Biology subjects

Zhang, Y.-z.

Publications and source records attributed to Zhang, Y.-z..

17 recordsLinked to original sources

bioRxiv2026

PHI-Reason: evidence-grounded species-level phage-host prediction from structured biological text profiles

Biological interaction inference often requires integrating heterogeneous evidence that differs in biological meaning, provenance, specificity, and reliability. Existing computational approaches typically compress such evidence into numerical features, aggregate scores or latent representations before prediction, obscuring the contribution of individual evidence records and limiting interpretation when evidence is conflicting or incomplete. Phage-host interaction prediction provides a demanding test case for this problem, requiring the integration of genomic, functional, and reference-derived evidence with differing coverage, specificity, and reliability. Here we introduce a structured-evidence inference paradigm that retains heterogeneous biological observations as modular, named, and experimentally perturbable records rather than compressing them before prediction. We implement this paradigm in PHI-Reason, which constructs structured evidence profiles from genomic annotations, receptor-binding protein homology, nucleotide-neighbour relationships, alignment-free genomic similarity, and available CRISPR spacer information. A general-purpose large language model (LLM) performs inference directly over these structured profiles without task-specific training or manually engineered evidence-fusion rules. To validate this paradigm, we evaluated it across two distinct biological interaction domains: prokaryotic phage-host prediction and eukaryotic virus-host prediction. Across phage-host benchmarks, PHI-Reason achieved species-level top-1 accuracies of 63.6% and 53.2% on RefSeq-634 and VHDB-3150, respectively, and a multi-host accuracy of 0.571 on the Hi-C cohort, outperforming established numerical methods. Besides, systematic evidence perturbations quantified how individual evidence supported, complemented, or misled inference, identifying nucleotide-neighbour context as the dominant signal. Analyses of LLM intermediate representations using a target-conditioned local Jacobian readout, together with output-grounding analyses, further characterized evidence-dependent inference and quantified the extent to which generated rationales departed from the supplied evidence profiles. Applying the same framework to eukaryotic virus-host prediction using domain-appropriate evidence also preserved high predictive accuracy (66.6%), supporting the generalizability of the proposed paradigm across biological prediction. These results demonstrate that LLMs can serve as an evidence-grounded inference interface, integrating heterogeneous biological evidence while making the boundaries of evidential support directly testable.

bioinformatics↗

bioRxiv2026

RiboPipe: efficient per-transcript codon-resolution ribo-seq coverage imputation for low-coverage transcripts

MotivationRibosome profiling (Ribo-seq) provides codon-resolution measurements of translation; however, many transcripts exhibit sparse or low read coverage, which limits downstream quantitative analyses. Reliable prediction and imputation of codon-resolution coverage for low-coverage transcripts remain computationally challenging. ResultsWe present RiboPipe, an efficient framework for per-transcript codon-resolution Ribo-seq coverage imputation for low-coverage transcripts. RiboPipe is designed around three key principles. First, it jointly optimizes transcript-level mean ribosome load (MRL) prediction and codon-level coverage modeling within a unified objective, enabling consistent learning across both local and transcript-level scales. Second, it introduces a peak-weighted loss that emphasizes high-signal codon positions associated with translational pausing, improving the recovery of functionally relevant coverage peaks. Third, the framework is lightweight and data-efficient, achieving stable performance even when trained on only a small fraction of high-coverage transcripts. Based on two publicly available Ribo-seq datasets (GSE233886 and GSE133393) as instances, we demonstrate stable convergence and consistent prediction accuracy across multiple train-test split ratios. Comparative evaluation of embedding strategies shows that simple one-hot representations achieve superior performance compared with pre-trained language model embeddings under identical training conditions. Overall, RiboPipe provides a computationally efficient and scalable framework for Ribo-seq coverage imputation in low-coverage transcripts. Availability and ImplementationThe source code and associated data can be accessed at https://github.com/yaozhong/riboPipe Contactyaozhong@ims.u-tokyo.ac.jp or imoto@hgc.jp

bioinformatics↗

bioRxiv2025

Phylogenetic tree-aware positive-unlabeled deep metric learning for phage-host interaction identification

Phages are viruses that infect bacteria and play essential roles in shaping microbial communities. Identifying phage-host interactions (PHIs) is crucial for understanding infection dynamics and developing phage-based therapeutic strategies. Recent deep learning approaches have shown great promise for PHI prediction; however, their performance remains constrained by the limited number of experimentally validated positive pairs and the overwhelming abundance of unlabeled or non-validated samples. Moreover, most existing models overlook higher-level phylogenetic relationships among hosts, which could provide valuable structural priors for guiding representation learning. To address these challenges, we propose a phylogenetic tree-aware positive-unlabeled deep metric learning framework for phage-host interaction (PHI) identification. Unlike traditional approaches that train classification models to strictly separate positive and negative phage-host pairs, the proposed method learns representations under supervision from both confirmed positive PHIs and host phylogenetic tree constraints on non-positive samples. The proposed method can seamlessly formalize contrastive learning and deep metric learning within the same framework that explicitly optimizes PHI encoders with biological constraints in the learning functions. We show that this metric learning formulation outperforms conventional contrastive learning approaches that enforce separation between positive and negative samples without consistently aligning the learned representations with evolutionary distances. Experiments on the Cherry benchmark dataset and metagenome Hi-C multi-host dataset demonstrate that our approach enhances species-level prediction accuracy, improves cross-host generalization, and yields more interpretable representations of phage-host relationships.

bioinformatics↗

bioRxiv2025

Rethinking Representation Complexity in Drug-Target Prediction via Supervised Vector Quantization

Accurate prediction of drug-target interactions (DTIs) is crucial for computational drug discovery. Although pretrained language models offer richer molecular and protein representations, their increasing complexity does not always lead to better predictive performance. In many cases, the inclusion of redundant or irrelevant features may obscure biologically relevant patterns. In this study, we systematically evaluate the contribution of complex features in DTI prediction and demonstrate that only a portion of these features is truly informative. Based on this insight, we propose a Vector Quantization (VQ)-based module that functions as a plug-and-play feature selection layer within deep learning architectures. When combined with a simple fully connected classifier, this supervised VQ (SVQ) framework not only surpasses recent state-of-the-art DTI methods in performance, but also enhances interpretability through the learning of discriminative codewords. This work highlights the importance of input feature selection in deep learning and offers a new perspective for constructing robust and interpretable DTI prediction models.

bioinformatics↗

bioRxiv2024

Tissue-Specific Cell Type Annotation with Supervised Representation Learning using Split Vector Quantization and Its Comparisons with Single-cell Foundation Models

Cell-type annotation in single-cell data involves identifying and labeling the cell types based on their gene expression profiles or molecular features. Recently, with advances in single-cell foundation models (FMs), unsupervised annotation and transfer learning with FMs have been explored for cell-type annotation tasks. However, because FMs are usually pre-trained in an unsupervised manner on data spanning a wide variety of tissues and cell types, their representations for specific tissues may lack specificity and become overly generalized. In this work, we propose a novel supervised representation learning method using split-vector-quantization, single-cell Vector-Quantization Classifier (scVQC). We evaluated scVQC against both supervised and unsupervised representation learning approaches, with a focus on foundation models pretrained on large-scale single-cell datasets, such as scBERT and scGPT. The experimental results highlight the importance of label supervision in cell-type annotation tasks and demonstrate that the learned codebook effectively profiles and distinguishes different cell types.

bioinformatics↗

bioRxiv2024

Weakly Supervised Vector Quantization for Whole Slide Images Classification

Whole Slide Images (WSIs) are gigapixel, high-resolution digital scans of microscope slides, providing detailed tissue profiles for pathological analysis. Due to their gigapixel size and lack of detailed annotations, Multiple Instance Learning (MIL) becomes the primary technique for WSI analysis. However, current MIL methods for WSIs directly use embeddings extracted by a pretrained vision encoder, which are not task-specific and often exhibit high variability. To address this, we introduce a novel method, VQ-MIL, which maps the embeddings to a discrete space using weakly supervised vector quantization to refine the embeddings and reduce the variability. Additionally, the discrete embeddings from our methods provides clearer visualizations compared to other methods. Our experiments show that VQ-MIL achieves state-of-the-art classification results on two benchmark datasets. The source code is available at https://github.com/aCoalBall/VQMIL.

pathology↗

bioRxiv2024

Enhanced spatially resolved transcriptomics analysis by matching between expression profiles and spatial topology

Spatially resolved transcriptomics (SRT) quantifies gene expression covering contextual heterogeneity for the tissue section. Exploratory data analysis using SRT data sheds light on diverse biomedical research fields. We propose STForte, a pairwise graph autoencoding-based approach for SRT data analysis, which is capable of matching the information between expression profiles and spatial topology in the latent space. STForte benefits from the designed framework to provide encodings with justifiable spatial correlations for the down-stream analysis of both homogeneous and heterogeneous SRT data. Moreover, STForte can unravel the biological patterns of unobserved locations or recover deficient measurements to enable spatial enhancement. Latent encodings generated by STForte can be used to perform spatial region identification or other downstream tasks to gain biological insights and elucidate biological processes. In this work, we presented various analyses to show that STForte is scalable for analyzing SRT data under different scenarios with considerable performance.

bioinformatics↗

bioRxiv2024

PharaCon: A new framework for identifying bacteriophages via conditional representation learning

MotivationIdentifying bacteriophages (phages) within metagenomic sequences is essential for understanding microbial community dynamics. Transformer-based foundation models have been successfully employed to address various biological challenges. However, these models are typically pre-trained with self-supervised tasks that do not consider label variance in the pre-training data. This presents a challenge for phage identification as pre-training on mixed bacterial and phage data may lead to information bias due to the imbalance between bacterial and phage samples. ResultsTo overcome this limitation, this study proposed a novel conditional BERT framework that incorporates labels during pre-training. We developed an approach using a conditional BERT model for pre-training labeled data, incorporating label constraints with modified language modeling tasks. This approach allows the model to acquire label-conditional sequence representations. Additionally, we proposed a solution that utilizes conditional BERT in the fine-tuning phase as a classifier. We applied this conditional BERT framework to identify phages using a novel fine-tuning strategy, introducing PharaCon. We evaluated PharaCon against several existing methods on both simulated sequence datasets and real metagenomic contig datasets. The results demonstrate PharaCon's potential as an effective and efficient method for phage identification, highlighting the effectiveness of conditional B ERT as a solution for learning label-specific representations during pre-training on mixed sequence data. AvailabilityThe codes of PharaCon are now available in: https://github.com/Celestial-Bai/PharaCon. Contactyaozhong@ims.u-tokyo.ac.jp and imoto@hgc.jp

bioinformatics↗

bioRxiv2024

Prognostic Significance of preoperative serum CA125, CA19-9, CA72-4, CEA, and AFP in Patients with Endometrial cancer

ObjectiveTo determine preoperative serum CA125, CA19-9, CA72-4, CEA, and AFP with prognostic value, and to establish a risk score based on CA125, CEA, AFP levels for predicting the overall survival (OS) and progression-free survival (PFS) of endometrial cancer (EC) patients. MethodsA retrospective cohort study with 2081 EC patients was conducted at Shengjing Hospital of China Medical University. Patient baseline information, tumor characteristics, and data on five serum biomarkers (CA125, CA19-9, CA72-4, CEA, and AFP) were collected. Hazard ratios (HRs) and 95% confidence intervals (CIs) were determined using univariate or multivariate Cox proportional hazard models. log-rank test and Kaplan-Meier analysis were used to compared survival, Data were randomly divided into a training cohort (50%, N = 1041) and an external validation cohort (50%, n = 1040). the least absolute shrinkage and selection operator (Lasso)-Cox regression model was used to screen the independent factors for establishing risk score. And develop nomograms for survival rate prediction. ResultsMultivariate analysis showed Elevated CA125 (P<0.0001) AFP (P <0.0001) and CEA(P=0.037) were identified as independent biomarkers for PFS. Increased CA125 (P = 0.003) AFP (P <0.0001) and CEA(P=0.014) were independent factors associated with OS. CA125, AFP and CEA were thus incorporated in an innovative Risk score (RS) by Lasso-Cox regression model, The RS was also an independent indicator for PFS (P<0.0001) and OS (P<0.0001). Furthermore, we developed and validated nomogram based on Cox regression models. The discriminative ability and calibration of the nomograms revealed good predictive ability, as indicated by the calibration plots. ConclusionThis study suggests that the risk score based on preoperative serum levels of CA125, CEA, and AFP was prognostic biomarkers for predicting progression-free survival and overall survival for EC patients. Nomograms based on the RS and clinicopathological features accurately predict Prognosis of EC patients.

cancer biology↗

bioRxiv2023

Predicting cell types with supervised contrastive learning on cells and their types

Single-cell RNA-sequencing (scRNA-seq) is a powerful technique that provides high-resolution expression profiling of individual cells. It significantly advances our understanding of cellular diversity and function. Despite its potential, the analysis of scRNA-seq data poses considerable challenges relate to multicollinearity, data imbalance, and batch effect. One of the pivotal tasks in single-cell data analysis is cell type annotation, which classifies cells into discrete types based on their gene expression profiles. In this work, we propose a novel modeling formalism for cell type annotation with a supervised contrastive learning method, named SCLSC (Supervised Contrastive Learning for Single Cell). Different from the previous usage of contrastive learning in single cell data analysis, we employed the contrastive learning for instance-type pairs instead of instance-instance pairs. More specifically, in the cell type annotation task, the contrastive learning is applied to learn cell and cell type representation that render cells of the same type to be clustered in the new embedding space. Through this approach, the knowledge derived from annotated cells is transferred to the feature representation for scRNA-seq data. The whole training process becomes more efficient when conducting contrastive learning for cell and their types. Our experiment results demonstrate that the proposed SCLSC method consistently achieves superior accuracy in predicting cell types compared to four state-of-the-art methods. SCLSC also performs well in identifying cell types in different batch groups. The simplicity of our method allows for scalability, making it suitable for analyzing datasets with a large number of cells. In a real-world application of SCLSC to monitor the dynamics of immune cell subpopulations over time, SCLSC demonstrates a capability to discriminate cell subtypes of CD19+ B cells that were not present in the training dataset.

bioinformatics↗

bioRxiv2022

Dysfunctional analysis of the pre-training model on nucleotide sequences and the evaluation of different k-mer embeddings

Pre-training has attracted much attention in recent years. Although significant performance improvements have been achieved in many downstream tasks using pre-training, the mechanism of how a pre-training method works for downstream tasks is not fully illustrated. In this work, focusing on nucleotide sequences, we decompose a pre-training model of Bidirectional Encoder Representations from Transformers (BERT) into embedding and encoding modules to illustrate what a pre-trained model learns from pre-training data. Through dysfunctional analysis on both data and model level, we demonstrate that the context-consistent k-mer representation is the primary product that a typical BERT model learns in the embedding layer. Surprisingly, single usage of the k-mer embedding pre-trained on the random data can achieve comparable performance to that of the k-mer embedding pre-trained on actual biological sequences. We further compare the learned k-mer embeddings with other commonly used k-mer representations in downstream tasks of sequence-based functional predictions and propose a novel solution to accelerate the pre-training process. Contactyaozhong@ims.u-tokyo.ac.jp or imoto@hgc.jp Supplementary informationThe source code and relevant data are available at https://github.com/yaozhong/bert_investigation.

bioinformatics↗

bioRxiv2022

Microbial Gene Ontology informed deep neural network for microbe functionality discovery in human diseases

BackgroundThe human microbiome plays a crucial role in human health and is associated with a number of human diseases. Determining microbiome functional roles in human diseases remains a biological challenge due to the high dimensionality of metagenome gene features. However, existing models were limited in providing biological interpretability, where the functional role of microbes in human diseases is unexplored. Here we propose to utilize a neural network-based model incorporating Gene Ontology (GO) relationship network to discover the microbe functionality in human diseases. ResultsWe use four benchmark datasets, including diabetes, liver cirrhosis, inflammatory bowel disease, and colorectal cancer, to explore the microbe functionality in the human diseases. Our model discovered and visualized the novel candidates important microbiome genes and their functions by calculating the important score of each gene and GO term in the network. Furthermore, we demonstrate that our model achieves a competitive performance in predicting the disease by comparison with other non-Gene Ontology informed models. ConclusionsThe discovered candidates important microbiome genes and their functions provide novel insights into microbe functional contribution.

bioinformatics↗

bioRxiv2021

Identification of Bacteriophages Using Deep Representation Model with Pre-training

MotivationBacteriophages/Phages are the viruses that infect and replicate within bacteria and archaea, and rich in human body. To investigate the relationship between phages and microbial communities, the identification of phages from metagenome sequences is the first step. Currently, there are two main methods for identifying phages: database-based (alignment-based) methods and alignment-free methods. Database-based methods typically use a large number of sequences as references; alignment-free methods usually learn the features of the sequences with machine learning and deep learning models. ResultsWe propose INHERIT which uses a deep representation learning model to integrate both database-based and alignment-free methods, combining the strengths of both. Pre-training is used as an alternative way of acquiring knowledge representations from existing databases, while the BERT-style deep learning framework retains the advantage of alignment-free methods. We compare INHERIT with four existing methods on a third-party benchmark dataset. Our experiments show that INHERIT achieves a better performance with the F1-score of 0.9932. In addition, we find that pre-training two species separately helps the non-alignment deep learning model make more accurate predictions. AvailabilityThe codes of INHERIT are now available in: https://github.com/Celestial-Bai/INHERIT. Contactyaozhong@ims.u-tokyo.ac.jp and imoto@hgc.jp Supplementary informationSupplementary data are available at BioRxiv online.

bioinformatics↗

bioRxiv2021

NF-kB-repressed Sirt3 mediates testicular cholesterol metabolism and cytoskeleton assembly by P450scc/SOD2 deacetylation

Testicular homeostasis requires the balanced interplay between specific molecules in Sertoli cells, Leydig cells, germ cells. Loss of this coordination can lead to the disruption of spermatogenesis, even male infertility. By operating the upregulation and downregulation of Sirt3 in our male subfertility rats model and two testicular cells models, we indicated that Sirt3 overexpression and activator ameliorated cholesterol metabolism via P450scc deacetylation in Leydig cells, and cytoskeleton assembly via PDLIM1 with SOD2 deacetylation in Sertoli cells and elongating spermatids. In terms of the upstream regulator of Sirt3, the phosphorylation of NF-{kappa}B p65Ser536 stimulated the nuclear translocation of NF-{kappa}B subunits (p50, p65, RelB), which bound to TFBS1 and TFBS2 synchronously in the promoter of Sirt3, repressing Sirt3 transcription. This study demonstrates that NF-{kappa}B-repressed SIRT3 acts directly on cholesterol metabolism of Leydig cells and cytoskeleton assembly of Sertoli cells via P450scc/SOD2 deacetylation to regulate sperm differentiation, influencing spermatogenesis, even male fertility. Research organism: Rat, mouse

developmental biology↗

bioRxiv2021

On the application of BERT models for nanopore methylation detection

MotivationDNA methylation is a common epigenetic modification, which is widely associated with various biological processes, such as gene expression, aging, and disease. Nanopore sequencing provides a promising methylation detection approach through monitoring abnormal signal shifts for detecting modified bases in target motif regions. Recently, model-based approaches, especially those with deep learning models, have achieved significant performance improvements on nanopore methylation detection. In this work, we explore using bidirectional encoder representations from transformers (BERT) for doing the task, which can provide non-recurrent neural structures for fast parallel computation. ResultsWe find original BERT architecture does not work as well as the bidirectional recurrent neural network (biRNN) on the nanopore methylation prediction task. Through further analysis, we observe recurrent patterns of positional-signal-shift in the context window surrounding target 5-methylcytosine (5mC) and N6-methyladenine (6mA) motifs. We propose a refined BERT with relative position representation and center hidden units concatenation, which takes account of task-specific characters into modeling. We perform systematic evaluations in-sample and cross-sample. The experiment results show that the refined BERT model can achieve competitive or even better results than the state-of-the-art biRNN model, while the model inference speed is about 6x faster. Besides, on the cross-sample evaluation of datasets from the different research groups, BERT models demonstrate a good generalization performance. AvailabilityThe source code and data are available at https://github.com/yaozhong/methBERT Contactyaozhong@ims.u-tokyo.ac.jp

bioinformatics↗

bioRxiv2019

Nanopore base-calling from a perspective of instance segmentation

BackgroundNanopore sequencing is a rapidly developing third-generation sequencing technology, which can generate long nucleotide reads of molecules within a portable device in real time. Through detecting the change of ion currency signals during a DNA/RNA fragments pass through a nanopore, genotypes are determined. Currently, the accuracy of nanopore base-calling has a higher error rate than short-read base-calling. Through utilizing deep neural networks, the-state-of-the art nanopore base-callers achieve base-calling accuracy in a range from 85% to 95%.\n\nResultIn this work, we proposed a novel base-calling approach from a perspective of instance segmentation. Different from the previous sequence labeling approaches, we formulated the base-calling problem as a multi-label segmentation task. Meanwhile, we proposed a refined U-net model which we call UR-net that can model sequential dependencies for a one-dimensional segmentation task. The experiment results show that the proposed base-caller URnano achieves competitive results compared to recently proposed CTC-featured base-caller Chiron, on the same amount of training and test data for in-domain evaluation. Our results show that formulating the base-calling problem as a one-dimensional segmentation task is a promising approach.\n\nAvailabilityThe source code and data are available at https://github.com/yaozhong/URnano\n\nContactyaozhong@ims.u-tokyo.ac.jp\n\nSupplementary informationSupplementary data are available at attachment online.

bioinformatics↗

bioRxiv2019

Intra-bin structural variant detection for whole genome sequencing data using U-net

MotivationFor short-read sequencing, read-depth based structural variant (SV) callers are difficult to find single-nucleotide-resolution breakpoints due to the bin-size limitation. ResultsIn this paper, we present RDBKE to enhance the breakpoint resolution of read-depth SV callers using deep segmentation model UNet. We show that UNet can be trained with a small amount of data and applied for breakpoint enhancement both in-sample and cross-sample. On both simulation and real data, RDBKE significantly increases the number of SVs with more precise breakpoints. Availabilitysource code of RDBKE is available at https://github.com/yaozhong/deepIntraSV Contactyaozhong@ims.u-tokyo.ac.jp

bioinformatics↗