Bidirectional Long Short-Term Memory (BiLSTM) Neural Networks with Conjoint Fingerprints: Application in Predicting Skin-Sensitizing Agents in Natural Compounds

Journal of Chemical Information and Modeling, Journal Year: 2025, Volume and Issue: unknown

Published: March 3, 2025

Skin sensitization, or allergic contact dermatitis, represents a critical end point in toxicity assessment, with profound implications for drug safety and regulatory decision-making. This study aims to develop robust deep-learning-based quantitative structure-activity relationship framework accurately predicting skin sensitization toxicity, particularly the context of natural-product-derived compounds. To achieve this, we explored advanced recurrent neural network architectures, including long short-term memory (LSTM), bidirectional LSTM (BiLSTM), gated unit (GRU), GRU, model intricate structure-toxicity relationships inherent molecular We aim optimize improve predictive performance by training cohort 55 models diverse set fingerprints. Notably, BiLSTM model, which integrates SMILES tokens RDKit fingerprints, achieved superior performance, underscoring its capability effectively capture key determinants sensitization. An extensive applicability domain analysis coupled an in-depth evaluation feature importance provided new insights into attributes that influence propensity. further evaluated using natural product data set, where it demonstrated exceptional generalization capabilities. The accuracy 86.5%, Matthews correlation coefficient 75.2%, sensitivity 100%, area under curve 88%, specificity 75%, F1-score 88.8%. Remarkably, categorized products discriminating sensitizing from non-sensitizing agents across various subcategories. These results underscore potential BiLSTM-based as powerful silico tools modern discovery efforts assessments, especially field products.

Language: Английский

Protecting your skin: a highly accurate LSTM network integrating conjoint features for predicting chemical-induced skin irritation

Journal of Cheminformatics, Journal Year: 2025, Volume and Issue: 17(1)

Published: March 27, 2025

Abstract Skin irritation is a significant adverse effect associated with chemicals and drug substances. Quantitative structure-activity relationship (QSAR) an alternative method bypassing in vivo assay for filling data gaps chemical risk assessment. In this study, we developed QSAR models based on recurrent neural networks (RNNs) to classify skin caused by compounds. We utilized language notation, molecular substructures, descriptors, combination of these features named conjoint fingerprints model construction. A simple RNN, long short-term memory (LSTM), bidirectional (BiLSTM), gated units (GRU), (BiGRU) architectures were used build the models. found that LSTM descriptors outperformed other significantly 80% accuracy, 60% MCC, 85% AUC external test set evaluation. Thereby, selected generalizability testing sets beyond our ensuring can be sets. Furthermore, applicability domain purposed was developed, enabling trustable prediction will made compound. This OECD guidelines assessment development, assuring compliance all required standards. The source codes study are publicly available, facilitating design safety evaluation, particularly assessing potential chemicals.

Language: Английский

Integrating ensemble machine learning and multi-omics approaches to identify Dp44mT as a novel anti-Candida albicans agent targeting cellular iron homeostasis

Frontiers in Pharmacology, Journal Year: 2025, Volume and Issue: 16

Published: April 24, 2025

Candidiasis, mainly caused by Candida albicans, poses a serious threat to human health. The escalating drug resistance in C. albicans and the limited antifungal options highlight critical need for novel therapeutic strategies. We evaluated 12 machine learning models on self-constructed dataset with known anti-C. activity. Based their performance, optimal model was selected screen our separate in-house compound library unknown activity potential agents. of compounds confirmed through vitro susceptibility assays, hyphal growth biofilm formation assays. Through transcriptomics, proteomics, iron rescue experiments, CTC staining, JC-1 DAPI molecular docking, dynamics simulations, we elucidated mechanism underlying compound. Among models, best predictive an ensemble constructed from Random Forests Categorical Boosting using soft voting. It predicts that Dp44mT exhibits potent tests further verified this finding can inhibit planktonic growth, formation, albicans. Mechanistically, exerts disrupting cellular homeostasis, leading collapse mitochondrial membrane ultimately causing apoptosis. This study presents practical approach predicting com-pounds provides new insights into development homeostasis

Language: Английский