Journal of Cheminformatics最新文献

Chemical Language Models (CLMs) have demonstrated capabilities in extracting patterns and predicting from vast volume of the Simplified Molecular Input Line Entry System (SMILES), a notation used to represent molecular structures. Different CLMs, developed from various architectures, can provide unique insights into molecular properties. To harness the uniqueness of different CLMs, we propose FusionCLM, a novel stacking-ensemble learning algorithm that integrate the outputs of multiple CLMs into a unified framework. FusionCLM first generates SMILES embeddings, predictions, and losses from each CLM. Auxiliary models are trained on these first-level predictions and embeddings to estimate test losses during inference. The losses and predictions are then concatenated to create an integrated feature matrix, which trains second-level meta-models for final predictions. Empirical testing on five datasets demonstrates that FusionCLM have better performance than individual CLM at the first level and three advanced multimodal deep learning frameworks, showcasing FusionCLM’s potential in advancing molecular property prediction.

Quantitative structure–activity relationship (QSAR) modeling has become a critical tool in drug design. Recently proposed Topological Regression (TR), a computationally efficient and highly interpretable QSAR model that maps distances in the chemical domain to distances in the activity domain, has shown predictive performance comparable to state-of-the-art deep learning-based models. However, TR’s dependence on simple random sampling-based anchor selection and utilization of radial basis function for response reconstruction constrain its interpretability and predictive capacity. To address these limitations, we propose Adaptive Topological Regression (AdapToR) with adaptive anchor selection and optimization-based reconstruction. We evaluated AdapToR on the NCI60 GI50 dataset, which consists of over 50,000 drug responses across 60 human cancer cell lines, and compared its performance to Transformer CNN, Graph Transformer, TR, and other baseline models. The results demonstrate that AdapToR outperforms competing QSAR models for drug response prediction with significantly lower computational cost and greater interpretability as compared to deep learning-based models.

Retrosynthesis—the process of deconstructing complex molecules into simpler, more accessible precursors—is a cornerstone of drug discovery and material design. While machine learning has improved single-step retrosynthesis prediction, generating complete multi-step retrosynthetic routes remains challenging. In this study, we explore the integration of single-step retrosynthesis models with various planning algorithms to improve multi-step retrosynthetic route generation. We expand the exploration space beyond previously limited settings by incorporating combinations of planning algorithms and single-step retrosynthesis models and diverse datasets, enabling a more comprehensive assessment of retrosynthetic strategies. We evaluated synthetic routes based on both solvability, the ability to generate a complete route, and route feasibility, which reflects their practical executability in the laboratory. Our findings show that the model combination with the highest solvability does not always produce the most feasible routes, underscoring the need for more nuanced evaluation. Through a systematic analysis of combinations of planning algorithms and single-step retrosynthesis models, their performance across different datasets, and various practical metrics, our study provides a more comprehensive evaluation of retrosynthetic planning strategies. These insights contribute to a better understanding of computational retrosynthesis and its alignment with real-world applicability.

When data availability is limited, the prediction of properties through purely data-driven machine learning (ML) is challenging. Integrating physically-based modeling techniques into ML methods may lead to better performance. In a recent work by Chew et al. (“Advancing material property prediction: using physics-informed machine learning models for viscosity”) descriptors from classical molecular dynamics (MD) simulations were included into a quantitative structure–property relationship to accurately predict temperature-dependent viscosity of pure liquids. Through feature importance analysis, the authors found that heat of vaporization was the most relevant descriptor for the prediction of viscosity. In this comment, we would like to discuss the physical origin of this finding by referring to Eyring’s rate theory, and develop an alternative modeling approach using a thermodynamic-based architecture that requires less input data.

Cyclic peptides are promising drug candidates due to their ability to modulate intracellular protein–protein interactions, a property often inaccessible to small molecules. However, their typically poor membrane permeability limits therapeutic applicability. Accurate computational prediction of permeability can accelerate the identification of cell-permeable candidates, reducing reliance on time-consuming and costly experimental screening. Although deep learning has shown potential in predicting molecular properties, its application in permeability prediction remains underexplored. A systematic evaluation of these models is important to assess current capabilities and guide future development. In this study, we conduct a comprehensive benchmark of 13 machine learning models for predicting cyclic peptide membrane permeability. These models cover four types of molecular representations: fingerprints, SMILES strings, molecular graphs, and 2D images. We use experimentally measured PAMPA permeability data from the CycPeptMPDB database, comprising nearly 6000 cyclic peptides, and evaluate performance across three prediction tasks: regression, binary classification, and soft-label classification. Two data-splitting strategies, random split and scaffold split, are used to assess the generalizability of trained models. Our results show that model performance depends strongly on molecular representation and model architecture. Graph-based models, particularly the Directed Message Passing Neural Network (DMPNN), consistently achieve top performance across tasks. Regression generally outperforms classification. Scaffold-based splitting, although intended to more rigorously assess generalization, yields substantially lower model generalizability compared to random splitting. Comparing prediction errors with experimental variability highlights the practical value of current models while also indicating room for further improvement.

Bitter peptides (BPs), derived from the hydrolysis of proteins in food, play a crucial role in both food science and biomedicine by influencing taste perception and participating in various physiological processes. Accurate identification of BPs is essential for understanding food quality and potential health impacts. Traditional machine learning approaches for BP identification have relied on conventional feature descriptors, achieving moderate success but struggling with the complexities of biological sequence data. Recent advances utilizing protein language model embedding and meta-learning approaches have improved the accuracy, but frequently neglect the molecular representations of peptides and lack interpretability. In this study, we propose xBitterT5, a novel multimodal and interpretable framework for BP identification that integrates pretrained transformer-based embeddings from BioT5+ with the combination of peptide sequence and its SELFIES molecular representation. Specifically, incorporating both peptide sequences and their molecular strings, xBitterT5 demonstrates superior performance compared to previous methods on the same benchmark datasets. Importantly, the model provides residue-level interpretability, highlighting chemically meaningful substructures that significantly contribute to its bitterness, thus offering mechanistic insights beyond black-box predictions. A user-friendly web server (https://balalab-skku.org/xBitterT5/) and a standalone version (https://github.com/cbbl-skku-org/xBitterT5/) are freely available to support both computational biologists and experimental researchers in peptide-based food and biomedicine.

Accurate chemical reaction prediction is critical for reducing both cost and time in drug development. This study introduces ReactionT5, a transformer-based chemical reaction foundation model pre-trained on the Open Reaction Database—a large publicly available reaction dataset. In benchmarks for product prediction, retrosynthesis, and yield prediction, ReactionT5 outperformed existing models. Specifically, ReactionT5 achieved 97.5% accuracy in product prediction, 71.0% in retrosynthesis, and a coefficient of determination of 0.947 in yield prediction. Remarkably, ReactionT5, when fine-tuned with only a limited dataset of reactions, achieved performance on par with models fine-tuned on the complete dataset. Additionally, the visualization of ReactionT5 embeddings illustrates that the model successfully captures and represents the chemical reaction space, indicating effective learning of reaction properties.

Graphical Abstract

Purpose

Drug-induced liver injury (DILI) is a significant concern in drug development, often leading to the discontinuation of clinical trials and the withdrawal of drugs from the market. This study explores the application of graph neural networks (GNNs) for DILI prediction, using molecular graph representations as the primary input.

Methods

We evaluated several GNN architectures, including Graph Convolutional Networks (GCNs), Graph Attention Networks (GATs), Graph Sample and Aggregation (GraphSAGE), and Graph Isomorphism Networks (GINs), using the latest FDA DILI dataset and other molecular property prediction datasets. We introduce a novel approach that creates a custom graph dataset, driven by molecular optimisation, that incorporates detailed and realistic chemical features such as bond lengths and partial charges as input into the GNN models. We have named our model approach DILIGeNN.

Results

DILIGeNN achieved an AUC of 0.897 on the DILI dataset, surpassing the current state-of-the-art model in the DILI prediction task. Furthermore, DILIGeNN outperformed the state-of-the-art in other graph-based molecular prediction tasks, achieving an AUC of 0.918 on the Clintox dataset, 0.993 on the BBBP dataset, and 0.953 on the BACE dataset, indicating strong generalisation and performance across different datasets.

Conclusion

DILIGeNN, utilising a single graph representation as input, outperforms the state-of-the-art methods in DILI prediction that incorporate both molecular fingerprint and graph-structured data. These findings highlight the effectiveness of our molecular graph generation and the GNN training approach as a powerful tool for early-stage drug development and drug repurposing pipeline.

Scientific Contribution: DILIGeNN is a GNN framework that extracts graph features from 3D optimised molecular structures as is done in target-based drug discovery and molecular docking simulation. Our method is the first to encode spatial and electrostatic information into a single graph representation, as opposed to other work that require multiple graphs or additional chemical descriptors for feature representation. Our approach, using warm starts following repeated early stopping during training, outperforms the current state-of-the-art methods in liver toxicity (DILI), permeability (BBBP) and activity (BACE) prediction tasks.

Graphic Abstract

Inference of molecules with desired activities/properties is one of the key and challenging issues in cheminformatics and bioinformatics. For that purpose, our research group has recently developed a state-of-the-art framework mol-infer for molecular inference. This framework first constructs a prediction function for a fixed property using machine learning models, which is then simulated by mixed-integer linear programming to infer desired molecules. The accuracy of the framework heavily relies on the representation power of the descriptors. In this study, we highlight a typical class of non-isomorphic chemical graphs with reasonably different property values that cannot be distinguished by the standard “two-layered (2L) model" of mol-infer. To address this distinguishability problem of the 2L model, we propose a novel family of descriptors, named cycle-configuration (CC), which captures the notion of ortho/meta/para patterns that appear in aromatic rings, which was impossible in the framework so far. Extensive computational experiments show that with the new descriptors, we can construct prediction functions with similar or better performance for all 44 tested chemical properties, including 27 regression datasets and 17 classification datasets comparing with our previous studies, confirming the effectiveness of the CC descriptors. For inference, we also provide a system of linear constraints to formulate the CC descriptors as linear constraints. We demonstrate that a chemical graph with up to 50 non-hydrogen vertices can be inferred within a practical time frame.