Shanfeng Zhu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Shanfeng Zhu has a total of 403 co-authors. Here are the top co-authors by number of shared publications:
Ronghui You
School of Statistics and Data Science, LPMC, Nankai University , Tianjin 300071
25 shared publications- Xiaodi Huang
Charles Sturt University
16 shared publications Hiroshi Mamitsuka
Bioinformatics Center, Institute for Chemical Research, Kyoto University , Uji, Kyoto Prefecture 611-0011
12 shared publications- Hiroshi Mamitsuka
Kyoto University
12 shared publications Ziye Wang
School of Mathematical Sciences and LPMC, Nankai University , Tianjin 300071
10 shared publicationsWeiqi Zhai
Institute of Science and Technology for Brain-Inspired Intelligence and and MOE Frontiers Center for Brain Science, Fudan University, Shanghai, 200433 China
8 shared publications
Latest Publications
- Open Access
- 24 October 2026
- Published by Oxford University Press (OUP) in Bioinformatics
- Vol. 42 (11)
Large-scale semantic indexing of Spanish biomedical literature using contrastive transfer learning
Motivation: The exponential growth of biomedical literature has made automatic indexing essential for advancing biomedical research. While automatic indexing has made strides in English biomedical literature, there has been limited research on non-English biomedical texts due to insufficient high-quality training data. Results: We propose BERTDeCS, a novel deep learning framework for automatically indexing Spanish biomedical literature using contrastive transfer learning. BERTDeCS utilizes a multilingual BERT (M-BERT) to generate multi-language representations and adapts M-BERT for Spanish biomedical literature domain through contrastive learning. Additionally, BERTDeCS enhances its semantic indexing capabilities on Spanish biomedical literature by leveraging enriched English annotated literature in MEDLINE through transfer learning. Experimental results on Spanish datasets demonstrate that BERTDeCS outperforms state-of-the-art indexing methods, achieving top performance in the MESINESP and MESINESP2 Tasks on medical semantic indexing in Spanish within the BioASQ challenge. Additional experiments on Portuguese provide preliminary evidence that BERTDeCS can transfer to a closely related language sharing substantial lexical and biomedical terminology with Spanish. Availability: The source code is available at https://github.com/yourh/BERTDeCS and at https://doi.org/10.5281/zenodo.23212468.
- 24 October 2026
- Published by Oxford University Press (OUP) in Bioinformatics
- Vol. 42 (11)
Open AccessLarge-scale semantic indexing of Spanish biomedical literature using contrastive transfer learning
Motivation: The exponential growth of biomedical literature has made automatic indexing essential for advancing biomedical research. While automatic indexing has made strides in English biomedical literature, there has been limited research on non-English biomedical texts due to insufficient high-quality training data. Results: We propose BERTDeCS, a novel deep learning framework for automatically indexing Spanish biomedical literature using contrastive transfer learning. BERTDeCS utilizes a multilingual BERT (M-BERT) to generate multi-language representations and adapts M-BERT for Spanish biomedical literature domain through contrastive learning. Additionally, BERTDeCS enhances its semantic indexing capabilities on Spanish biomedical literature by leveraging enriched English annotated literature in MEDLINE through transfer learning. Experimental results on Spanish datasets demonstrate that BERTDeCS outperforms state-of-the-art indexing methods, achieving top performance in the MESINESP and MESINESP2 Tasks on medical semantic indexing in Spanish within the BioASQ challenge. Additional experiments on Portuguese provide preliminary evidence that BERTDeCS can transfer to a closely related language sharing substantial lexical and biomedical terminology with Spanish. Availability: The source code is available at https://github.com/yourh/BERTDeCS and at https://doi.org/10.5281/zenodo.23212468.
- Open Access
Advancing long-read metagenomic binning via single-copy-gene guided contrastive learning
Long-read sequencing advances metagenomics by producing highly contiguous assemblies and more complete metagenome-assembled genomes (MAGs). However, current long-read metagenomic binners fail to incorporate the rich information of long-read assemblies into representation learning and exhibit limited performance on complex datasets. Here, we show that a higher proportion of long-read assembled contigs contain single-copy genes (SCGs) and more SCGs per contig. Therefore, we developed SCGBinner, which leverages SCG-guided contrastive learning to exploit the advantage of long-read data for learning high-quality contig embeddings. SCGBinner consistently outperforms other binning methods across five simulated and seven real-world long-read datasets, especially on real-world high-diversity samples. For a deep agricultural soil metagenome, SCGBinner recovered 71% more high-quality MAGs and 38% more near-complete MAGs than the second-best method. Notably, SCGBinner uniquely recovered 449 novel high-quality species, which shed light on the predicted ecological roles of 65 uncharacterised families and 93 novel genera. Overall, SCGBinner could harness the potential of long-read sequencing to provide unprecedented insights into the microbial dark matter of complex microbial communities.
Open AccessAdvancing long-read metagenomic binning via single-copy-gene guided contrastive learning
Long-read sequencing advances metagenomics by producing highly contiguous assemblies and more complete metagenome-assembled genomes (MAGs). However, current long-read metagenomic binners fail to incorporate the rich information of long-read assemblies into representation learning and exhibit limited performance on complex datasets. Here, we show that a higher proportion of long-read assembled contigs contain single-copy genes (SCGs) and more SCGs per contig. Therefore, we developed SCGBinner, which leverages SCG-guided contrastive learning to exploit the advantage of long-read data for learning high-quality contig embeddings. SCGBinner consistently outperforms other binning methods across five simulated and seven real-world long-read datasets, especially on real-world high-diversity samples. For a deep agricultural soil metagenome, SCGBinner recovered 71% more high-quality MAGs and 38% more near-complete MAGs than the second-best method. Notably, SCGBinner uniquely recovered 449 novel high-quality species, which shed light on the predicted ecological roles of 65 uncharacterised families and 93 novel genera. Overall, SCGBinner could harness the potential of long-read sequencing to provide unprecedented insights into the microbial dark matter of complex microbial communities.
- Open Access
First community challenge for automated virus taxonomy
The rapid rate of virus discovery renders manual curation by taxonomy experts increasingly impractical, creating a need for reliable software that can reproducibly assign viral contigs to taxa at all fifteen ranks of the virus taxonomy. We led an open community challenge for the computational taxonomic classification of viruses and assembled a dataset of virus sequences combining expert-curated and metagenomic sequences. Seventeen teams contributed a total of thirty-four automated, fully reproducible classification pipelines. Most tools correctly assigned viruses belonging to established species, genera, or families, but viruses that are unclassified at those lower ranks remain challenging. This study provides datasets, open-source software, novel approaches, and recommendations to benchmark computational taxonomic classification of viruses, and support organizing the many viruses discovered in big omics data.
Open AccessFirst community challenge for automated virus taxonomy
The rapid rate of virus discovery renders manual curation by taxonomy experts increasingly impractical, creating a need for reliable software that can reproducibly assign viral contigs to taxa at all fifteen ranks of the virus taxonomy. We led an open community challenge for the computational taxonomic classification of viruses and assembled a dataset of virus sequences combining expert-curated and metagenomic sequences. Seventeen teams contributed a total of thirty-four automated, fully reproducible classification pipelines. Most tools correctly assigned viruses belonging to established species, genera, or families, but viruses that are unclassified at those lower ranks remain challenging. This study provides datasets, open-source software, novel approaches, and recommendations to benchmark computational taxonomic classification of viruses, and support organizing the many viruses discovered in big omics data.
- Open Access
MyoScore: a Genetically Anchored Transcriptomic Scoring System for Quantifying Human Skeletal Muscle Health
Muscle health varies continuously from optimal function to severe pathology, yet no unified genetic framework quantifies this spectrum objectively. Here we develop MyoScore, a transcriptomic scoring system derived from transcriptome-wide association studies of 27 muscle-related phenotypes in over one million participants. TWAS selects genes whose genetically regulated expression in skeletal muscle associates with muscle-related traits, providing the genetic anchoring of the scoring system, while MyoScore itself is computed from measured bulk RNA-seq expression in new samples. From 1,116 transcriptome-wide association study (TWAS)-significant genes, 417 are expressed in skeletal muscle and form the basis of the scoring system. These genes are organized into five dimensions of muscle biology (Strength, Mass, LeanMuscle, Youth and Resilience), each scored from 0 to 100. Across 1,722 human skeletal muscle transcriptomes from four independent cohorts, MyoScore defines a continuous four-stage muscle health spectrum, discriminates healthy from diseased muscle (area under the curve 0.751–0.873), and correlates with histopathological severity, quantitative MRI and clinical outcomes. Functional validation through iPSC-to-myotube differentiation supports predicted expression changes for novel MyoScore genes. UK Biobank analysis of blood biomarker proxies in 467,123 participants demonstrates concordant associations with muscle phenotypes, and two-sample Mendelian randomization using skeletal muscle cis-eQTL supports causal directionality for 78% of gene-outcome pairs tested. Single-cell validation across 475,584 cells from two independent muscle ageing atlases shows that pseudobulk MyoScore declines with age, with type II myofibre nuclei most affected. Together, MyoScore establishes the first genetically anchored, dimension-resolved quantification of human muscle health, enabling objective assessment, patient stratification and biomarker discovery across the full spectrum from optimal function to severe disease.
Open AccessMyoScore: a Genetically Anchored Transcriptomic Scoring System for Quantifying Human Skeletal Muscle Health
Muscle health varies continuously from optimal function to severe pathology, yet no unified genetic framework quantifies this spectrum objectively. Here we develop MyoScore, a transcriptomic scoring system derived from transcriptome-wide association studies of 27 muscle-related phenotypes in over one million participants. TWAS selects genes whose genetically regulated expression in skeletal muscle associates with muscle-related traits, providing the genetic anchoring of the scoring system, while MyoScore itself is computed from measured bulk RNA-seq expression in new samples. From 1,116 transcriptome-wide association study (TWAS)-significant genes, 417 are expressed in skeletal muscle and form the basis of the scoring system. These genes are organized into five dimensions of muscle biology (Strength, Mass, LeanMuscle, Youth and Resilience), each scored from 0 to 100. Across 1,722 human skeletal muscle transcriptomes from four independent cohorts, MyoScore defines a continuous four-stage muscle health spectrum, discriminates healthy from diseased muscle (area under the curve 0.751–0.873), and correlates with histopathological severity, quantitative MRI and clinical outcomes. Functional validation through iPSC-to-myotube differentiation supports predicted expression changes for novel MyoScore genes. UK Biobank analysis of blood biomarker proxies in 467,123 participants demonstrates concordant associations with muscle phenotypes, and two-sample Mendelian randomization using skeletal muscle cis-eQTL supports causal directionality for 78% of gene-outcome pairs tested. Single-cell validation across 475,584 cells from two independent muscle ageing atlases shows that pseudobulk MyoScore declines with age, with type II myofibre nuclei most affected. Together, MyoScore establishes the first genetically anchored, dimension-resolved quantification of human muscle health, enabling objective assessment, patient stratification and biomarker discovery across the full spectrum from optimal function to severe disease.
- Open Access
On the state of protein function prediction: a report on the fourth CAFA challenge
Background: The Critical Assessment of Functional Annotation (CAFA) is a community effort held to understand the field of computational protein function prediction. Every three years, since 2010, the organizers initiate an experiment to collect function predictions on a large set of proteins and then evaluate the performance of predicting methods on a subset of proteins that have accumulated experimental annotations between the submission deadline and the evaluation time. CAFA provides an independent and rigorous assessment of the current state of the art, thus leveling the playing field, highlighting successes, revealing bottlenecks, and offering a forum for the exchange of ideas in protein science. Here, we report the results of the fourth CAFA experiment (CAFA4). Results: CAFA4 featured the participation of 148 methods from 70 research groups on a total of 46,205 unique proteins over a 5-year annotation accumulation phase, the longest in any CAFA. In a comparison across CAFA2-CAFA4 methods, the prediction of Gene Ontology (GO) terms has clearly improved across all three GO aspects and traditional evaluation settings. While not achieving the first rank, several CAFA2 and CAFA3 methods featured in the top ten methods in many evaluations, suggesting that earlier methods still hold relevance. The performance is weaker in the newly introduced “partial knowledge” evaluation category (proteins with experimental annotations before submission deadline that gained additional annotations in the same GO aspect during the annotation accumulation phase), highlighting the need for a new class of methods. The rankings of the methods were stable over the years in traditional evaluation settings, but less so in the new partial knowledge evaluation. Overall, the field continues to progress with some influx of new participants. Sustained efforts will be necessary to substantially advance it.
Open AccessOn the state of protein function prediction: a report on the fourth CAFA challenge
Background: The Critical Assessment of Functional Annotation (CAFA) is a community effort held to understand the field of computational protein function prediction. Every three years, since 2010, the organizers initiate an experiment to collect function predictions on a large set of proteins and then evaluate the performance of predicting methods on a subset of proteins that have accumulated experimental annotations between the submission deadline and the evaluation time. CAFA provides an independent and rigorous assessment of the current state of the art, thus leveling the playing field, highlighting successes, revealing bottlenecks, and offering a forum for the exchange of ideas in protein science. Here, we report the results of the fourth CAFA experiment (CAFA4). Results: CAFA4 featured the participation of 148 methods from 70 research groups on a total of 46,205 unique proteins over a 5-year annotation accumulation phase, the longest in any CAFA. In a comparison across CAFA2-CAFA4 methods, the prediction of Gene Ontology (GO) terms has clearly improved across all three GO aspects and traditional evaluation settings. While not achieving the first rank, several CAFA2 and CAFA3 methods featured in the top ten methods in many evaluations, suggesting that earlier methods still hold relevance. The performance is weaker in the newly introduced “partial knowledge” evaluation category (proteins with experimental annotations before submission deadline that gained additional annotations in the same GO aspect during the annotation accumulation phase), highlighting the need for a new class of methods. The rankings of the methods were stable over the years in traditional evaluation settings, but less so in the new partial knowledge evaluation. Overall, the field continues to progress with some influx of new participants. Sustained efforts will be necessary to substantially advance it.
- 28 April 2026
- Published by Elsevier in Pattern Recognition
- Vol. 179
MoE-TCR: Mixture-of-experts framework for pan-specific TCR-epitope binding prediction
- 28 April 2026
- Published by Elsevier in Pattern Recognition
- Vol. 179
MoE-TCR: Mixture-of-experts framework for pan-specific TCR-epitope binding prediction
- 18 April 2026
- Published by Springer Nature in Health Information Science and Systems
- Vol. 14 (1) , 58
PhenoRareAI: Phenotype-based intelligent diagnosis for rare neuromuscular disorders of glycogen storage disease and spinal muscular atrophy
Rare neuromuscular diseases pose significant diagnostic challenges due to their genetic complexity and varied clinical presentations. Current diagnostic methods are often costly and inefficient, with primary care physician slacking adequate knowledge and effective diagnostic algorithms needed for accurate identification. This study addresses these challenges by developing PhenoRareAI, a model designed to improve the diagnosis of rare neuromuscular diseases such as Glycogen Storage Disease II and Spinal Muscular Atrophy. A total of 109 electronic health records from Huashan Hospital and external centers were used to develop and validate PhenoRareAI. The model integrates phenotype recognition algorithms and disease priority ranking models, which were rigorously tested against baseline models and validated with external datasets. The key outcome measures included the accuracy of phenotype recognition and disease diagnosis. PhenoRareAI demonstrated signifi cant improvement in phenotype recognition, achieving a micro F1-score and a macro F1-score of 27.40% and 25.16%, compared to the baseline PhenoPro model's scores of 11.44% and 14.65%. In terms of disease prioritization, the model achieved a top 1 ranking for 15 cases of Glycogen Storage Disease II and 17 cases of Spinal Muscular Atrophy after phenotype enhancement, with a mean reciprocal rank (MRR) of 37.37 and 57.21, respectively, significantly outperforming existing methods. PhenoRareAI signifi cantly enhances the diagnosis process for rare neuromuscular diseases, demonstrating strong potential for clinical application. These findings highlight the importance of integrating AI-driven diagnostic tools into healthcare to better navigate the complexities of rare disease diagnosis. The online version contains supplementary material available at 10.1007/s13755-026-00455-w.
- 18 April 2026
- Published by Springer Nature in Health Information Science and Systems
- Vol. 14 (1) , 58
PhenoRareAI: Phenotype-based intelligent diagnosis for rare neuromuscular disorders of glycogen storage disease and spinal muscular atrophy
Rare neuromuscular diseases pose significant diagnostic challenges due to their genetic complexity and varied clinical presentations. Current diagnostic methods are often costly and inefficient, with primary care physician slacking adequate knowledge and effective diagnostic algorithms needed for accurate identification. This study addresses these challenges by developing PhenoRareAI, a model designed to improve the diagnosis of rare neuromuscular diseases such as Glycogen Storage Disease II and Spinal Muscular Atrophy. A total of 109 electronic health records from Huashan Hospital and external centers were used to develop and validate PhenoRareAI. The model integrates phenotype recognition algorithms and disease priority ranking models, which were rigorously tested against baseline models and validated with external datasets. The key outcome measures included the accuracy of phenotype recognition and disease diagnosis. PhenoRareAI demonstrated signifi cant improvement in phenotype recognition, achieving a micro F1-score and a macro F1-score of 27.40% and 25.16%, compared to the baseline PhenoPro model's scores of 11.44% and 14.65%. In terms of disease prioritization, the model achieved a top 1 ranking for 15 cases of Glycogen Storage Disease II and 17 cases of Spinal Muscular Atrophy after phenotype enhancement, with a mean reciprocal rank (MRR) of 37.37 and 57.21, respectively, significantly outperforming existing methods. PhenoRareAI signifi cantly enhances the diagnosis process for rare neuromuscular diseases, demonstrating strong potential for clinical application. These findings highlight the importance of integrating AI-driven diagnostic tools into healthcare to better navigate the complexities of rare disease diagnosis. The online version contains supplementary material available at 10.1007/s13755-026-00455-w.
- Open Access
- 1 December 2025
- Published by Springer Nature in Nature Communications
- Vol. 16 (1) , 11198
VACmap: an accurate long-read aligner for unraveling complex genomic rearrangements
Sequence alignment is essential for genomic research and clinical diagnostics, yet detecting complex rearrangements such as inversions, duplications, and gene conversions remains challenging due to allele complexity and limitations of current methods. We introduce VACmap, a non-linear mapping approach to enhance the detection and representation of all genetic variations. VACmap improves duplication detection from 20% to 90% in the Challenging Medically-Relevant Genes (CMRG) benchmark and improves characterization of complex inversions in repetitive regions and gene conversion events. It improves resolving clinically significant loci, including the LPA gene (with repetitive KIV-2 units linked to coronary heart disease), GBA1 and STRC genes (risk factors for Parkinson’s disease and hearing loss, respectively, affected by pseudogene recombination with GBAP1 and STRCP1). Here, we show that VACmap delivers better alignment accuracy and SV detection, providing a robust tool for genomic analysis and clinical insights, with potential to advance understanding of genetic diversity and disease mechanisms.
- 1 December 2025
- Published by Springer Nature in Nature Communications
- Vol. 16 (1) , 11198
Open AccessVACmap: an accurate long-read aligner for unraveling complex genomic rearrangements
Sequence alignment is essential for genomic research and clinical diagnostics, yet detecting complex rearrangements such as inversions, duplications, and gene conversions remains challenging due to allele complexity and limitations of current methods. We introduce VACmap, a non-linear mapping approach to enhance the detection and representation of all genetic variations. VACmap improves duplication detection from 20% to 90% in the Challenging Medically-Relevant Genes (CMRG) benchmark and improves characterization of complex inversions in repetitive regions and gene conversion events. It improves resolving clinically significant loci, including the LPA gene (with repetitive KIV-2 units linked to coronary heart disease), GBA1 and STRC genes (risk factors for Parkinson’s disease and hearing loss, respectively, affected by pseudogene recombination with GBAP1 and STRCP1). Here, we show that VACmap delivers better alignment accuracy and SV detection, providing a robust tool for genomic analysis and clinical insights, with potential to advance understanding of genetic diversity and disease mechanisms.
- 21 November 2025
- Published by Springer Nature in Health Information Science and Systems
- Vol. 14 (1) , 4
Vaner2: towards more general biomedical named entity recognition using multi-task large language model encoders
Biomedical named entity recognition (BioNER) serves as the foundation for many downstream tasks, such as relation extraction, question answering, and clinical text analysis. BioNER was previously dominated by BERT-based models pretrained on large biomedical corpora. However, BERT-based models finetuned on specific BioNER datasets exhibit limited ability to generalize to other datasets. With the recent advances of large language models (LLMs), several works have fine-tuned autoregressive LLMs that are not inherently suitable for BioNER tasks, which limits model performance. In this study, building upon our previous work VANER, we utilized LLMs with the causal attention mask removed as a text encoder for sequence labeling. Using 39 BioNER datasets, we trained a multi-task NER model that extracts all entity types with one LLM forward pass. We also proposed a token-wise loss rescaling technique to deal with the data imbalance between different tags and entity types. Extensive experiments on independent test datasets demonstrated that our VANER2 model achieved the best generalization results compared with BERT-based baselines and several recent BioNER methods. VANER2 is freely available at https://github.com/ZhuLab-Fudan/VANER2.
- 21 November 2025
- Published by Springer Nature in Health Information Science and Systems
- Vol. 14 (1) , 4
Vaner2: towards more general biomedical named entity recognition using multi-task large language model encoders
Biomedical named entity recognition (BioNER) serves as the foundation for many downstream tasks, such as relation extraction, question answering, and clinical text analysis. BioNER was previously dominated by BERT-based models pretrained on large biomedical corpora. However, BERT-based models finetuned on specific BioNER datasets exhibit limited ability to generalize to other datasets. With the recent advances of large language models (LLMs), several works have fine-tuned autoregressive LLMs that are not inherently suitable for BioNER tasks, which limits model performance. In this study, building upon our previous work VANER, we utilized LLMs with the causal attention mask removed as a text encoder for sequence labeling. Using 39 BioNER datasets, we trained a multi-task NER model that extracts all entity types with one LLM forward pass. We also proposed a token-wise loss rescaling technique to deal with the data imbalance between different tags and entity types. Extensive experiments on independent test datasets demonstrated that our VANER2 model achieved the best generalization results compared with BERT-based baselines and several recent BioNER methods. VANER2 is freely available at https://github.com/ZhuLab-Fudan/VANER2.