Mohammad Saleh Refahi
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Mohammad Saleh Refahi has a total of 23 co-authors. Here are the top co-authors by number of shared publications:
- Gail L. Rosen
Drexel University
9 shared publications - Bahrad Ali Sokhansanj
Drexel University
6 shared publications - James R. Brown
Drexel University, Philadelphia, PA, USA
6 shared publications - Hyunwoo Yoo
Drexel University
4 shared publications Gavin Hearne
2 shared publications- Gavin Hearne
Drexel University, Philadelphia, PA, USA
2 shared publications
Latest Publications
- 2 July 2026
- Published by Springer Nature
- p. 89-134
Evaluating Genomic Language Models: Experimental Validation, Multi-Omics Integration, and Ethical-Legal Challenges
- 2 July 2026
- Published by Springer Nature
- p. 89-134
Evaluating Genomic Language Models: Experimental Validation, Multi-Omics Integration, and Ethical-Legal Challenges
- Open Access
Nevermore: Target-Conditioned Protein–Ligand Representation Learning for Multi-Objective Lead Optimization with Database-Grounded Retrieval
Recently, there has been great interest in AI-based approaches for de novo design of novel drug candidates. However, the generation of useful lead drug candidate compounds requires more than predicting engagement with the desired protein target. Candidate molecules must also be anchored in the real world of medicinal chemistry for their synthesis and modification as well as satisfying multiple drug development-related criteria. Here, we present Nevermore, an AI target-conditioned, database-grounded workflow for prioritizing candidate ligands from large compound libraries. Nevermore uses a geometry-aware protein–ligand affinity oracle to score target-specific binding and perform sparse integer edits in count-based Morgan fingerprint space. Nevermore then retrieves the most structurally similar molecules from public chemical databases. This design enables multi-objective search over predicted affinity and absorption, distribution, metabolism, excretion, and toxicity (ADMET) proxies while keeping all candidates anchored to valid database compounds. We evaluated Nevermore’s performance across three biologically distinct targets: Menin, a protein-interaction target relevant to leukemia; SARS-CoV-2 Mpro, a viral cysteine protease relevant to antiviral discovery; and epidermal growth factor receptor (EGFR), a kinase-superfamily oncology target with extensive experimentally tested compounds. Nevermore retrieved candidate sets with favorable predicted affinity–property trade-offs. These results support database-grounded fingerprint steering as a practical computational strategy for lead prioritization and for generating testable molecular hypotheses, although the prioritized candidates remain predictions, requiring follow-up experimental validation.
Open AccessNevermore: Target-Conditioned Protein–Ligand Representation Learning for Multi-Objective Lead Optimization with Database-Grounded Retrieval
Recently, there has been great interest in AI-based approaches for de novo design of novel drug candidates. However, the generation of useful lead drug candidate compounds requires more than predicting engagement with the desired protein target. Candidate molecules must also be anchored in the real world of medicinal chemistry for their synthesis and modification as well as satisfying multiple drug development-related criteria. Here, we present Nevermore, an AI target-conditioned, database-grounded workflow for prioritizing candidate ligands from large compound libraries. Nevermore uses a geometry-aware protein–ligand affinity oracle to score target-specific binding and perform sparse integer edits in count-based Morgan fingerprint space. Nevermore then retrieves the most structurally similar molecules from public chemical databases. This design enables multi-objective search over predicted affinity and absorption, distribution, metabolism, excretion, and toxicity (ADMET) proxies while keeping all candidates anchored to valid database compounds. We evaluated Nevermore’s performance across three biologically distinct targets: Menin, a protein-interaction target relevant to leukemia; SARS-CoV-2 Mpro, a viral cysteine protease relevant to antiviral discovery; and epidermal growth factor receptor (EGFR), a kinase-superfamily oncology target with extensive experimentally tested compounds. Nevermore retrieved candidate sets with favorable predicted affinity–property trade-offs. These results support database-grounded fingerprint steering as a practical computational strategy for lead prioritization and for generating testable molecular hypotheses, although the prioritized candidates remain predictions, requiring follow-up experimental validation.
- 12 June 2026
- Published by Drexel University Libraries
Enhancing nucleotide sequence representation for functional, taxonomy and phenotype prediction
Modern sequencing technologies generate vast amounts of nucleotide sequence data across genomics, metagenomics, and biotechnology. However, the central challenge has shifted from data generation to biological interpretation, particularly when sequences are fragmented, noisy, weakly labeled, or derived from organisms poorly represented in reference databases. Conventional pipelines based on alignment, k-mer matching, or task-specific feature engineering remain essential, but they often struggle under novelty, distribution shift, and incomplete sequence context. This dissertation addresses these challenges by developing foundation-model-based methods for learning transferable nucleotide sequence representations that support functional, taxonomic, and phenotype-related prediction. The dissertation introduces a progression of representation-learning systems for biological sequences. First, MetaBERTa is developed as a fragment-scale nucleotide foundation model pretrained on diverse prokaryotic genomes to support taxonomic classification and sequence-level representation learning. Second, Carmania extends nucleotide modeling to long genomic contexts using transition-regularized self-supervision, improving the ability to capture sequence organization beyond short fragments. Third, Scorpio introduces hierarchical contrastive adaptation to reshape embedding geometry across gene and taxonomic levels, enabling retrieval, classification, confidence estimation, and biological interpretation across full genes, short fragments, simulated reads, promoter sequences, and antimicrobial resistance tasks. Finally, Nevermore extends this framework toward multimodal biological modeling by integrating protein and ligand representations for target-conditioned, database-grounded molecular optimization. Together, these contributions advance nucleotide sequence modeling from isolated supervised pipelines toward reusable biological representation systems. The results demonstrate that learned embeddings capture meaningful biological structure, improve downstream prediction, support interpretation under uncertainty, and provide a foundation for future multimodal models connecting genomic sequences with proteins, molecules, and phenotype-related objectives. This work contributes to scalable and interpretable biological artificial intelligence systems for genomics, metagenomics, and computational discovery.
- 12 June 2026
- Published by Drexel University Libraries
Enhancing nucleotide sequence representation for functional, taxonomy and phenotype prediction
Modern sequencing technologies generate vast amounts of nucleotide sequence data across genomics, metagenomics, and biotechnology. However, the central challenge has shifted from data generation to biological interpretation, particularly when sequences are fragmented, noisy, weakly labeled, or derived from organisms poorly represented in reference databases. Conventional pipelines based on alignment, k-mer matching, or task-specific feature engineering remain essential, but they often struggle under novelty, distribution shift, and incomplete sequence context. This dissertation addresses these challenges by developing foundation-model-based methods for learning transferable nucleotide sequence representations that support functional, taxonomic, and phenotype-related prediction. The dissertation introduces a progression of representation-learning systems for biological sequences. First, MetaBERTa is developed as a fragment-scale nucleotide foundation model pretrained on diverse prokaryotic genomes to support taxonomic classification and sequence-level representation learning. Second, Carmania extends nucleotide modeling to long genomic contexts using transition-regularized self-supervision, improving the ability to capture sequence organization beyond short fragments. Third, Scorpio introduces hierarchical contrastive adaptation to reshape embedding geometry across gene and taxonomic levels, enabling retrieval, classification, confidence estimation, and biological interpretation across full genes, short fragments, simulated reads, promoter sequences, and antimicrobial resistance tasks. Finally, Nevermore extends this framework toward multimodal biological modeling by integrating protein and ligand representations for target-conditioned, database-grounded molecular optimization. Together, these contributions advance nucleotide sequence modeling from isolated supervised pipelines toward reusable biological representation systems. The results demonstrate that learned embeddings capture meaningful biological structure, improve downstream prediction, support interpretation under uncertainty, and provide a foundation for future multimodal models connecting genomic sequences with proteins, molecules, and phenotype-related objectives. This work contributes to scalable and interpretable biological artificial intelligence systems for genomics, metagenomics, and computational discovery.
- Open Access
Nevermore: Target-Conditioned Protein–Ligand Representation Learning for Multi-Objective Lead Optimization with Database-Grounded Retrieval
Target-conditioned molecular design requires optimizing binding affinity to a proposed therapeutic protein target while balancing competing developability constraints (e.g., absorption, distribution, metabolism, excretion, and toxicity; ADMET). Yet many computational pipelines either optimize a single objective or rely on fully de novo generation that can be difficult to control and interpret. We present Nevermore, a target-conditioned, database-grounded framework that combines a geometry-aware protein–ligand affinity oracle with Pareto-aware multi-objective search over an explicit molecular feature space. A central design choice is to optimize in count-based Morgan fingerprint space, where each feature corresponds to a chemically meaningful substructure count, enabling discrete, interpretable “bucket-level” edits. Nevermore learns target-conditioned scores by aligning protein and ligand representations under contrastive objectives and using a similarity-based prediction head; the resulting affinity oracle improves over previously reported benchmark baselines, providing a stronger scoring signal for downstream optimization. Nevermore then steers candidate selection by proposing sparse fingerprint edits, re-ranking candidates under multiple objectives, and projecting edited fingerprints back to valid molecules via nearest-neighbor retrieval from a large compound library. This yields efficient screening without exhaustive enumeration and provides transparent attributions that connect optimization steps to concrete chemical motifs. We evaluate Nevermore on two target case studies (Menin and SARS-CoV-2 Mpro). Across targets, the closed-loop search consistently retrieves candidate sets with improved affinity–property trade-offs compared with random sampling and similarity-only retrieval baselines, while maintaining explicit control and interpretability through discrete feature-space edits. These results support database-grounded, feature-space steering as a practical route to target-conditioned multi-objective lead refinement without relying on fully de novo generation.
Open AccessNevermore: Target-Conditioned Protein–Ligand Representation Learning for Multi-Objective Lead Optimization with Database-Grounded Retrieval
Target-conditioned molecular design requires optimizing binding affinity to a proposed therapeutic protein target while balancing competing developability constraints (e.g., absorption, distribution, metabolism, excretion, and toxicity; ADMET). Yet many computational pipelines either optimize a single objective or rely on fully de novo generation that can be difficult to control and interpret. We present Nevermore, a target-conditioned, database-grounded framework that combines a geometry-aware protein–ligand affinity oracle with Pareto-aware multi-objective search over an explicit molecular feature space. A central design choice is to optimize in count-based Morgan fingerprint space, where each feature corresponds to a chemically meaningful substructure count, enabling discrete, interpretable “bucket-level” edits. Nevermore learns target-conditioned scores by aligning protein and ligand representations under contrastive objectives and using a similarity-based prediction head; the resulting affinity oracle improves over previously reported benchmark baselines, providing a stronger scoring signal for downstream optimization. Nevermore then steers candidate selection by proposing sparse fingerprint edits, re-ranking candidates under multiple objectives, and projecting edited fingerprints back to valid molecules via nearest-neighbor retrieval from a large compound library. This yields efficient screening without exhaustive enumeration and provides transparent attributions that connect optimization steps to concrete chemical motifs. We evaluate Nevermore on two target case studies (Menin and SARS-CoV-2 Mpro). Across targets, the closed-loop search consistently retrieves candidate sets with improved affinity–property trade-offs compared with random sampling and similarity-only retrieval baselines, while maintaining explicit control and interpretability through discrete feature-space edits. These results support database-grounded, feature-space steering as a practical route to target-conditioned multi-objective lead refinement without relying on fully de novo generation.
- Open Access
- 11 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 1
Fast and Scalable Gene Embedding Search: A Comparative Study of FAISS and ScaNN
- 11 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 1
Open AccessFast and Scalable Gene Embedding Search: A Comparative Study of FAISS and ScaNN
- Open Access
Learning to Align Molecules and Proteins: A Geometry-Aware Approach to Binding Affinity
Accurate prediction of drug-target binding affinity can accelerate drug discovery by prioritizing promising compounds before costly wet-lab screening. While deep learning has advanced this task, most models fuse ligand and protein representations via simple concatenation and lack explicit geometric regularization, resulting in poor generalization across chemical space and time. We introduce FIRM-DTI, a lightweight framework that conditions molecular embeddings on protein embeddings through a feature-wise linear modulation (FiLM) layer and enforces metric structure with a triplet loss. An RBF regression head operating on embedding distances yields smooth, interpretable affinity predictions. Despite its modest size, FIRM-DTI achieves state-of-the-art performance on the Therapeutics Data Commons DTI-DG benchmark, as demonstrated by an extensive ablation study and out-of-domain evaluation. Our results underscore the value of conditioning and metric learning for robust drug-target affinity prediction.
Open AccessLearning to Align Molecules and Proteins: A Geometry-Aware Approach to Binding Affinity
Accurate prediction of drug-target binding affinity can accelerate drug discovery by prioritizing promising compounds before costly wet-lab screening. While deep learning has advanced this task, most models fuse ligand and protein representations via simple concatenation and lack explicit geometric regularization, resulting in poor generalization across chemical space and time. We introduce FIRM-DTI, a lightweight framework that conditions molecular embeddings on protein embeddings through a feature-wise linear modulation (FiLM) layer and enforces metric structure with a triplet loss. An RBF regression head operating on embedding distances yields smooth, interpretable affinity predictions. Despite its modest size, FIRM-DTI achieves state-of-the-art performance on the Therapeutics Data Commons DTI-DG benchmark, as demonstrated by an extensive ablation study and out-of-domain evaluation. Our results underscore the value of conditioning and metric learning for robust drug-target affinity prediction.
- Open Access
Fast and Scalable Gene Embedding Search: A Comparative Study of FAISS and ScaNN
The exponential growth of DNA sequencing data has outpaced traditional heuristic-based methods, which struggle to scale effectively. Efficient computational approaches are urgently needed to support large-scale similarity search, a foundational task in bioinformatics for detecting homology, functional similarity, and novelty among genomic and proteomic sequences. Although tools like BLAST have been widely used and remain effective in many scenarios, they suffer from limitations such as high computational cost and poor performance on divergent sequences. In this work, we explore embedding-based similarity search methods that learn latent representations capturing deeper structural and functional patterns beyond raw sequence alignment. We systematically evaluate two state-of-the-art vector search libraries, FAISS and ScaNN, on biologically meaningful gene embeddings. Unlike prior studies, our analysis focuses on bioinformatics-specific embeddings and benchmarks their utility for detecting novel sequences, including those from uncharacterized taxa or genes lacking known homologs. Our results highlight both computational advantages (in memory and runtime efficiency) and improved retrieval quality, offering a promising alternative to traditional alignment-heavy tools.
Open AccessFast and Scalable Gene Embedding Search: A Comparative Study of FAISS and ScaNN
The exponential growth of DNA sequencing data has outpaced traditional heuristic-based methods, which struggle to scale effectively. Efficient computational approaches are urgently needed to support large-scale similarity search, a foundational task in bioinformatics for detecting homology, functional similarity, and novelty among genomic and proteomic sequences. Although tools like BLAST have been widely used and remain effective in many scenarios, they suffer from limitations such as high computational cost and poor performance on divergent sequences. In this work, we explore embedding-based similarity search methods that learn latent representations capturing deeper structural and functional patterns beyond raw sequence alignment. We systematically evaluate two state-of-the-art vector search libraries, FAISS and ScaNN, on biologically meaningful gene embeddings. Unlike prior studies, our analysis focuses on bioinformatics-specific embeddings and benchmarks their utility for detecting novel sequences, including those from uncharacterized taxa or genes lacking known homologs. Our results highlight both computational advantages (in memory and runtime efficiency) and improved retrieval quality, offering a promising alternative to traditional alignment-heavy tools.
- Open Access
- 29 March 2025
- Published by Springer Nature in Communications Biology
- Vol. 8 (1) , 1-18
Enhancing nucleotide sequence representations in genomic analysis with contrastive optimization
Analysis of genomic and metagenomic sequences is inherently more challenging than that of amino acid sequences due to the higher divergence among evolutionarily related nucleotide sequences, variable k-mer and codon usage within and among genomes of diverse species, and poorly understood selective constraints. We introduce Scorpio (Sequence Contrastive Optimization for Representation and Predictive Inference on DNA), a versatile framework designed for nucleotide sequences that employ contrastive learning to improve embeddings. By leveraging pre-trained genomic language models and k-mer frequency embeddings, Scorpio demonstrates competitive performance in diverse applications, including taxonomic and gene classification, antimicrobial resistance (AMR) gene identification, and promoter detection. A key strength of Scorpio is its ability to generalize to novel DNA sequences and taxa, addressing a significant limitation of alignment-based methods. Scorpio has been tested on multiple datasets with DNA sequences of varying lengths (long and short) and shows robust inference capabilities. Additionally, we provide an analysis of the biological information underlying this representation, including correlations between codon adaptation index as a gene expression factor, sequence similarity, and taxonomy, as well as the functional and structural information of genes.
- 29 March 2025
- Published by Springer Nature in Communications Biology
- Vol. 8 (1) , 1-18
Open AccessEnhancing nucleotide sequence representations in genomic analysis with contrastive optimization
Analysis of genomic and metagenomic sequences is inherently more challenging than that of amino acid sequences due to the higher divergence among evolutionarily related nucleotide sequences, variable k-mer and codon usage within and among genomes of diverse species, and poorly understood selective constraints. We introduce Scorpio (Sequence Contrastive Optimization for Representation and Predictive Inference on DNA), a versatile framework designed for nucleotide sequences that employ contrastive learning to improve embeddings. By leveraging pre-trained genomic language models and k-mer frequency embeddings, Scorpio demonstrates competitive performance in diverse applications, including taxonomic and gene classification, antimicrobial resistance (AMR) gene identification, and promoter detection. A key strength of Scorpio is its ability to generalize to novel DNA sequences and taxa, addressing a significant limitation of alignment-based methods. Scorpio has been tested on multiple datasets with DNA sequences of varying lengths (long and short) and shows robust inference capabilities. Additionally, we provide an analysis of the biological information underlying this representation, including correlations between codon adaptation index as a gene expression factor, sequence similarity, and taxonomy, as well as the functional and structural information of genes.
- Open Access
- 22 November 2024
- Published by Association for Computing Machinery (ACM)
- p. 1
Normalized Compression Distance for DNA Classification
The increasingly common use of next-generation sequencing has enabled greater access to large-scale (meta-)genomic datasets than ever before. The resulting deluge of data has made the quest for efficient DNA sequence classification methods an urgent challenge for downstream analyses. Traditional sequence alignment-based methods for DNA sequence classification struggle when presented with increasingly large volumes of sequence data due to the computational complexity of alignment. Subsequently, there is a need for methods capable of sequence identification without alignment. Normalized compression distance (NCD) has demonstrated capabilities in the field of text classification as a low-resource alternative to deep neural networks by leveraging compression algorithms to approximate Kolmogorov information distance. In an effort to apply this technique toward genomics tasks akin to tools such as Many-against-Many sequence searching (MMseqs) and Kraken2, we have explored the use of a gzip-based NCD towards both gene labeling of ORFs (open reading frames) and taxonomic classification of short reads. This demonstrates the efficacy of NCD in diverse multitask classification, and we further explore the capacity for NCD to classify larger libraries of metagenomic reads.
- 22 November 2024
- Published by Association for Computing Machinery (ACM)
- p. 1
Open AccessNormalized Compression Distance for DNA Classification
The increasingly common use of next-generation sequencing has enabled greater access to large-scale (meta-)genomic datasets than ever before. The resulting deluge of data has made the quest for efficient DNA sequence classification methods an urgent challenge for downstream analyses. Traditional sequence alignment-based methods for DNA sequence classification struggle when presented with increasingly large volumes of sequence data due to the computational complexity of alignment. Subsequently, there is a need for methods capable of sequence identification without alignment. Normalized compression distance (NCD) has demonstrated capabilities in the field of text classification as a low-resource alternative to deep neural networks by leveraging compression algorithms to approximate Kolmogorov information distance. In an effort to apply this technique toward genomics tasks akin to tools such as Many-against-Many sequence searching (MMseqs) and Kraken2, we have explored the use of a gzip-based NCD towards both gene labeling of ORFs (open reading frames) and taxonomic classification of short reads. This demonstrates the efficacy of NCD in diverse multitask classification, and we further explore the capacity for NCD to classify larger libraries of metagenomic reads.