Mahdi Belcaid
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Mahdi Belcaid has a total of 52 co-authors. Here are the top co-authors by number of shared publications:
- Andrew Johnson
Department of Computer Science, Electronic Visualization Laboratory, University of Illinois, Chicago, Chicago, USA
3 shared publications Barbra D. Ferrell
Department of Computer and Information Sciences, University of Delaware, Newark, DE, United States of America
3 shared publications- Chris North
Department of Computer Science, Virginia Tech, Blacksburg, Virginia, USA
3 shared publications - Jason Leigh
Information an Computer Science/Natural Science/LAVA, University of Hawaii at Manoa, Honolulu, Hawaii, USA
3 shared publications - K. Eric Wommack
Department of Plant and Soil Sciences, University of Delaware, Newark, DE 19716, USA
3 shared publications - Luc Renambot
Department of Computer Science, Electronic Visualization Laboratory, University of Illinois, Chicago, Chicago, USA
3 shared publications
Latest Publications
- Open Access
- 17 March 2026
- Published by Springer Nature in SN Computer Science
- Vol. 7 (3) , 281
Building Collaborative Intelligence: The Translational Journey of the Smart Amplified Group Environment Across Research, Education, and Practice
- 17 March 2026
- Published by Springer Nature in SN Computer Science
- Vol. 7 (3) , 281
Open AccessBuilding Collaborative Intelligence: The Translational Journey of the Smart Amplified Group Environment Across Research, Education, and Practice
- Open Access
- 21 January 2026
- Published by Oxford University Press (OUP) in GigaScience
- Vol. 15
Extending protein language models to a viral genomic scale using biologically induced sparse attention
Aa Background: The transformer architecture in deep learning has revolutionized protein sequence analysis. Recent advancements in protein language models have paved the way for significant progress across various domains, including protein function and structure prediction, multiple sequence alignments, and mutation effect prediction. A protein language model is commonly trained on individual proteins, ignoring the interdependencies between sequences within a genome. However, biological understanding reveals that protein–protein interactions span entire genomic regions, underscoring the limitations of focusing solely on individual proteins. Ab Findings: To address these limitations, we propose a novel approach that extends the context size of transformer models across the entire viral genome. By training on large genomic fragments, our method captures putative long-range dependencies consistent with inter-protein relationships and encodes protein sequences with integrated information from distant proteins within the same genome, offering benefits across downstream tasks. Viruses, with their densely packed genomes, minimal intergenic regions, and protein annotation challenges, are ideal candidates for genome-wide learning. We introduce a long-context protein language model, trained on entire viral genomes, leveraging a biologically informed sparse attention mechanism in which inter-protein links are inferred computationally and used as sparsity priors. Our semi-supervised approach supports long sequences of up to 61,000 amino acids (aa). Ac Conclusion: Our evaluations show improved prediction of masked aa and improved downstream discrimination relative to single-protein models and long-context baselines, with additional validation that our inferred links correlate with independently curated interaction resources.
- 21 January 2026
- Published by Oxford University Press (OUP) in GigaScience
- Vol. 15
Open AccessExtending protein language models to a viral genomic scale using biologically induced sparse attention
Aa Background: The transformer architecture in deep learning has revolutionized protein sequence analysis. Recent advancements in protein language models have paved the way for significant progress across various domains, including protein function and structure prediction, multiple sequence alignments, and mutation effect prediction. A protein language model is commonly trained on individual proteins, ignoring the interdependencies between sequences within a genome. However, biological understanding reveals that protein–protein interactions span entire genomic regions, underscoring the limitations of focusing solely on individual proteins. Ab Findings: To address these limitations, we propose a novel approach that extends the context size of transformer models across the entire viral genome. By training on large genomic fragments, our method captures putative long-range dependencies consistent with inter-protein relationships and encodes protein sequences with integrated information from distant proteins within the same genome, offering benefits across downstream tasks. Viruses, with their densely packed genomes, minimal intergenic regions, and protein annotation challenges, are ideal candidates for genome-wide learning. We introduce a long-context protein language model, trained on entire viral genomes, leveraging a biologically informed sparse attention mechanism in which inter-protein links are inferred computationally and used as sparsity priors. Our semi-supervised approach supports long sequences of up to 61,000 amino acids (aa). Ac Conclusion: Our evaluations show improved prediction of masked aa and improved downstream discrimination relative to single-protein models and long-context baselines, with additional validation that our inferred links correlate with independently curated interaction resources.
- Open Access
Machine Learning-Based Characterization of Bacillus anthracis Phenotypes from pXO1 Plasmid Proteins
The Bacillus anthracis pXO1 plasmid, encoding ~143 proteins, presents a compact model for exploring protein function and evolutionary patterns using protein language models. Due to the organism’s slow evolutionary rate, its limited amino acid variation enhances detection of physiologically relevant patterns in plasmid protein composition. In this study, we applied embedding-based analyses and machine learning methods to characterize pXO1 protein modules across diverse B. anthracis lineages. We generated protein sequence embeddings, constructed phylogenies, and compared plasmid content with whole genome variation. While whole genome and plasmid-based phylogenies diverge, the composition of proteins encoded along the pXO1 plasmid revealed lineage specific structure. Association rule mining combined with decision tree classification produced plasmid-encoded targets for assessing anthrax sublineage, which yielded functionally redundant protein modules that reflected geographic and phylogenetic patterns. A conserved DNA replication module exhibited both shared and B. anthracis lineage specific features. These results show that pXO1 plasmid protein modules contain biologically meaningful and evolutionarily informative signatures, exemplifying their value in phylogeographic characterizations of bacterial pathogens. This framework can be extended to study additional virulence plasmids across Bacillus and other environmental pathogens using scalable protein language model tools.
Open AccessMachine Learning-Based Characterization of Bacillus anthracis Phenotypes from pXO1 Plasmid Proteins
The Bacillus anthracis pXO1 plasmid, encoding ~143 proteins, presents a compact model for exploring protein function and evolutionary patterns using protein language models. Due to the organism’s slow evolutionary rate, its limited amino acid variation enhances detection of physiologically relevant patterns in plasmid protein composition. In this study, we applied embedding-based analyses and machine learning methods to characterize pXO1 protein modules across diverse B. anthracis lineages. We generated protein sequence embeddings, constructed phylogenies, and compared plasmid content with whole genome variation. While whole genome and plasmid-based phylogenies diverge, the composition of proteins encoded along the pXO1 plasmid revealed lineage specific structure. Association rule mining combined with decision tree classification produced plasmid-encoded targets for assessing anthrax sublineage, which yielded functionally redundant protein modules that reflected geographic and phylogenetic patterns. A conserved DNA replication module exhibited both shared and B. anthracis lineage specific features. These results show that pXO1 plasmid protein modules contain biologically meaningful and evolutionarily informative signatures, exemplifying their value in phylogeographic characterizations of bacterial pathogens. This framework can be extended to study additional virulence plasmids across Bacillus and other environmental pathogens using scalable protein language model tools.
- Open Access
Fine-tuning protein language models unlocks the potential of underrepresented viral proteomes
Protein language models (pLMs) have revolutionized computational biology by generating rich protein vector representations, or embeddings—enabling major advancements in de novo protein design, structure prediction, variant effect analysis, and evolutionary studies. Despite these breakthroughs, current pLMs often exhibit biases against proteins from underrepresented species, with viral proteins being particularly affected, frequently referred to as the “dark matter” of the biological world due to their vast diversity and ubiquity, yet sparse representation in training datasets. Here, we show that fine-tuning pre-trained pLMs on viral protein sequences, using diverse learning frameworks and parameter-efficient strategies, significantly enhances representation quality and improves performance on downstream tasks. To support further research, we provide source code for fine-tuning pLMs and benchmarking embedding quality. By enabling more accurate modeling of viral proteins, our approach advances tools for understanding viral biology, combating emerging infectious diseases, and driving biotechnological innovation.
Open AccessFine-tuning protein language models unlocks the potential of underrepresented viral proteomes
Protein language models (pLMs) have revolutionized computational biology by generating rich protein vector representations, or embeddings—enabling major advancements in de novo protein design, structure prediction, variant effect analysis, and evolutionary studies. Despite these breakthroughs, current pLMs often exhibit biases against proteins from underrepresented species, with viral proteins being particularly affected, frequently referred to as the “dark matter” of the biological world due to their vast diversity and ubiquity, yet sparse representation in training datasets. Here, we show that fine-tuning pre-trained pLMs on viral protein sequences, using diverse learning frameworks and parameter-efficient strategies, significantly enhances representation quality and improves performance on downstream tasks. To support further research, we provide source code for fine-tuning pLMs and benchmarking embedding quality. By enabling more accurate modeling of viral proteins, our approach advances tools for understanding viral biology, combating emerging infectious diseases, and driving biotechnological innovation.
- Open Access
Extending Protein Language Models to a Viral Genomic Scale Using Biologically Induced Sparse Attention
The transformer architecture in deep learning has revolutionized protein sequence analysis. Recent advancements in protein language models have paved the way for significant progress across various domains, including protein function and structure prediction, multiple sequence alignments and mutation effect prediction. A protein language model is commonly trained on individual proteins, ignoring the interdependencies between sequences within a genome. However, biological understanding reveals that protein–protein interactions span entire genomic regions, underscoring the limitations of focusing solely on individual proteins. To address these limitations, we propose a novel approach that extends the context size of transformer models across the entire viral genome. By training on large genomic fragments, our method captures long-range interprotein interactions and encodes protein sequences with integrated information from distant proteins within the same genome, offering substantial benefits in various tasks. Viruses, with their densely packed genomes, minimal intergenic regions, and protein annotation challenges, are ideal candidates for genome-wide learning. We introduce a long-context protein language model, trained on entire viral genomes, leveraging a sparse attention mechanism based on protein–protein interactions. Our semi-supervised approach supports long sequences of up to 61,000 amino acids (aa). Our evaluations demonstrate that the resulting embeddings significantly surpass those generated by single-protein models and outperform alternative large-context architectures that rely on static masking or non-transformer frameworks.
Open AccessExtending Protein Language Models to a Viral Genomic Scale Using Biologically Induced Sparse Attention
The transformer architecture in deep learning has revolutionized protein sequence analysis. Recent advancements in protein language models have paved the way for significant progress across various domains, including protein function and structure prediction, multiple sequence alignments and mutation effect prediction. A protein language model is commonly trained on individual proteins, ignoring the interdependencies between sequences within a genome. However, biological understanding reveals that protein–protein interactions span entire genomic regions, underscoring the limitations of focusing solely on individual proteins. To address these limitations, we propose a novel approach that extends the context size of transformer models across the entire viral genome. By training on large genomic fragments, our method captures long-range interprotein interactions and encodes protein sequences with integrated information from distant proteins within the same genome, offering substantial benefits in various tasks. Viruses, with their densely packed genomes, minimal intergenic regions, and protein annotation challenges, are ideal candidates for genome-wide learning. We introduce a long-context protein language model, trained on entire viral genomes, leveraging a sparse attention mechanism based on protein–protein interactions. Our semi-supervised approach supports long sequences of up to 61,000 amino acids (aa). Our evaluations demonstrate that the resulting embeddings significantly surpass those generated by single-protein models and outperform alternative large-context architectures that rely on static masking or non-transformer frameworks.
- Open Access
Fine-Tuning Protein Language Models Unlocks the Potential of Underrepresented Viral Proteomes
Protein language models (pLMs) have revolutionized computational biology by generating rich protein vector representations—or embeddings—enabling major advancements in de novo protein design, structure prediction, variant effect analysis, and evolutionary studies. Despite these breakthroughs, current pLMs often exhibit biases against proteins from underrepresented species, with viral proteins being particularly affected—frequently referred to as the “dark matter” of the biological world due to their vast diversity and ubiquity, yet sparse representation in training datasets. Here, we show that fine-tuning pre-trained pLMs on viral protein sequences, using diverse learning frameworks and parameter-efficient strategies, significantly enhances representation quality and improves performance on downstream tasks. To support further research, we provide source code for fine-tuning pLMs and benchmarking embedding quality. By enabling more accurate modeling of viral proteins, our approach advances tools for understanding viral biology, combating emerging infectious diseases, and driving biotechnological innovation.
Open AccessFine-Tuning Protein Language Models Unlocks the Potential of Underrepresented Viral Proteomes
Protein language models (pLMs) have revolutionized computational biology by generating rich protein vector representations—or embeddings—enabling major advancements in de novo protein design, structure prediction, variant effect analysis, and evolutionary studies. Despite these breakthroughs, current pLMs often exhibit biases against proteins from underrepresented species, with viral proteins being particularly affected—frequently referred to as the “dark matter” of the biological world due to their vast diversity and ubiquity, yet sparse representation in training datasets. Here, we show that fine-tuning pre-trained pLMs on viral protein sequences, using diverse learning frameworks and parameter-efficient strategies, significantly enhances representation quality and improves performance on downstream tasks. To support further research, we provide source code for fine-tuning pLMs and benchmarking embedding quality. By enabling more accurate modeling of viral proteins, our approach advances tools for understanding viral biology, combating emerging infectious diseases, and driving biotechnological innovation.
- Open Access
Space to Teach: Content-Rich Canvases for Visually-Intensive Education
With the decreasing cost of consumer display technologies making it easier for universities to have larger displays in classrooms, and the ubiquitous use of online tools such as collaborative whiteboards for remote learning during the COVID-19 pandemic, combining the two can be useful in higher education. This is especially true in visually intensive classes, such as data visualization courses, that can benefit from additional "space to teach," coined after the "space to think" sense-making idiom. In this paper, we reflect on our approach to using SAGE3, a collaborative whiteboard with advanced features, in higher education to teach visually intensive classes, provide examples of activities from our own visually-intensive courses, and present student feedback. We gather our observations into usage patterns for using content-rich canvases in education.
Open AccessSpace to Teach: Content-Rich Canvases for Visually-Intensive Education
With the decreasing cost of consumer display technologies making it easier for universities to have larger displays in classrooms, and the ubiquitous use of online tools such as collaborative whiteboards for remote learning during the COVID-19 pandemic, combining the two can be useful in higher education. This is especially true in visually intensive classes, such as data visualization courses, that can benefit from additional "space to teach," coined after the "space to think" sense-making idiom. In this paper, we reflect on our approach to using SAGE3, a collaborative whiteboard with advanced features, in higher education to teach visually intensive classes, provide examples of activities from our own visually-intensive courses, and present student feedback. We gather our observations into usage patterns for using content-rich canvases in education.
- Open Access
- 25 April 2024
- Published by Springer Nature in BMC Bioinformatics
- Vol. 25 (1) , 1-19
Improvements in viral gene annotation using large language models and soft alignments
Background: The annotation of protein sequences in public databases has long posed a challenge in molecular biology. This issue is particularly acute for viral proteins, which demonstrate limited homology to known proteins when using alignment, k-mer, or profile-based homology search approaches. A novel methodology employing Large Language Models (LLMs) addresses this methodological challenge by annotating protein sequences based on embeddings. Results: Central to our contribution is the soft alignment algorithm, drawing from traditional protein alignment but leveraging embedding similarity at the amino acid level to bypass the need for conventional scoring matrices. This method not only surpasses pooled embedding-based models in efficiency but also in interpretability, enabling users to easily trace homologous amino acids and delve deeper into the alignments. Far from being a black box, our approach provides transparent, BLAST-like alignment visualizations, combining traditional biological research with AI advancements to elevate protein annotation through embedding-based analysis while ensuring interpretability. Tests using the Virus Orthologous Groups and ViralZone protein databases indicated that the novel soft alignment approach recognized and annotated sequences that both blastp and pooling-based methods, which are commonly used for sequence annotation, failed to detect. Conclusion: The embeddings approach shows the great potential of LLMs for enhancing protein sequence annotation, especially in viral genomics. These findings present a promising avenue for more efficient and accurate protein function inference in molecular biology.
- 25 April 2024
- Published by Springer Nature in BMC Bioinformatics
- Vol. 25 (1) , 1-19
Open AccessImprovements in viral gene annotation using large language models and soft alignments
Background: The annotation of protein sequences in public databases has long posed a challenge in molecular biology. This issue is particularly acute for viral proteins, which demonstrate limited homology to known proteins when using alignment, k-mer, or profile-based homology search approaches. A novel methodology employing Large Language Models (LLMs) addresses this methodological challenge by annotating protein sequences based on embeddings. Results: Central to our contribution is the soft alignment algorithm, drawing from traditional protein alignment but leveraging embedding similarity at the amino acid level to bypass the need for conventional scoring matrices. This method not only surpasses pooled embedding-based models in efficiency but also in interpretability, enabling users to easily trace homologous amino acids and delve deeper into the alignments. Far from being a black box, our approach provides transparent, BLAST-like alignment visualizations, combining traditional biological research with AI advancements to elevate protein annotation through embedding-based analysis while ensuring interpretability. Tests using the Virus Orthologous Groups and ViralZone protein databases indicated that the novel soft alignment approach recognized and annotated sequences that both blastp and pooling-based methods, which are commonly used for sequence annotation, failed to detect. Conclusion: The embeddings approach shows the great potential of LLMs for enhancing protein sequence annotation, especially in viral genomics. These findings present a promising avenue for more efficient and accurate protein function inference in molecular biology.
- 25 August 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in Computing in Science & Engineering
- Vol. 25 (2) , 50-56
Reflecting on the Scalable Adaptive Graphics Environment Team’s 20-Year Translational Research Endeavor in Digital Collaboration Tools
Translational software research bridges the gap between scientific innovations and practical applications, driving impactful societal advancements. However, developing such software is challenging due to interdisciplinary collaboration, technology adoption, and postfunding sustainability. This article presents the experiences and insights of the Scalable Adaptive Graphics Environment (SAGE) team, which has spent two decades developing translational, cross-disciplinary, collaboration tools to benefit computational science research. With a focus on SAGE and its next-generation iterations, we explore the inherent challenges in translational research, such as fostering cross-disciplinary collaboration, motivating technology adoption, and ensuring postfunding product sustainability. We also discuss the roles of funding agencies, policymakers, and academic institutions in promoting translational research. Although the journey is fraught with challenges, the societal impact and satisfaction derived from translational research underscore its significance in the broader scientific landscape. This article aims to encourage further conversation and the development of effective models for translational software projects.
- 25 August 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in Computing in Science & Engineering
- Vol. 25 (2) , 50-56
Reflecting on the Scalable Adaptive Graphics Environment Team’s 20-Year Translational Research Endeavor in Digital Collaboration Tools
Translational software research bridges the gap between scientific innovations and practical applications, driving impactful societal advancements. However, developing such software is challenging due to interdisciplinary collaboration, technology adoption, and postfunding sustainability. This article presents the experiences and insights of the Scalable Adaptive Graphics Environment (SAGE) team, which has spent two decades developing translational, cross-disciplinary, collaboration tools to benefit computational science research. With a focus on SAGE and its next-generation iterations, we explore the inherent challenges in translational research, such as fostering cross-disciplinary collaboration, motivating technology adoption, and ensuring postfunding product sustainability. We also discuss the roles of funding agencies, policymakers, and academic institutions in promoting translational research. Although the journey is fraught with challenges, the societal impact and satisfaction derived from translational research underscore its significance in the broader scientific landscape. This article aims to encourage further conversation and the development of effective models for translational software projects.
- Open Access
- 5 April 2021
- Published by Oxford University Press (OUP) in Bioinformatics
- Vol. 37 (18) , 2803-2810
CoCoNet: an efficient deep learning tool for viral metagenome binning
Metagenomic approaches hold the potential to characterize microbial communities and unravel the intricate link between the microbiome and biological processes. Assembly is one of the most critical steps in metagenomics experiments. It consists of transforming overlapping DNA sequencing reads into sufficiently accurate representations of the community’s genomes. This process is computationally difficult and commonly results in genomes fragmented across many contigs. Computational binning methods are used to mitigate fragmentation by partitioning contigs based on their sequence composition, abundance or chromosome organization into bins representing the community’s genomes. Existing binning methods have been principally tuned for bacterial genomes and do not perform favorably on viral metagenomes. We propose Composition and Coverage Network (CoCoNet), a new binning method for viral metagenomes that leverages the flexibility and the effectiveness of deep learning to model the co-occurrence of contigs belonging to the same viral genome and provide a rigorous framework for binning viral contigs. Our results show that CoCoNet substantially outperforms existing binning methods on viral datasets. CoCoNet was implemented in Python and is available for download on PyPi (https://pypi.org/). The source code is hosted on GitHub at https://github.com/Puumanamana/CoCoNet and the documentation is available at https://coconet.readthedocs.io/en/latest/index.html. CoCoNet does not require extensive resources to run. For example, binning 100k contigs took about 4 h on 10 Intel CPU Cores (2.4 GHz), with a memory peak at 27 GB (see Supplementary Fig. S9 ). To process a large dataset, CoCoNet may need to be run on a high RAM capacity server. Such servers are typically available in high-performance or cloud computing settings. Supplementary data are available at Bioinformatics online.
- 5 April 2021
- Published by Oxford University Press (OUP) in Bioinformatics
- Vol. 37 (18) , 2803-2810
Open AccessCoCoNet: an efficient deep learning tool for viral metagenome binning
Metagenomic approaches hold the potential to characterize microbial communities and unravel the intricate link between the microbiome and biological processes. Assembly is one of the most critical steps in metagenomics experiments. It consists of transforming overlapping DNA sequencing reads into sufficiently accurate representations of the community’s genomes. This process is computationally difficult and commonly results in genomes fragmented across many contigs. Computational binning methods are used to mitigate fragmentation by partitioning contigs based on their sequence composition, abundance or chromosome organization into bins representing the community’s genomes. Existing binning methods have been principally tuned for bacterial genomes and do not perform favorably on viral metagenomes. We propose Composition and Coverage Network (CoCoNet), a new binning method for viral metagenomes that leverages the flexibility and the effectiveness of deep learning to model the co-occurrence of contigs belonging to the same viral genome and provide a rigorous framework for binning viral contigs. Our results show that CoCoNet substantially outperforms existing binning methods on viral datasets. CoCoNet was implemented in Python and is available for download on PyPi (https://pypi.org/). The source code is hosted on GitHub at https://github.com/Puumanamana/CoCoNet and the documentation is available at https://coconet.readthedocs.io/en/latest/index.html. CoCoNet does not require extensive resources to run. For example, binning 100k contigs took about 4 h on 10 Intel CPU Cores (2.4 GHz), with a memory peak at 27 GB (see Supplementary Fig. S9 ). To process a large dataset, CoCoNet may need to be run on a high RAM capacity server. Such servers are typically available in high-performance or cloud computing settings. Supplementary data are available at Bioinformatics online.