Article
Biochemical Research Methods
Fengqing Lu, Mufei Li, Xiaoping Min, Chunyan Li, Xiangxiang Zeng
Summary: This study introduces a computational framework called DLGN for generating bioactive molecules towards two specific targets. DLGN utilizes adversarial training and reinforcement learning to explore chemical spaces and encourage the generation of molecules that belong to the intersection of two bioactive compound distributions. The proposed model shows promise in generating novel compounds with high similarity to multiple bioactive datasets.
BRIEFINGS IN BIOINFORMATICS
(2021)
Article
Engineering, Chemical
Yifan Liu, Runze Liu, Jinyu Duan, Li Wang, Xiangwen Zhang, Guozhu Li
Summary: Commercial fuel discovery is facing decreasing return of investment due to stricter environmental criteria and reducing potential uses for each new fuel. This study proposes a deep generative model called LIGANDS, which screens desired fuel molecules in a large chemical space without manually setting design rules. LIGANDS integrates a variational autoencoder, a generative adversarial network, and a stacking model to generate new fuel molecules with similar properties and improved energy performance. The model imitates key properties of target fuel to expand and enrich the fuel-relevant chemical space with innovative molecular entities.
CHEMICAL ENGINEERING SCIENCE
(2023)
Article
Pharmacology & Pharmacy
Wentao Shi, Manali Singha, Gopal Srivastava, Limeng Pu, J. Ramanujam, Michal Brylinski
Summary: Computational modeling is crucial in modern drug discovery, particularly in predicting binding molecules. Pocket2Drug is a promising computational approach that uses data mining and machine learning to predict binding molecules for a given ligand binding site.
FRONTIERS IN PHARMACOLOGY
(2022)
Article
Chemistry, Physical
Yong Zhao, Edirisuriya M. Dilanga Siriwardane, Zhenyao Wu, Nihang Fu, Mohammed Al-Fahdi, Ming Hu, Jianjun Hu
Summary: We propose a deep learning based Physics Guided Crystal Generative Model (PGCGM) for efficient crystal material design. Our model increases the generation validity by more than 700% compared to FTCP, one of the latest structure generators and by more than 45% compared to our previous CubicGAN model. The generated structures are validated using Density Functional Theory (DFT) calculations, with 1869 out of 2000 materials successfully optimized and deposited into the Carolina Materials Database, showing thermodynamic stability and potential synthesizability with negative formation energy and energy-above-hull less than 0.25 eV/atom for 39.6% and 5.3% of the materials, respectively.
NPJ COMPUTATIONAL MATERIALS
(2023)
Article
Chemistry, Multidisciplinary
Wei Lu, David L. Kaplan, Markus J. Buehler
Summary: This study proposes a custom generative language model to design novel spider silk protein sequences with complex combinations of target mechanical properties. The model is fine-tuned on major ampullate spidroin (MaSp) sequences and enables the creation of silk sequences with unique combinations of properties. The study provides insights into the mechanistic roles of sequence patterns in achieving key mechanical properties and has implications for expanding the silkome dataset and synthetic silk design and optimization.
ADVANCED FUNCTIONAL MATERIALS
(2023)
Article
Engineering, Multidisciplinary
Renjie Lin, Shide Du, Shiping Wang, Wenzhong Guo
Summary: This paper proposes a multi-channel augmented graph embedding convolutional network (MAGEC-Net) and its extended end-to-end model (EMAGEC-Net) for multi-view clustering. The method utilizes generative adversarial networks to obtain augmented graphs and deep fusion networks to fuse the augmented views. Feature extraction is then performed on the fused consistent graphs, resulting in better clustering performance. Experimental results on six real datasets demonstrate the effectiveness and superiority of the proposed method.
IEEE TRANSACTIONS ON NETWORK SCIENCE AND ENGINEERING
(2023)
Article
Automation & Control Systems
Binyu Zhang, Junfeng Wan, Yanyun Zhao, Zhihang Tong, Yunhao Du
Summary: The study introduces a Multi-actor Activity Detection Framework (MADF) for modeling the interactive relationship among multiple actors in extended videos. By utilizing detection, classification, and post-processing stages, the system achieves accurate detection and classification of multi-actor activities.
ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE
(2022)
Article
Engineering, Multidisciplinary
Satheesh C. Chandran, Suraj Kamal, A. Mujeeb, M. H. Supriya
Summary: This paper addresses the challenging task of detecting and classifying passive acoustic targets in the ocean using generative models. It introduces a data-efficient underwater target classifier based on Auxiliary Classifier Generative Adversarial Network (ACGAN), and demonstrates promising results in terms of data efficiency, class confidence, and classification accuracy using target instances collected from the Indian Ocean.
ENGINEERING SCIENCE AND TECHNOLOGY-AN INTERNATIONAL JOURNAL-JESTECH
(2022)
Article
Engineering, Electrical & Electronic
Parinaz Naseri, Sean Hum
Summary: This article introduces a machine learning-based approach to automate the design of dual- and triple-layer metasurfaces, solving the difficulties encountered in traditional synthesis methods. This method allows for the synthesis of thin structures with specific scattering properties, applied in various metasurface applications.
IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION
(2021)
Article
Multidisciplinary Sciences
Jin-Hong Du, Zhanrui Cai, Kathryn Roeder
Summary: In this study, the researchers propose a probabilistic variational autoencoder model, scVAEIT, for integrating and imputing multimodal datasets with mosaic measurements. The model effectively combines different panels of measurements and accurately imputes missing molecular layers. Validation results show that scVAEIT robustly imputes missing modalities and features of cells biologically different from the training data, and it adjusts for batch effects while maintaining biological variation. The study demonstrates that scVAEIT significantly improves integration and imputation across unseen cell types, different technologies, and different tissues.
PROCEEDINGS OF THE NATIONAL ACADEMY OF SCIENCES OF THE UNITED STATES OF AMERICA
(2022)
Article
Computer Science, Interdisciplinary Applications
Fei Hu, Chunlei Wu, Jiangwei Shang, Yiming Yan, Leiquan Wang, Huan Zhang
Summary: This paper explores a flexible approach to generate synthetic realizations with desired geological styles by introducing numerical codes and extending the existing mapping model. The resulting generative model can synthesize images that respect both hard data and exhibit specific geological styles.
COMPUTERS & GEOSCIENCES
(2023)
Article
Computer Science, Artificial Intelligence
Gaoming Yang, Mingwei Li, Xianjing Fang, Ji Zhang, Xingzhu Liang
Summary: Adversarial examples pose a security threat to deep learning models, and the proposed Attack Without a Target Model (AWTM) method achieves high attack success rate with low time cost.
PEERJ COMPUTER SCIENCE
(2021)
Article
Computer Science, Artificial Intelligence
Jianping Gou, Xiangshuo Xiong, Baosheng Yu, Lan Du, Yibing Zhan, Dacheng Tao
Summary: Knowledge distillation is an effective technique for compressing deep models by transferring knowledge from a large teacher model to a small student model. Existing methods mainly focus on unidirectional knowledge transfer, overlooking the effectiveness of students' self-reflection in real-world education scenarios. To address this, we propose a new framework called MTKD-SSR that enhances the teacher's ability to transfer knowledge and improves the student's capacity to absorb knowledge through self-reflection.
INTERNATIONAL JOURNAL OF COMPUTER VISION
(2023)
Article
Mathematical & Computational Biology
Sofia D'Souza, K. V. Prema, S. Balaji, Ronak Shah
Summary: Chemogenomics, or proteochemometrics, uses computational methods to predict drug-target interactions based on large-scale data. This study develops a deep learning CNN model using one-dimensional SMILES for drugs and protein binding pocket sequences as inputs to predict unknown ligand-target interactions. The proposed method outperforms shallow machine learning methods in terms of prediction accuracy and computational efficiency.
INTERDISCIPLINARY SCIENCES-COMPUTATIONAL LIFE SCIENCES
(2023)
Article
Automation & Control Systems
Nikita O. Starodubcev, Nikolay O. Nikitin, Elizaveta A. Andronova, Konstantin G. Gavaza, Denis O. Sidorenko, Anna Kalyuzhnaya
Summary: Generative design techniques have been widely applied in various fields, particularly in engineering, to automate initial stages of designing structures, minimizing routine work. However, existing approaches are limited by problem specificity and lack flexibility in method selection. To address these issues, a general approach named GEFEST (Generative Evolution For Encoded STructure) was proposed, providing sampling, estimation, and optimization principles for adaptable problem solutions. Experimental studies confirmed the effectiveness of GEFEST in coastal engineering, microfluidics, thermodynamics, and oil field planning, achieving improvements of 12%, 9%, 8%, and 7% respectively over baseline and state-of-the-art solutions.
ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE
(2023)
Article
Biochemistry & Molecular Biology
Atsushi Yoshimori, Filip Miljkovic, Juergen Bajorath
Summary: Deep machine learning is applied to expand the capacity of computational compound design for covalent protein kinase inhibitors. A computational approach combining fragment-based design and deep generative modeling augmented by three-dimensional pharmacophore screening is devised. The approach is successfully applied to generate novel candidate inhibitors for Bruton's tyrosine kinase.
Article
Biochemistry & Molecular Biology
Raquel Rodriguez-Perez, Juergen Bajorath
Summary: Support Vector Machine (SVM) algorithm, as one of the most widely used machine learning methods, has been applied in predicting active compounds and molecular properties for over a decade. It operates in feature spaces of increasing dimensionality, distinguishing itself from many other methods. SVM is applicable to compound classification, ranking, multi-class predictions, and regression modeling. It remains relevant in the emerging era of deep learning and stands as one of the premier ML methods in chemoinformatics.
JOURNAL OF COMPUTER-AIDED MOLECULAR DESIGN
(2022)
Editorial Material
Chemistry, Medicinal
Atsushi Yoshimori, Hengwei Chen, Juergen Bajorath
FUTURE MEDICINAL CHEMISTRY
(2023)
Article
Biochemistry & Molecular Biology
Alec Lamens, Jurgen Bajorath
Summary: Compounds with multiple targets (MT-CPDs) are important in drug discovery, and computational approaches have been used to design or identify them. Machine learning models have been derived to distinguish between MT-CPDs and compounds with single-target activity (ST-CPDs). Surprisingly, it was found that models derived for ST-CPDs can accurately predict MT-CPDs, revealing a relationship between the two types of compounds.
Article
Biochemistry & Molecular Biology
Wojciech Pietrus, Rafal Kurczab, Dawid Warszycki, Andrzej J. J. Bojarski, Jurgen Bajorath
Summary: In this study, a total of 898 F-containing isomeric analog sets were identified and analyzed for structure-activity relationship (SAR) in the ChEMBL database. The results showed significant differences in affinity for some isomeric compounds against different aminergic GPCRs, and the change of fluorine position could lead to a significant change in potency. Additionally, a computational workflow was proposed to score the fluorine positions in the molecule.
Editorial Material
Chemistry, Medicinal
Juergen Bajorath
FUTURE MEDICINAL CHEMISTRY
(2023)
Article
Chemistry, Medicinal
Elena Xerxa, Filip Miljkovic, Juergen Bajorath
Summary: Large-scale analysis of public human and mouse protein kinase inhibitor (PKI) data revealed a significant number of human and mouse PKIs with reliable activity measurements. Human PKIs were found to exhibit coverage against a large number of kinases, with a substantial growth in the past years. Covalent PKIs containing acrylamide or heterocyclic urea warheads were particularly abundant, and they showed higher potency compared to structurally analogous PKIs. These findings have important implications for medicinal chemistry.
JOURNAL OF MEDICINAL CHEMISTRY
(2023)
Article
Biochemistry & Molecular Biology
Atsushi Yoshimori, Juergen Bajorath
Summary: In the field of drug design, very few studies have attempted the prediction of new active compounds from protein sequence data. This is mainly due to the challenging nature of this prediction task, as global protein sequence similarity has strong evolutionary and structural implications but is not directly related to ligand binding. However, the application of deep language models adapted from natural language processing provides new opportunities to attempt such predictions by linking amino acid sequences and chemical structures through textual molecular representations. In this study, a biochemical language model with a transformer architecture, named Motif2Mol, was introduced for the prediction of new active compounds based on sequence motifs of ligand binding sites. In a proof-of-concept application on inhibitors of more than 200 human kinases, Motif2Mol exhibited promising learning characteristics and an unprecedented ability to consistently reproduce known inhibitors of different kinases.
Article
Biochemistry & Molecular Biology
Tiago Janela, Kosuke Takeuchi, Juergen Bajorath
Summary: Prediction of potency of bioactive compounds is typically done using linear or nonlinear quantitative structure-activity relationship (QSAR) models. In this study, a novel approach called structure-potency fingerprint (SPFP) is introduced, which combines structural features and potency values into a single bit string representation. By using a conditional variational autoencoder (CVAE) with SPFPs, the potency module of test compounds can be accurately predicted using only their structure module. The SPFP-CVAE approach achieves comparable accuracy to support vector regression (SVR) and deep neural networks in predicting compounds' activity classes and potency values.
Article
Chemistry, Medicinal
Nicola Gambacorta, Fulvio Ciriaco, Nicola Amoroso, Cosimo Damiano Altomare, Ju''rgen Bajorath, Orazio Nicolotti
Summary: In this study, a CIRCE compound prediction platform was developed based on explainable machine learning to support the design of selective CB1R and CB2R ligands. The platform achieved about 80% accuracy in test calculations and provided rationalized explanations for the predictions. CIRCE is freely available as a web-based prediction platform.
JOURNAL OF CHEMICAL INFORMATION AND MODELING
(2023)
Article
Biochemistry & Molecular Biology
Alec Lamens, Juergen Bajorath
Summary: In pharmaceutical research, black box predictions from machine learning models hinder their application in guiding experimental work. This study presents a test system that combines two explainable artificial intelligence methods to better understand prediction outcomes and provide chemically intuitive explanations for model decisions.
Article
Biochemistry & Molecular Biology
Elena Xerxa, Oliver Laufkoetter, Jurgen Bajorath
Summary: Protein kinase inhibitors (PKIs) are widely studied as drug candidates in drug discovery, with increasing interest in allosteric PKIs (APKIs) and covalent PKIs (CPKIs). The popularity of PKIs for therapeutic intervention has led to the accumulation of large volumes of activity data, enabling large-scale computational analysis. CPKIs with certain reactive groups showed significantly higher potency compared to non-covalent PKIs, but there was no general increase in promiscuity. Additionally, new APKIs have been identified, allowing for structure-based investigation of PK inhibition by covalent-allosteric mechanisms.
Article
Medicine, Research & Experimental
Elena Xerxa, Juergen Bajorath
Summary: The aim of this study was to generate high-quality data sets of protein kinase inhibitors (PKIs). Publicly available PKIs with reliable activity data were curated, and weakly active PKIs were classified as inactive. Analogue series and PKIs containing reactive groups enabling covalent inhibition were systematically identified. The study obtained a total of 155,579 human and 3057 mouse PKIs, with human PKIs showing activity against 440 kinases and including 13,949 covalent PKIs. The collection of qualifying PKIs and corresponding inactive compounds is made available as an open access deposition.
Review
Chemistry, Medicinal
Juergen Bajorath
Summary: This article discusses the opportunities and methods of applying chemical language models (CLMs) in drug discovery. CLMs can be developed using recurrent neural networks or transformer architectures, and attention mechanisms are used to improve predictive performance. CLMs can be used for constrained generative modeling and the prediction of chemical reactions or drug-target interactions. Since CLMs can learn mappings of different types of sequences and are applicable to any compound or target data presented in a sequential format and tokenized, they have a wide range of applications.
MOLECULAR INFORMATICS
(2023)
Article
Computer Science, Artificial Intelligence
Andrea Mastropietro, Giuseppe Pasculli, Juergen Bajorath
Summary: In drug design, graph neural networks (GNNs) are used to predict compound potency by analyzing graph representations of protein-ligand interactions. The study reveals that GNNs are influenced by ligand memorization during learning and certain GNN architectures prioritize interaction information for predicting high affinities. While GNNs do not comprehensively account for protein-ligand interactions and physical reality, they provide a helpful balance between ligand memorization and learning of interaction patterns.
NATURE MACHINE INTELLIGENCE
(2023)