Biology subjects

Yu, D.-J.

Publications and source records attributed to Yu, D.-J..

10 recordsLinked to original sources

bioRxiv2025

De novo design of functionally diverse and druggable antimicrobial peptides by diffusion and multimodal deep learning

The escalating crisis of antibiotic resistance underscores the urgent need for innovative anti-infective agents, such as antimicrobial peptides (AMPs), though their discovery and optimization remain challenging. To address this, we introduce AMP-D3, a comprehensive framework that integrates AMP generation, identification, and screening into a cohesive workflow. Central to AMP-D3 is PACD, an advanced generator that utilizes contrastive diffusion to produce diverse, functionally tailored AMPs. For precise recognition, the CAST model combines ESM-2 embeddings with localized sequence features through cross-modal attention, yielding an accuracy of 0.891--a 2.2% enhancement over current state-of-the-art approaches. Additionally, a robust multi-attribute prediction module evaluates key properties, including antifungal and anticancer potential, across twenty-two distinct characteristics. Furthermore, a meticulously designed screening pipeline identifies high-efficacy AMP candidates against critical pathogens, such as C. albicans, E. coli, P. aeruginosa, and S. aureus. This process assessed 10,000 peptides generated by PACD alongside 43,000 from prior studies, selecting the top 100 candidates per pathogen for further analysis. Experimental validation confirms that these peptides exhibit structural diversity and binding affinities comparable to naturally occurring AMPs, as evidenced by AlphaFold2 structural predictions and molecular docking studies. Thus, AMP-D3 represents a transformative approach to the design and expedited discovery of clinically viable AMPs.

bioinformatics↗

bioRxiv2025

A Deep Learning-based Method for Drug Molecule Representation and Property Prediction

Accurately and robustly representing drug molecule features, prediction of drug-target biomacromolecule interactions, and determining drug molecule physicochemical properties are crucial in drug development. However, due to issues such as insufficient generalization ability of single-modal representation, lack of multi-task prediction frameworks, and weak adaptability in cold-start scenarios, these tasks remain challenging. Here, we introduce DrugDL, a framework designed for drug molecule representation and the prediction of multiple downstream tasks, including drug-target interactions, binding affinities, binding sites, physicochemical properties, toxicity, and drug-drug interactions. DrugDL achieves joint representation learning of the drug chemical space and the target protein biological space and analyzes the multi-scale interaction mechanisms between drug molecules and target proteins by introducing cross-modal contrastive learning and single-modal feature enhancement algorithms. It employs a multi-task prediction framework to predict multiple properties of drug molecules. In practical applications, DrugDL outperforms state-of-the-art methods, especially in cold-start tasks. Its successfully applied to high-throughput screening, identifying inhibitors for SARS-CoV-2 and metabolic enzymes, and aids in predicting cancer-targeted drugs. Validations for EGFR and ALK targets confirm its efficiency as a precise drug discovery tool. Leveraging accurate molecular representation and multi-property prediction, DrugDL provides full-chain technical support for drug development, significantly accelerating the drug discovery process.

bioinformatics↗

bioRxiv2025

MKFGO: Integrating Multi-Source Knowledge Fusion with Pre-Trained Language Model for High-Accuracy Protein Function Prediction

Accurately identifying protein functions is essential to understand life mechanisms and thus advance drug discovery. Although biochemical experiments are the gold standard for determining protein functions, they are often time-consuming and labor-intensive. Here, we proposed a novel composite deep-learning method, MKFGO, to infer Gene Ontology (GO) attributes through integrating five complementary pipelines built on multi-source biological data. MKFGO was rigorously benchmarked on 1522 non-redundant proteins, demonstrating superior performance over 11 state-of-the-art function prediction methods. Comprehensive data analyses revealed that the major advantage of MKFGO lies in its two deep-learning components, HFRGO and PLMGO, which derive handcraft features and protein large language model (PLM)-based features, respectively, from protein sequences in different biological views, with effective knowledge fusion at the decision-level. HFRGO leverages an LSTM-attention network embedded with handcraft features, in which the triplet loss-based guilt-by-association strategy is designed to enhance the correlation between feature similarity and function similarity. PLMGO employs the PLM to capture feature embeddings with discriminative functional patterns from sequences. Meanwhile, another three components provide complementary insights for further improving prediction accuracy, driven by protein-protein interaction, GO term probability, and protein-coding gene sequence, respectively. The source codes and models of MKFGO are freely available at https://github.com/yiheng-zhu/MKFGO.

bioinformatics↗

bioRxiv2025

GerNA-Bind: Geometric-enhanced RNA-ligand Binding Specificity Prediction with Deep Learning

RNA molecules are essential regulators of biological processes and promising therapeutic targets for various diseases. Discovering small molecules that selectively bind to specific RNA conformations remains challenging due to RNAs structural complexity and the limited availability of high-resolution data. Herein, we introduce GerNA-Bind, a geometric deep learning framework to predict RNA-ligand binding specificity by integrating multi-state RNA-ligand representations and interactions. GerNA-Bind achieves state-of-the-art performance on multiple benchmark datasets and excels in predicting interactions for low-homology RNA-ligand pairs. It achieves a 20.8% improvement in precision for binding site prediction compared to AlphaFold3. Furthermore, it offers informative, well-calibrated predictions with built-in uncertainty quantification. In a large-scale virtual screening application, GerNA-Bind identified 18 structurally diverse compounds targeting the oncogenic MALAT1 RNA, with experimentally confirmed submicromolar affinities. Among them, one lead compound selectively binds the MALAT1 triple-helix, reduces its transcript levels, and inhibits cancer cell migration. These findings highlight GerNA-Binds potential as a powerful tool for RNA-focused drug discovery, offering both accuracy and biological insight.

bioinformatics↗

bioRxiv2025

Self-iterative multiple instance learning enables the prediction of CD4+ T cell immunogenic epitopes

Accurately predicting the antigen presentation to CD4+ T cells and subsequent induction of immune response is fundamentally important for vaccine development, autoimmune disease treatments, and cancer neoepitope identification. In immunopeptidomics, single-allelic data are highly specific but limited in allele scope, while multi-allelic data contain broader coverage at the cost of weakly labeling. Existing computational approaches either overlook the massive multi-allelic data or introduce label ambiguity due to inadequate modeling strategies. Here, we introduce ImmuScope, a weakly supervised deep-learning framework integrating precise MHC-II antigen presentation, CD4+ T cell epitopes, and immunogenicity predictions. ImmuScope leverages self-iterative multiple-instance learning with positive-anchor triplet loss to explore peptide-MHC-II (pMHC-II) binding from weakly labeled multi-allelic data and single-allelic data, comprising over 600,000 ligands across 142 alleles. Moreover, ImmuScope can also interpret the MHC-II binding specificity and motif deconvolution of immunopeptidomics data. We successfully applied ImmuScope to discover melanoma neoantigens, revealing variations in pMHC-II binding and immunogenicity upon epitope mutations. We further employed ImmuScope to assess the effects of SARS-CoV-2 epitope mutations on immune escape, with its predictions aligned well with experimentally determined immune escape dynamics. Overall, ImmuScope provides a comprehensive solution for CD4+ T cell antigen recognition and immunogenicity assessment, with broad potential for advancing vaccine design and personalized immunotherapy.

bioinformatics↗

bioRxiv2023

ConBoTNet: supervised contrastive learning enhances MHC-II peptide binding affinity prediction

Accurate prediction of major histocompatibility complex (MHC)-peptide binding affinity can improve our understanding of cellular immune responses and guide personalized immunotherapies. Nevertheless, the existing deep learning-based approaches for predicting MHC-II peptide interactions fall short of satisfactory performance and offer restricted model interpretability. In this study, we propose a novel deep neural network, termed ConBoTNet, to address the above issues by introducing the designed supervised contrastive learning and bottleneck transformer extractors. Specifically, the supervised contrastive learning pre-training enhances the models representative and generalizable capabilities on MHC-II peptides by pulling positive pairs closer and pushing negative pairs further in the feature space, while the bottleneck transformer module focuses on MHC-II peptide interactions to precisely identify binding cores and anchor positions in an unsupervised manner. Extensive experiments on benchmark datasets under 5-fold cross-validation, leave-one-molecule-out validation, independent testing, and binding core prediction settings highlighted the superiority of our proposed ConBoTNet over current state-of-the-art methods. Data distribution analysis in the latent feature space demonstrated that supervised contrastive learning can aggregate MHC-II-peptide samples with similar affinity labels and learn common features of similar affinity. Additionally, we interpreted the trained neural network by associating the attention weights with peptides and innovatively find both well-established and potential peptide motifs. This work not only introduces an innovative tool for accurately predicting MHC-II peptide affinity, but also provides new insights into a new paradigm for modeling essential biological interactions, advancing data-driven discovery in biomedicine.

bioinformatics↗

bioRxiv2023

ULDNA: Integrating Unsupervised Multi-Source Language Models with LSTM-Attention Network for Protein-DNA Binding Site Prediction

Accurate identification of protein-DNA interactions is critical to understand the molecular mechanisms of proteins and design new drugs. We proposed a novel deeplearning method, ULDNA, to predict DNA-binding sites from protein sequences through a LSTM-attention architecture embedded with three unsupervised language models pretrained in multiple large-scale sequence databases. The method was systematically tested on 1287 proteins with DNA-binding site annotation from Protein Data Bank. Experimental results showed that ULDNA achieved a significant increase of the DNA-binding site prediction accuracy compared to the state-of-the-art approaches. Detailed data analyses showed that the major advantage of ULDNA lies in the utilization of three pre-trained transformer language models which can extract the complementary DNA-binding patterns buried in evolution diversity-based feature embeddings in residue-level. Meanwhile, the designed LSTM-attention network could further enhance the correlation between evolution diversity and protein-DNA interaction. These results demonstrated a new avenue for high-accuracy deep-learning DNA-binding site prediction that is applicable to large-scale protein-DNA binding annotation from sequence alone.

molecular biology↗

bioRxiv2022

Integrating Self-Attention Transformer with Triplet Neural Networks for Protein Gene Ontology Prediction

Accurate identification of protein function is critical to elucidate life mechanism and design new drugs. We proposed a novel deep-learning method, ATGO, to predict Gene Ontology (GO) attributes of proteins through a triplet neural-network architecture embedded with pre-trained self-attention transformer models. The method was systematically tested on 1068 non-redundant benchmarking proteins and 3328 targets from the third Critical Assessment of Protein Function Annotation (CAFA) challenge. Experimental results showed that ATGO achieved a significant increase of the GO prediction accuracy compared to the state-of-the-art approaches in all aspects of molecular function, biological process, and cellular component. Detailed data analyses showed that the major advantage of ATGO lies in the utilization of attention transformer models which can extract discriminative functional pattern from the feature embeddings. Meanwhile, the proposed triplet network helps enhance the association of functional similarity with feature similarity in the sequence embedding space. In addition, it was found that the combination of the network scores with the complementary homology-based inferences could further improve the accuracy and coverage of the predicted models. These results demonstrated a new avenue for high-accuracy deep-learning function prediction that is applicable to large-scale protein function annotations from sequence alone. AvailabilityThe benchmark dataset, standalone package, and online server for ATGO are available at https://zhanggroup.org/ATGO/. Author SummaryIn the post-genome sequencing era, a major challenge in computational molecular biology is to annotate the biological functions of all gene and gene products, which have been classified, in the context of the widely used Gene Ontology (GO), into three aspects of molecular function, biological process, and cellular component. In this work, we proposed a new open-source deep-learning architecture, ATGO, to deduce GO terms of proteins from the primary amino acid sequence, through the integration of the triplet neural-network with attention transformer models. Large benchmark tests showed that, when powered with a pre-trained self-attention transformer model, ATGO achieved a significantly improved performance than other state-of-the-art approaches in all the GO aspect predictions. Following the rapid progress of self-attention neural network techniques, which have demonstrated remarkable impacts on language processing and multi-sensory data process, and most recently on protein structure prediction, this study showed the significant potential of attention transformer models on protein function annotations.

bioinformatics↗

bioRxiv2021

TripletGO: Integrating Transcript Expression Profiles with Protein Homology Inferences for High-Accuracy Gene Function Annotations

Gene Ontology (GO) has been widely used to annotate functions of genes and gene products. We proposed a new method (TripletGO) to deduce GO terms of protein-coding and non-coding genes, through the integration of four complementary pipelines built on transcript expression profiling, genetic sequence alignment, protein sequence alignment and naive probability, respectively. TripletGO was tested on a large set of 5,754 genes from 8 species (human, mouse, arabidopsis, rat, fly, budding yeast, fission yeast, and nematoda) and 2,433 proteins with available expression data from the CAFA3 experiment and achieved function annotation accuracy significantly beyond the current state-of-the-art approaches. Detailed analyses show that the major advantage of TripletGO lies in the coupling of a new triplet-network based profiling method with the feature space mapping technique which can accurately recognize function patterns from transcript expressions. Meanwhile, the combination of multiple complementary models, especially those from transcript expression and protein-level alignments, improves the coverage and accuracy of the final GO annotation results. The standalone package and an online server of TripletGO are freely available at https://zhanglab.ccmb.med.umich.edu/TripletGO/.

bioinformatics↗

bioRxiv2020

Deducing high-accuracy protein contact-maps from a triplet of coevolutionary matrices through deep residual convolutional networks

The topology of protein folds can be specified by the inter-residue contact-maps and accurate contact-map prediction can help ab initio structure folding. We developed TripletRes to deduce protein contact-maps from discretized distance profiles by end-to-end training of deep residual neural-networks. Compared to previous approaches, the major advantage of TripletRes is in its ability to learn and directly fuse a triplet of coevolutionary matrices extracted from the whole-genome and metagenome databases and therefore minimize the information loss during the course of contact model training. TripletRes was tested on a large set of 245 non-homologous proteins from CASP and CAMEO experiments, and outperformed other state-of-the-art methods by at least 58.4% for the CASP 11&12 and 44.4% for the CAMEO targets in the top-L long-range contact precision. On the 31 FM targets from the latest CASP13 challenge, TripletRes achieved the highest precision (71.6%) for the top-L/5 long-range contact predictions. These results demonstrate a novel efficient approach to extend the power of deep convolutional networks for high-accuracy medium- and long-range protein contact-map predictions starting from primary sequences, which are critical for constructing 3D structure of proteins that lack homologous templates in the PDB library. AvailabilityThe training and testing data, standalone package, and the online server for TripletRes are available at https://zhanglab.ccmb.med.umich.edu/TripletRes/. Author SummaryAb initio protein folding has been a major unsolved problem in computational biology for more than half a century. Recent community-wide Critical Assessment of Structure Prediction (CASP) experiments have witnessed exciting progress on ab initio structure prediction, which was mainly powered by the boosting of contact-map prediction as the latter can be used as constraints to guide ab initio folding simulations. In this work, we proposed a new open-source deep-learning architecture, TripletRes, built on the residual convolutional neural networks for high-accuracy contact prediction. The large-scale benchmark and blind test results demonstrate significant advancement of the proposed methods over other approaches in predicting medium- and long-range contact-maps that are critical for guiding protein folding simulations. Detailed data analyses showed that the major advantage of TripletRes lies in the unique protocol to fuse multiple evolutionary feature matrices which are directly extracted from whole-genome and metagenome databases and therefore minimize the information loss during the contact model training.

bioinformatics↗