based on analogy using a rules-based penalty score scheme associated with atom types and chemical connectivity. Accordingly, the accuracy of CGenFF is related to the extent of the training set of available parameters. In the present study that training set is extended by 1390 molecules selected to represent connectivities new to CGenFF training compounds. Quantum mechanical (QM) data for optimized geometries, bond, valence angle, and dihedral angle potential energy scans, interactions with water, molecular dipole moments, and electrostatic potentials were used as target data. The resultant bonded parameters and partial atomic charges were used to train a new version of the CGenFF program, v5.0, which was used to generate parameters for a validation set of molecules, including drug-like molecules approved by the FDA, which were then benchmarked against both experimental and QM data. CGenFF v5.0 shows overall improvements with respect to QM intramolecular geometries, vibrations, dihedral potential energy scans, dipole moments and interactions with water. Tests of pure solvent properties of 216 molecules show small improvements versus the previous release of CGenFF v2.5.1 reflecting the high quality of the Lennard-Jones parameters that were explicitly optimized during the initial optimization of both the CGenFF and the CHARMM36 force field. CGenFF v5.0 represents an improvement that is anticipated to more accurately model intramolecular geometries and strain energies as well as noncovalent interactions of drug-like and other organic molecules.

Large language models open new way of AI-assisted molecule design for chemists

Recent advancements in artificial intelligence (AI)-based molecular design methodologies have offered synthetic chemists new ways to design functional molecules with their desired properties. While various AI-based molecule generators have significantly advanced toward practical applications, their effective use still requires specialized knowledge and skills concerning AI techniques. Here, we develop a large language model (LLM)-powered chatbot, ChatChemTS, that assists users in designing new molecules using an AI-based molecule generator through only chat interactions, including automated construction of reward functions for the specified properties. Our study showcases the utility of ChatChemTS through de novo design cases involving chromophores and anticancer drugs (epidermal growth factor receptor inhibitors), exemplifying single- and multiobjective molecule optimization scenarios, respectively. ChatChemTS is provided as an open-source package on GitHub at https://github.com/molecule-generator-collection/ChatChemTS . Scientific contribution ChatChemTS is an open-source application that assists users in utilizing an AI-based molecule generator, ChemTSv2, solely through chat interactions. This study demonstrates that LLMs possess the potential to utilize advanced software, such as AI-based molecular generators, which require specialized knowledge and technical skills.

Going digital to boost safe and sustainable materials innovation markets. The digital safe-and-sustainability-by-design innovation approach of the PINK project

In this innovation report, we present the vision of the PINK project to foster Safe-and-Sustainable-by-Design (SSbD) advanced materials and chemicals (AdMas&Chems) development by integrating state-of-the-art computational modelling, simulation tools and data resources. PINK proposes a novel approach for the use of the SSbD Framework, whose innovative approach is based on the application of a multi-objective optimisation procedure for the criteria of functionality, safety, sustainability and cost efficiency. At the core is the PINK open innovation platform, a distributed system that integrates all relevant modelling resources enriched with advanced data visualisation and an AI-driven decision support system. Data and modelling tools from the, in large parts, independently developed areas of functional design, safety assessment, life cycle assessment & costing are brought together based on a newly created Interoperability Framework. The PINK In Silico Hub, as the user Interface to the platform, finally guides the user through the complete AdMas&Chems development process from idea creation to market introduction. Guided by two Developmental Case Studies, the process of building of the PINK Platform is iterative, ensuring industry readiness to implement and apply it. Additionally, the Industrial Demonstrator programme will be introduced as part of the final project phase, which allows industry partners and especially small and medium enterprises (SMEs) to become part of the PINK consortium. Feedback from the Demonstrators as well as other stakeholder-engagement activities and collaborations will shape the platform's final look and feel and, even more important, activities to assure long-term technical sustainability.

Crystal Structure Prediction Meets Artificial Intelligence

Crystal structure prediction (CSP) represents a fundamental research frontier in computational materials science and chemistry, aiming to predict thermodynamically stable periodic structures from given chemical compositions. Traditional methods often face challenges such as high computational costs and local minima trapping. Recently, artificial intelligence methods, represented by generative adversarial networks (GANs), variational autoencoders (VAEs), diffusion models, and large language models (LLMs), have revolutionized the traditional prediction paradigm. These computational frameworks efficiently extract chemical rules and structural features from crystal databases, significantly reducing computational costs while maintaining prediction accuracy. This Perspective systematically evaluates the advantages and limitations of various generative models, explores their synergies with conventional approaches, and discusses their future prospects in accelerating materials discovery and development, providing new insights for future research directions.

Mol-AIR: Molecular Reinforcement Learning with Adaptive Intrinsic Rewards for Goal-Directed Molecular Generation

Optimizing techniques for discovering molecular structures with desired properties is crucial in artificial intelligence (AI)-based drug discovery. Combining deep generative models with reinforcement learning has emerged as an effective strategy for generating molecules with specific properties. Despite its potential, this approach is ineffective in exploring the vast chemical space and optimizing particular chemical properties. To overcome these limitations, we present Mol-AIR, a reinforcement learning-based framework using adaptive intrinsic rewards for effective goal-directed molecular generation. Mol-AIR leverages the strengths of both history-based and learning-based intrinsic rewards by exploiting random distillation network and counting-based strategies. In benchmark tests, Mol-AIR demonstrates improved performance over existing approaches in generating molecules having the desired properties, including penalized LogP, QED, and celecoxib similarity, without any prior knowledge. We believe that Mol-AIR represents a significant advancement in drug discovery, offering a more efficient path to discovering novel therapeutics.

High performance-oriented computer aided drug design approaches in the exascale era

INTRODUCTION

In 2023, the first exascale supercomputer was opened to the public in the US. With a demonstrated 1.1 exaflops of performance, Frontier represents an unprecedented breakthrough in high-performance computing (HPC). Currently, more (and more powerful) machines are being installed worldwide. Computer-aided drug design (CADD) is one of the fields of computational science that can greatly benefit from exascale computing for the benefit of the whole society. However, scaling CADD approaches to exploit exascale machines require new algorithmic and software solutions.

AREAS COVERED

Here, the authors consider physics-based and machine learning (ML)-aided techniques for the design of small molecule binders capable of leveraging modern parallel computer architectures. Specifically, the authors focus on HPC-oriented large-scale applications from the past 3 years that were enabled by (pre)exascale supercomputers by running on up tothousands of accelerated nodes.

EXPERT OPINION

In the area of ML, exascale computers can enable the training of generative models with unprecedented predictive power to design novel ligands, provided large amounts of high-quality data are available. Exascale computers could also unlock the potential of accurate ML-aided physics-based methods to boost the success rate of structure-based drug design campaigns. Currently, however, methodological developments are still required to allow routine large-scale applications of such rigorous approaches.

Chemical space as a unifying theme for chemistry

Chemistry has diversified from a basic understanding of the elements to studying millions of highly diverse molecules and materials, which together are conceptualized as the chemical space. A map of this chemical space where distances represent similarities between compounds can represent the mutual relationships between different subfields of chemistry and help the discipline to be viewed and understood globally.

Large property models: a new generative machine-learning formulation for molecules

Generative models for the inverse design of molecules with particular properties have been heavily hyped, but have yet to demonstrate significant gains over machine-learning-augmented expert intuition. A major challenge of such models is their limited accuracy in predicting molecules with targeted properties in the data-scarce regime, which is the regime typical of the prized outliers that it is hoped inverse models will discover. For example, activity data for a drug target or stability data for a material may only number in the tens to hundreds of samples, which is insufficient to learn an accurate and reasonably general property-to-structure inverse mapping from scratch. We've hypothesized that the property-to-structure mapping becomes unique when a sufficient number of properties are supplied to the models during training. This hypothesis has several important corollaries if true. It would imply that data-scarce properties can be completely determined using a set of more accessible molecular properties. It would also imply that a generative model trained on multiple properties would exhibit an accuracy phase transition after achieving a sufficient size-a process analogous to what has been observed in the context of large language models. To interrogate these behaviors, we have built the first transformers trained on the property-to-molecular-graph task, which we dub "large property models" (LPMs). A key ingredient is supplementing these models during training with relatively basic but abundant chemical property data. The motivation for the large-property-model paradigm, the model architectures, and case studies are presented here.

Spiers Memorial Lecture: How to do impactful research in artificial intelligence for chemistry and materials science

Machine learning has been pervasively touching many fields of science. Chemistry and materials science are no exception. While machine learning has been making a great impact, it is still not reaching its full potential or maturity. In this perspective, we first outline current applications across a diversity of problems in chemistry. Then, we discuss how machine learning researchers view and approach problems in the field. Finally, we provide our considerations for maximizing impact when researching machine learning for chemistry.

CoarsenConf: Equivariant Coarsening with Aggregated Attention for Molecular Conformer Generation

Molecular conformer generation (MCG) is an important task in cheminformatics and drug discovery. The ability to efficiently generate low-energy 3D structures can avoid expensive quantum mechanical simulations, leading to accelerated virtual screenings and enhanced structural exploration. Several generative models have been developed for MCG, but many struggle to consistently produce high-quality conformers for meaningful downstream applications. To address these issues, we introduce CoarsenConf, which coarse-grains molecular graphs based on torsional angles and integrates them into an SE(3)-equivariant hierarchical variational autoencoder. Through equivariant coarse-graining, we aggregate the fine-grained atomic coordinates of subgraphs connected via rotatable bonds, creating a variable-length coarse-grained latent representation. Our model uses a novel aggregated attention mechanism to restore fine-grained coordinates from the coarse-grained latent representation, enabling efficient generation of accurate conformers. Furthermore, we evaluate the chemical and biochemical quality of our generated conformers on multiple downstream applications, including property prediction and large-scale oracle-based protein docking. Overall, CoarsenConf generates more accurate conformer ensembles compared to prior generative models.

BindingDB in 2024: a FAIR knowledgebase of protein-small molecule binding data

BindingDB (bindingdb.org) is a public, web-accessible database of experimentally measured binding affinities between small molecules and proteins, which supports diverse applications including medicinal chemistry, biochemical pathway annotation, training of artificial intelligence models and computational chemistry methods development. This update reports significant growth and enhancements since our last review in 2016. Of note, the database now contains 2.9 million binding measurements spanning 1.3 million compounds and thousands of protein targets. This growth is largely attributable to our unique focus on curating data from US patents, which has yielded a substantial influx of novel binding data. Recent improvements include a remake of the website following responsive web design principles, enhanced search and filtering capabilities, new data download options and webservices and establishment of a long-term data archive replicated across dispersed sites. We also discuss BindingDB's positioning relative to related resources, its open data sharing policies, insights gleaned from the dataset and plans for future growth and development.

Critical review on in silico methods for structural annotation of chemicals detected with LC/HRMS non-targeted screening

Non-targeted screening with liquid chromatography coupled to high-resolution mass spectrometry (LC/HRMS) is increasingly leveraging in silico methods, including machine learning, to obtain candidate structures for structural annotation of LC/HRMS features and their further prioritization. Candidate structures are commonly retrieved based on the tandem mass spectral information either from spectral or structural databases; however, the vast majority of the detected LC/HRMS features remain unannotated, constituting what we refer to as a part of the unknown chemical space. Recently, the exploration of this chemical space has become accessible through generative models. Furthermore, the evaluation of the candidate structures benefits from the complementary empirical analytical information such as retention time, collision cross section values, and ionization type. In this critical review, we provide an overview of the current approaches for retrieving and prioritizing candidate structures. These approaches come with their own set of advantages and limitations, as we showcase in the example of structural annotation of ten known and ten unknown LC/HRMS features. We emphasize that these limitations stem from both experimental and computational considerations. Finally, we highlight three key considerations for the future development of in silico methods.

Artificial intelligence and food flavor: How AI models are shaping the future and revolutionary technologies for flavor food development

The food flavor science, traditionally reliant on experimental methods, is now entering a promising era with the help of artificial intelligence (AI). By integrating existing technologies with AI, researchers can explore and develop new flavor substances in a digital environment, saving time and resources. More and more research will use AI and big data to enhance product flavor, improve product quality, meet consumer needs, and drive the industry toward a smarter and more sustainable future. In this review, we elaborate on the mechanisms of flavor recognition and their potential impact on nutritional regulation. With the increase of data accumulation and the development of internet information technology, food flavor databases and food ingredient databases have made great progress. These databases provide detailed information on the nutritional content, flavor molecules, and chemical properties of various food compounds, providing valuable data support for the rapid evaluation of flavor components and the construction of screening technology. With the popularization of AI in various fields, the field of food flavor has also ushered in new development opportunities. This review explores the mechanisms of flavor recognition and the role of AI in enhancing food flavor analysis through high-throughput omics data and screening technologies. AI algorithms offer a pathway to scientifically improve product formulations, thereby enhancing flavor and customized meals. Furthermore, it discusses the safety challenges of integrating AI into the food flavor industry.

Data Generation for Machine Learning Interatomic Potentials and Beyond

The field of data-driven chemistry is undergoing an evolution, driven by innovations in machine learning models for predicting molecular properties and behavior. Recent strides in ML-based interatomic potentials have paved the way for accurate modeling of diverse chemical and structural properties at the atomic level. The key determinant defining MLIP reliability remains the quality of the training data. A paramount challenge lies in constructing training sets that capture specific domains in the vast chemical and structural space. This Review navigates the intricate landscape of essential components and integrity of training data that ensure the extensibility and transferability of the resulting models. We delve into the details of active learning, discussing its various facets and implementations. We outline different types of uncertainty quantification applied to atomistic data acquisition and the correlations between estimated uncertainty and true error. The role of atomistic data samplers in generating diverse and informative structures is highlighted. Furthermore, we discuss data acquisition via modified and surrogate potential energy surfaces as an innovative approach to diversify training data. The Review also provides a list of publicly available data sets that cover essential domains of chemical space.

Kolmogorov-Arnold Network Made Learning Physics Laws Simple

In recent years, contrastive learning has gained widespread adoption in machine learning applications to physical systems primarily due to its distinctive cross-modal capabilities and scalability. Building on the foundation of Kolmogorov-Arnold Networks (KANs) [Liu, Z. et al. Kan: Kolmogorov-arnold networks. arXiv 2024, 2404.19756], we introduce a novel contrastive learning framework, Kolmogorov-Arnold Contrastive Crystal Property Pretraining (KCCP), which integrates the principles of CLIP and KAN to establish robust correlations between crystal structures and their physical properties. During the training process, we conducted a comparative analysis between Multilayer Perceptron (MLP) and KAN, revealing that KAN significantly outperforms MLP in both accuracy and convergence speed for this task. By extending the capabilities of contrastive learning to the realm of physical systems, KCCP offers a promising approach for constructing cross-data structural and cross-modal physical models, representing an area of considerable potential.

Unleashing the future: The revolutionary role of machine learning and artificial intelligence in drug discovery

Drug discovery is a complex and multifaceted process aimed at identifying new therapeutic compounds with the potential to treat various diseases. Traditional methods of drug discovery are often time-consuming, expensive, and characterized by low success rates. Because of this, there is an urgent need to improve the drug development process using new technologies. The integration of the current state-of-art of artificial intelligence (AI) and machine learning (ML) approaches with conventional methods will enhance the efficiency and effectiveness of pharmaceutical research. This review highlights the transformative impact of AI and ML in drug discovery, discussing current applications, challenges, and future directions in harnessing these technologies to accelerate the development of innovative therapeutics. We have discussed the latest developments in AI and ML technologies to streamline several stages of drug discovery, from target identification and validation to lead optimization and preclinical studies.

Molecular tweaking by generative cheminformatics and ligand-protein structures for rational drug discovery

The purpose of this review is two-fold: (1) to summarize artificial intelligence and machine learning approaches and document the role of ligand-protein structures in directing drug discovery; (2) to present examples of drugs from the recent literature (past decade) of case studies where such strategies have been applied to accelerate the discovery pipeline. Compared to 50 years ago when drug discovery was largely a synthetic chemist driven research exercise, today a holistic approach needs to be adopted with seamless integration between synthetic and medicinal chemistry, supramolecular complexes, computations, artificial intelligence, machine learning, structural biology, chemical biology, diffraction analytical tools, drugs databases, and pharmacology. The urgency for an integrated and collaborative platform to accelerate drug discovery in an academic setting is emphasized.

Molecular similarity: Theory, applications, and perspectives

Molecular similarity pervades much of our understanding and rationalization of chemistry. This has become particularly evident in the current data-intensive era of chemical research, with similarity measures serving as the backbone of many Machine Learning (ML) supervised and unsupervised procedures. Here, we present a discussion on the role of molecular similarity in drug design, chemical space exploration, chemical "art" generation, molecular representations, and many more. We also discuss more recent topics in molecular similarity, like the ability to efficiently compare large molecular libraries.

De novo Design of Biocompatible Nanomaterials Using Quasi-SMILES and Recurrent Neural Networks

Screening nanomaterials (NMs) with desired properties from the extensive chemical space presents significant challenges. The potential toxicity of NMs further limits their applications in biological systems. Traditional methods struggle with these complexities, but generative models offer a possible solution to producing new molecules without prior knowledge. However, converting complex 3D nanostructures into computer-readable formats remains a critical prerequisite. To overcome these challenges, we proposed an innovative deep-learning framework for the de novo design of biocompatible NMs. This framework comprises two predictive models and a generative model, utilizing a Quasi-SMILES representation to encode three-dimensional structural information on NMs. Our generative model successfully created 289 new NMs not previously seen in the training set. The predictive models identified a particularly promising NM characterized by high cellular uptake and low toxicity. This NM was successfully synthesized, and its predicted properties were experimentally validated. Our approach advances the application of artificial intelligence in NM design and provides a practical solution for balancing functionality and toxicity in NMs.

NGT: Generative AI with Synthesizability Guarantees Discovers MC2R Inhibitors from a Tera-Scale Virtual Screen

Commercially available, synthesis-on-demand virtual libraries contain upward of trillions of readily synthesizable compounds for drug discovery campaigns. These libraries are a critical resource for rapid cycles of in silico discovery, property optimization and in vitro validation. However, as these libraries continue to grow exponentially in size, traditional search strategies encounter significant limitations. Here we present NeuralGenThesis (NGT), an efficient reinforcement learning approach to generate compounds from ultralarge libraries that satisfy user-specified constraints. Our method first trains a generative model over a virtual library and subsequently trains a normalizing flow to learn a distribution over latent space that decodes constraint-satisfying compounds. NGT allows multiple constraints simultaneously without dictating how molecular properties are calculated. Using NGT, we generated potent and selective inhibitors for the melanocortin-2 receptor (MC2R) from a three trillion compound library. NGT offers a powerful and scalable solution for navigating ultralarge virtual libraries, accelerating drug discovery efforts.

AlphaFold Meets De Novo Drug Design: Leveraging Structural Protein Information in Multitarget Molecular Generative Models

Recent advancements in deep learning and generative models have significantly expanded the applications of virtual screening for drug-like compounds. Here, we introduce a multitarget transformer model, PCMol, that leverages the latent protein embeddings derived from AlphaFold2 as a means of conditioning a de novo generative model on different targets. Incorporating rich protein representations allows the model to capture their structural relationships, enabling the chemical space interpolation of active compounds and target-side generalization to new proteins based on embedding similarities. In this work, we benchmark against other existing target-conditioned transformer models to illustrate the validity of using AlphaFold protein representations over raw amino acid sequences. We show that low-dimensional projections of these protein embeddings cluster appropriately based on target families and that model performance declines when these representations are intentionally corrupted. We also show that the PCMol model generates diverse, potentially active molecules for a wide array of proteins, including those with sparse ligand bioactivity data. The generated compounds display higher similarity known active ligands of held-out targets and have comparable molecular docking scores while maintaining novelty. Additionally, we demonstrate the important role of data augmentation in bolstering the performance of generative models in low-data regimes. Software package and AlphaFold protein embeddings are freely available at https://github.com/CDDLeiden/PCMol.

Response Matching for Generating Materials and Molecules

Diffusion models have recently emerged as powerful tools for the generation of new molecular and material structures. The key insight is that the noise in these models is related to the response of the atoms to displacement, and the denoising step is thus analogous to the geometry relaxation of atomistic systems starting from a random structure. Building on this, we present a generative method called Response Matching (RM), which leverages the fact that each stable material or molecule exists at the minimum of its potential energy surface. Any perturbation induces a response in energy and stress, driving the structure back to equilibrium. Matching this response is closely related to score matching in diffusion models. Another important aspect of state-of-the-art diffusion models is the incorporation of physical symmetries such as translation, rotation, and periodicity. RM employs a machine learning interatomic potential and random structure search as the denoising model, inherently respecting these symmetries and exploiting the locality of atomic interactions. RM handles both molecules and bulk materials under the same framework. Its efficiency and generalization are demonstrated on three systems: a small organic molecular data set, stable crystals from the Materials Project, and one-shot learning on a single diamond configuration.

An approach for full space inverse materials design by combining universal machine learning potential, universal property model, and optimization algorithm

We present a full space inverse materials design (FSIMD) approach that fully automates the materials design for target physical properties without the need to provide the atomic composition, chemical stoichiometry, and crystal structure in advance. Here, we used density functional theory reference data to train a universal machine learning potential (UPot) and transfer learning to train a universal bulk modulus model (UBmod). Both UPot and UBmod were able to cover materials systems composed of any element among 42 elements. Interfaced with optimization algorithm and enhanced sampling, the FSIMD approach is applied to find the materials with the largest cohesive energy and the largest bulk modulus, respectively. NaCl-type ZrC was found to be the material with the largest cohesive energy. For bulk modulus, diamond was identified to have the largest value. The FSIMD approach is also applied to design materials with other multi-objective properties with accuracy limited principally by the amount, reliability, and diversity of the training data. The FSIMD approach provides a new way for inverse materials design with other functional properties for practical applications.

Recent advances from computer-aided drug design to artificial intelligence drug design

Computer-aided drug design (CADD), a cornerstone of modern drug discovery, can predict how a molecular structure relates to its activity and interacts with its target using structure-based and ligand-based methods. Fueled by ever-increasing data availability and continuous model optimization, artificial intelligence drug design (AIDD), as an enhanced iteration of CADD, has thrived in the past decade. AIDD demonstrates unprecedented opportunities in protein folding, property prediction, and molecular generation. It can also facilitate target identification, high-throughput screening (HTS), and synthetic route prediction. With AIDD involved, the process of drug discovery is greatly accelerated. Notably, AIDD offers the potential to explore uncharted territories of chemical space beyond current knowledge. In this perspective, we began by briefly outlining the main workflows and components of CADD. Then through showcasing exemplary cases driven by AIDD in recent years, we describe the evolving role of artificial intelligence (AI) in drug discovery from three distinct stages, that is, chemical library screening, linker generation, and de novo molecular generation. In this process, we attempted to draw comparisons between the features of CADD and AIDD.

Discovering Photoswitchable Molecules for Drug Delivery with Large Language Models and Chemist Instruction Training

Background: As large language models continue to expand in size and diversity, their substantial potential and the relevance of their applications are increasingly being acknowledged. The rapid advancement of these models also holds profound implications for the long-term design of stimulus-responsive materials used in drug delivery. Methods: The large model used Hugging Face's Transformers package with BigBird, Gemma, and GPT NeoX architectures. Pre-training used the PubChem dataset, and fine-tuning used QM7b. Chemist instruction training was based on Direct Preference Optimization. Drug Likeness, Synthetic Accessibility, and PageRank Scores were used to filter molecules. All computational chemistry simulations were performed using ORCA and Time-Dependent Density-Functional Theory. Results: To optimize large models for extensive dataset processing and comprehensive learning akin to a chemist's intuition, the integration of deeper chemical insights is imperative. Our study initially compared the performance of BigBird, Gemma, GPT NeoX, and others, specifically focusing on the design of photoresponsive drug delivery molecules. We gathered excitation energy data through computational chemistry tools and further investigated light-driven isomerization reactions as a critical mechanism in drug delivery. Additionally, we explored the effectiveness of incorporating human feedback into reinforcement learning to imbue large models with chemical intuition, enhancing their understanding of relationships involving -N=N- groups in the photoisomerization transitions of photoresponsive molecules. Conclusions: We implemented an efficient design process based on structural knowledge and data, driven by large language model technology, to obtain a candidate dataset of specific photoswitchable molecules. However, the lack of specialized domain datasets remains a challenge for maximizing model performance.

Augmenting genetic algorithms with machine learning for inverse molecular design

Evolutionary and machine learning methods have been successfully applied to the generation of molecules and materials exhibiting desired properties. The combination of these two paradigms in inverse design tasks can yield powerful methods that explore massive chemical spaces more efficiently, improving the quality of the generated compounds. However, such synergistic approaches are still an incipient area of research and appear underexplored in the literature. This perspective covers different ways of incorporating machine learning approaches into evolutionary learning frameworks, with the overall goal of increasing the optimization efficiency of genetic algorithms. In particular, machine learning surrogate models for faster fitness function evaluation, discriminator models to control population diversity on-the-fly, machine learning based crossover operations, and evolution in latent space are discussed. The further potential of these synergistic approaches in generative tasks is also assessed, outlining promising directions for future developments.

Catalysing (organo-)catalysis: Trends in the application of machine learning to enantioselective organocatalysis

Organocatalysis has established itself as a third pillar of homogeneous catalysis, besides transition metal catalysis and biocatalysis, as its use for enantioselective reactions has gathered significant interest over the last decades. Concurrent to this development, machine learning (ML) has been increasingly applied in the chemical domain to efficiently uncover hidden patterns in data and accelerate scientific discovery. While the uptake of ML in organocatalysis has been comparably slow, the last two decades have showed an increased interest from the community. This review gives an overview of the work in the field of ML in organocatalysis. The review starts by giving a short primer on ML for experimental chemists, before discussing its application for predicting the selectivity of organocatalytic transformations. Subsequently, we review ML employed for privileged catalysts, before focusing on its application for catalyst and reaction design. Concluding, we give our view on current challenges and future directions for this field, drawing inspiration from the application of ML to other scientific domains.

Leveraging experimental and computational tools for advancing carbon capture adsorbents research

CO2 emissions have been steadily increasing and have been a major contributor for climate change compelling nations to take decisive action fast. The average global temperature could reach 1.5 °C by 2035 which could cause a significant impact on the environment, if the emissions are left unchecked. Several strategies have been explored of which carbon capture is considered the most suitable for faster deployment. Among different carbon capture solutions, adsorption is considered both practical and sustainable for scale-up. But the development of adsorbents that can exhibit satisfactory performance is typically done through the experimental approach. This hit and trial method is costly and time consuming and often success is not guaranteed. Machine learning (ML) and other computational tools offer an alternate to this approach and is accessible to everyone. Often, the research towards materials focuses on maximizing its performance under simulated conditions. The aim of this study is to present a holistic view on progress in material research for carbon capture and the various tools available in this regard. Thus, in this review, we first present a context on the workflow for carbon capture material development before providing various machine learning and computational tools available to support researchers at each stage of the process. The most popular application of ML models is for predicting material performance and recommends that ML approaches can be utilized wherever possible so that experimentations can be focused on the later stages of the research and development.

Transformers for Molecular Property Prediction: Lessons Learned from the Past Five Years

Molecular Property Prediction (MPP) is vital for drug discovery, crop protection, and environmental science. Over the last decades, diverse computational techniques have been developed, from using simple physical and chemical properties and molecular fingerprints in statistical models and classical machine learning to advanced deep learning approaches. In this review, we aim to distill insights from current research on employing transformer models for MPP. We analyze the currently available models and explore key questions that arise when training and fine-tuning a transformer model for MPP. These questions encompass the choice and scale of the pretraining data, optimal architecture selections, and promising pretraining objectives. Our analysis highlights areas not yet covered in current research, inviting further exploration to enhance the field's understanding. Additionally, we address the challenges in comparing different models, emphasizing the need for standardized data splitting and robust statistical analysis.

Artificial intelligence-aiding lab-on-a-chip workforce designed oral [3.1.0] bi and [4.2.0] tricyclic catalytic interceptors inhibiting multiple SARS-CoV-2 protomers assisted by double-shell deep learning

While each massive pandemic has claimed the lives of millions of vulnerable populations over the centuries, one limitation exists: that the Edisonian approach (human-directed with trial errors) relies on repurposing pharmaceuticals, designing drugs, and herbal remedies with the violation of Lipinski's rule of five druglikeness. It may lead to adverse health effects with long-term health multimorbidity. Nevertheless, declining birth rates and aging populations will likely cause a shift in society due to a shortage of a scientific workforce to defend against the next pandemic incursion. The challenge of combating the ongoing post-COVID-19 pandemic has been exacerbated by the lack of gold standard drugs to deactivate multiple SARS-CoV-2 protein targets. Meanwhile, there are three FDA-approved antivirals, Remdesivir, Molnupiravir, and Paxlovid, with moderate clinical efficacy and drug resistance. There is a pressing need for additional antivirals and prepared omics technology to combat the current and future devastating coronavirus pandemics. While there is a limitation of existing contemporary inhibitors to deactivate viral RNA replication with minimal rotational bonds, one strategy is to create Lipinski inhibitors with less than 10 rotational bonds and precise halogen bond placement to destabilize multiple viral protomers. This work describes the efforts to design gold-standard oral inhibitors of bi- and tri-cyclic catalytic interceptors with electrophilic heads using double-shell deep learning. Here, KS1 with and KS2 compounds designed by lab-on-a-chip technology attain 5-fold novel filtered-Lipinski, GHOSE, VEBER, EGAN, and MUEGGE druglikeness. The graph neural network (GNN) relies on module-initiation, expansion, relabeling atom index, and termination (METORITE) iterations, while the deep neural network (DNN) engages pinning, extraction, convolution, pooling, and flattening (PROOF) operations. The cyclic compound's specific halogen atom location enhances the nitrile catalytic head, which deactivates several viral protein targets. Initiating this lab-on-a-chip that is not susceptible to the aging process for creating clinical compounds can leverage a new path to many valuable drugs with speedy oral drug discovery, especially to defend the loss of vulnerable population and prevent multimorbidity that is susceptible to hidden viral persistence in the continuing aging times.

Identification of potential PIM-2 inhibitors via ligand-based generative models, molecular docking and molecular dynamics simulations

Proviral Integrations of Moloney-2 (PIM-2) kinase is a promising target for various cancers and other diseases, and its inhibitors hold potential for treating related diseases. However, there is currently no clinically available PIM-2 inhibitor. In this study, we constructed a generative model for de novo PIM-2 inhibitor design based on artificial intelligence, performed molecular docking and molecular dynamics (MD) simulations to develop an efficient PIM-2 inhibitor generative model and discover potential PIM-2 inhibitors. First, we designed a generative model based on a Bi-directional Long Short-Term Memory (BiLSTM) framework combined with a transfer learning strategy and generated a new PIM-2 small molecule library using existing active drug databases. The generated compound library was then virtually screened by molecular docking and scaffold similarity comparison, identifying 10 initial hit compounds with better performance. Next, using the inhibitor in the crystal structure as a positive control, we performed two rounds of MD simulations, with lengths of 100 ns and 500 ns, respectively, to study the dynamic stability of the protein-ligand systems of the 10 compounds with PIM-2. Analyzed the interactions with key hinge region residues, binding free energies, and changes in the ATP pocket size. The generative model demonstrates good molecular generation capability and can generate efficient novel molecules with similar physicochemical properties as active PIM-2 drugs. Among the 10 initially selected hit compounds, 5 compounds C3 (- 29.69 kcal/mol), C4 (- 33.31 kcal/mol), C5 (- 28.59 kcal/mol), C8 (- 34.68 kcal/mol), and C9 (- 25.88 kcal/mol) have higher binding energies with PIM-2 than the positive drug 3YR (- 26.18 kcal/mol). The MD simulation results are consistent with the docking analysis, these compounds have lower and more stable RMSD values for the complex systems with the reported positive drug 3YR and PIM-2 complex system. They can form long-term stable interactions with active site and the hinge region of PIM-2, which suggests these compounds are likely to have potent inhibitory effects on PIM-2. This study provides an efficient generative model for PIM-2 inhibitor research and discovers 5 potential novel PIM-2 inhibitors.

Deductive machine learning models for product identification

Deductive solution strategies are required in prediction scenarios that are under determined, when contradictory information is available, or more generally wherever one-to-many non-functional mappings occur. In contrast, most contemporary machine learning (ML) in the chemical sciences is inductive learning from example, with a fixed set of features. Chemical workflows are replete with situations requiring deduction, including many aspects of lab automation and spectral interpretation. Here, a general strategy is described for designing and training machine learning models capable of deduction that consists of combining individual inductive models into a larger deductive network. The training and testing of these models is demonstrated on the task of deducing reaction products from a mixture of spectral sources. The resulting models can distinguish between intended and unintended reaction outcomes and identify starting material based on a mixture of spectral sources. The models also perform well on tasks that they were not directly trained on, like performing structural inference using real rather than simulated spectral inputs, predicting minor products from named organic chemistry reactions, identifying reagents and isomers as plausible impurities, and handling missing or conflicting information. A new dataset of 1 124 043 simulated spectra that were generated to train these models is also distributed with this work. These findings demonstrate that deductive bottlenecks for chemical problems are not fundamentally insuperable for ML models.

A Human Feedback Strategy for Photoresponsive Molecules in Drug Delivery: Utilizing GPT-2 and Time-Dependent Density Functional Theory Calculations

Photoresponsive drug delivery stands as a pivotal frontier in smart drug administration, leveraging the non-invasive, stable, and finely tunable nature of light-triggered methodologies. The generative pre-trained transformer (GPT) has been employed to generate molecular structures. In our study, we harnessed GPT-2 on the QM7b dataset to refine a UV-GPT model with adapters, enabling the generation of molecules responsive to UV light excitation. Utilizing the Coulomb matrix as a molecular descriptor, we predicted the excitation wavelengths of these molecules. Furthermore, we validated the excited state properties through quantum chemical simulations. Based on the results of these calculations, we summarized some tips for chemical structures and integrated them into the alignment of large-scale language models within the reinforcement learning from human feedback (RLHF) framework. The synergy of these findings underscores the successful application of GPT technology in this critical domain.

Efficacious human metapneumovirus vaccine based on AI-guided engineering of a closed prefusion trimer

The prefusion conformation of human metapneumovirus fusion protein (hMPV Pre-F) is critical for eliciting the most potent neutralizing antibodies and is the preferred immunogen for an efficacious vaccine against hMPV respiratory infections. Here we show that an additional cleavage event in the F protein allows closure and correct folding of the trimer. We therefore engineered the F protein to undergo double cleavage, which enabled screening for Pre-F stabilizing substitutions at the natively folded protomer interfaces. To identify these substitutions, we developed an AI convolutional classifier that successfully predicts complex polar interactions often overlooked by physics-based methods and visual inspection. The combination of additional processing, stabilization of interface regions and stabilization of the membrane-proximal stem, resulted in a Pre-F protein vaccine candidate without the need for a heterologous trimerization domain that exhibited high expression yields and thermostability. Cryo-EM analysis shows the complete ectodomain structure, including the stem, and a specific interaction of the newly identified cleaved C-terminus with the adjacent protomer. Importantly, the protein induces high and cross-neutralizing antibody responses resulting in near complete protection against hMPV challenge in cotton rats, making the highly stable, double-cleaved hMPV Pre-F trimer an attractive vaccine candidate.