Rohit Singh
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
Research Interests
Rohit Singh has a total of 123 co-authors. Here are the top co-authors by number of shared publications:
- Bonnie Berger
Massachusetts Institute of Technology
23 shared publications - Lenore Cowen
Tufts University
11 shared publications - Samuel Sledzieski
Center for Computational Biology, Flatiron Institute , New York, NY 10010
9 shared publications - Kapil Devkota
Duke University
7 shared publications Huan Liang
Department of Biostat. and Bioinfo., Duke University, Durham, NC, USA
5 shared publications- Purushothama Rao Tata
Department of Cell Biology, Duke University School of Medicine, Durham, NC 27710, USA
4 shared publications
Latest Publications
- Open Access
COMPASS: Component-Wise Inference of Shared and Gene-Specific Perturbation Response
Predicting how a genetic perturbation reshapes a cell’s transcriptome is a central goal of computational biology. Previous studies report that the mean response across training perturbations rivals specialized models on standard accuracy metrics, even though it cannot distinguish which perturbation occurred. Across 2,270 CRISPRi perturbations measured in each of six cell lines, we show that this apparent paradox reflects a conserved organization of perturbation responses. Perturbations span a continuum from responses strongly aligned with the mean to more targeted responses that depart from it. Crucially, a perturbation’s position along this continuum is conserved across cell lines (Kendall’s W = 0.59) and predictable from STRING protein-interaction embeddings ( R 2 = 0.35). We formalize this structure with COMPASS, an interpretable linear model that decomposes each response into shared and gene-specific components and estimates them separately. The shared-response component is modeled as a cell-line-wide response scaled by a perturbation-specific coefficient. This coefficient is strongly conserved across cell lines. The residual gene-specific component—which is moderately conserved across cell lines—recovers pathway-level programs. COMPASS outperforms scGPT, CPA, GEARS, GenePert, and S tate in both response accuracy (de-biased Pearson delta 0.34 vs. ≤ 0.32) and perturbation discrimination (cosine PDS gain 0.23 vs. ≤ 0.08). These results recast perturbation prediction across cellular contexts as component-wise inference, with each component estimated from the evidence best suited to it.
Open AccessCOMPASS: Component-Wise Inference of Shared and Gene-Specific Perturbation Response
Predicting how a genetic perturbation reshapes a cell’s transcriptome is a central goal of computational biology. Previous studies report that the mean response across training perturbations rivals specialized models on standard accuracy metrics, even though it cannot distinguish which perturbation occurred. Across 2,270 CRISPRi perturbations measured in each of six cell lines, we show that this apparent paradox reflects a conserved organization of perturbation responses. Perturbations span a continuum from responses strongly aligned with the mean to more targeted responses that depart from it. Crucially, a perturbation’s position along this continuum is conserved across cell lines (Kendall’s W = 0.59) and predictable from STRING protein-interaction embeddings ( R 2 = 0.35). We formalize this structure with COMPASS, an interpretable linear model that decomposes each response into shared and gene-specific components and estimates them separately. The shared-response component is modeled as a cell-line-wide response scaled by a perturbation-specific coefficient. This coefficient is strongly conserved across cell lines. The residual gene-specific component—which is moderately conserved across cell lines—recovers pathway-level programs. COMPASS outperforms scGPT, CPA, GEARS, GenePert, and S tate in both response accuracy (de-biased Pearson delta 0.34 vs. ≤ 0.32) and perturbation discrimination (cosine PDS gain 0.23 vs. ≤ 0.08). These results recast perturbation prediction across cellular contexts as component-wise inference, with each component estimated from the evidence best suited to it.
- Open Access
- 29 July 2026
- Published by Springer Nature in Nature
- Vol. 657 (8133) , 1064-1072
Miniaturizing and modifying natural proteins with Raygun
Proteins have evolved over billions of years through coordinated substitutions, insertions and deletions, yet computational protein design cannot fully replicate nature’s ability to engineer new proteins from existing templates. Protein language models1,2,3 generate informative per-residue representations, but harnessing them for large-scale, function-preserving sequence modifications has remained beyond reach. Here we introduce Raygun, a generative artificial intelligence framework that enables miniaturization, modification and augmentation of proteins, using a probabilistic encoding of protein sequences constructed from language model embeddings. Our key conceptual advance is to encode each protein not as a sequence of variable length in high-dimensional space, but as a probability distribution in fixed dimensions, making proteins of any length directly commensurable. Controlled by just two parameters governing substitutions and length changes, Raygun can shrink proteins by 10–25% (sometimes more than 50%), expand them beyond their natural size, and introduce extensive sequence diversity, all while preserving predicted structural integrity and functional sites. In cell-based validation, Raygun miniaturized fluorescent proteins (2 shorter than 96% of fluorescent proteins in FPbase) and TurboID, a synthetic biotin ligase that has been widely adopted for proteomics. It also expanded epidermal growth factor (EGF), generating variants with higher EGFR-binding affinity than the wild type. These results show that protein function can be faithfully captured in a length-agnostic representation, enabling the kind of coordinated, large-scale sequence modifications that characterize natural protein evolution.
- 29 July 2026
- Published by Springer Nature in Nature
- Vol. 657 (8133) , 1064-1072
Open AccessMiniaturizing and modifying natural proteins with Raygun
Proteins have evolved over billions of years through coordinated substitutions, insertions and deletions, yet computational protein design cannot fully replicate nature’s ability to engineer new proteins from existing templates. Protein language models1,2,3 generate informative per-residue representations, but harnessing them for large-scale, function-preserving sequence modifications has remained beyond reach. Here we introduce Raygun, a generative artificial intelligence framework that enables miniaturization, modification and augmentation of proteins, using a probabilistic encoding of protein sequences constructed from language model embeddings. Our key conceptual advance is to encode each protein not as a sequence of variable length in high-dimensional space, but as a probability distribution in fixed dimensions, making proteins of any length directly commensurable. Controlled by just two parameters governing substitutions and length changes, Raygun can shrink proteins by 10–25% (sometimes more than 50%), expand them beyond their natural size, and introduce extensive sequence diversity, all while preserving predicted structural integrity and functional sites. In cell-based validation, Raygun miniaturized fluorescent proteins (2 shorter than 96% of fluorescent proteins in FPbase) and TurboID, a synthetic biotin ligase that has been widely adopted for proteomics. It also expanded epidermal growth factor (EGF), generating variants with higher EGFR-binding affinity than the wild type. These results show that protein function can be faithfully captured in a length-agnostic representation, enabling the kind of coordinated, large-scale sequence modifications that characterize natural protein evolution.
- 17 July 2026
- Published by Elsevier in Cell Systems
- Vol. 17 (9) , 101675
Foundation model reveals the shared organization of transcription and topologically associating domains
- 17 July 2026
- Published by Elsevier in Cell Systems
- Vol. 17 (9) , 101675
Foundation model reveals the shared organization of transcription and topologically associating domains
- Open Access
TP53-mediated bidirectional lineage plasticity drives alveolar epithelial cell extrusion and tissue remodeling
Cell extrusion contributes to epithelial homeostasis, but its dysregulation can lead to tumorigenesis or degeneration. A fine balance in this process is therefore essential for tissue integrity. Yet the cell types and states vulnerable to extrusion, and the mechanisms that drive it, remain elusive. Here, using spatial maps of cell states in human idiopathic pulmonary fibrosis (IPF) we find that aberrant TP53 activation in alveolar epithelial cells drives cell extrusion. Genetic modulation of TP53 specifically in alveolar epithelial type 1 cells (AT1) was sufficient to induce plasticity and subsequent extrusion as demonstrated by lineage tracing and live imaging. Strikingly, single cell and bulk transcriptome profiling revealed aberrant TP53 drives AT1 cells to acquire a transitional state mirroring AT2-derived regeneration associated intermediate states. Critically, loss of AT1 derived transitional state triggers a compensatory AT2-derived regenerative response, establishing a bidirectional transitional state that activates myofibroblasts and remodels the alveolus. Together, our study implicates AT1 plasticity and their reversion as an unrecognized driver of epithelial cell loss and establishes bidirectional transitional state as a central mechanism underlying progression of fibrotic remodeling.
Open AccessTP53-mediated bidirectional lineage plasticity drives alveolar epithelial cell extrusion and tissue remodeling
Cell extrusion contributes to epithelial homeostasis, but its dysregulation can lead to tumorigenesis or degeneration. A fine balance in this process is therefore essential for tissue integrity. Yet the cell types and states vulnerable to extrusion, and the mechanisms that drive it, remain elusive. Here, using spatial maps of cell states in human idiopathic pulmonary fibrosis (IPF) we find that aberrant TP53 activation in alveolar epithelial cells drives cell extrusion. Genetic modulation of TP53 specifically in alveolar epithelial type 1 cells (AT1) was sufficient to induce plasticity and subsequent extrusion as demonstrated by lineage tracing and live imaging. Strikingly, single cell and bulk transcriptome profiling revealed aberrant TP53 drives AT1 cells to acquire a transitional state mirroring AT2-derived regeneration associated intermediate states. Critically, loss of AT1 derived transitional state triggers a compensatory AT2-derived regenerative response, establishing a bidirectional transitional state that activates myofibroblasts and remodels the alveolus. Together, our study implicates AT1 plasticity and their reversion as an unrecognized driver of epithelial cell loss and establishes bidirectional transitional state as a central mechanism underlying progression of fibrotic remodeling.
- Open Access
- 31 March 2026
- Published by Oxford University Press (OUP) in Bioinformatics
- Vol. 42 (5)
Local transcriptional covariation produces accurate estimates of cell phenotype
Summary: The utility of single-cell RNA sequencing (scRNA-seq) is premised on the notion that transcriptional state can faithfully reflect cell phenotype. However, scRNA-seq measurements are noisy and sparse, with individual transcript counts showing limited correlation with cell phenotype markers such as protein expression. To better characterize cell states from scRNA-seq data, researchers analyze gene programs—sets of covarying genes—rather than individual transcripts. We hypothesized that more accurate estimation of gene covariation, especially at a local (i.e. cell-state) rather than global (i.e. experimental) scale, could better capture cell phenotypes. However, the field lacks appropriate mathematical frameworks for analyzing gene covariation: coexpression is quantified as a symmetric positive-definite matrix, where even basic operations like arithmetic differences lack biological interpretability. Here, we introduce Sceodesic, which exploits the Riemannian manifold structure of gene coexpression matrices to quantify cell state-specific coexpression patterns using the log-Euclidean metric from differential geometry. Unlike principal components analysis and non-negative matrix factorization, which infer only global covariation, Sceodesic efficiently discovers local covariation patterns and organizes them into interpretable, linear gene programs. Sceodesic outperforms existing approaches in predicting protein expression levels, distinguishing transcriptional responses to gene perturbations, and identifying biologically meaningful programs in fetal development. By respecting the mathematical structure of gene coexpression, Sceodesic bridges the gap between biological variability and statistical analysis of scRNA-seq data, enabling more accurate characterization of cell phenotypes. Availability and implementation: https://singhlab.net/Sceodesic.
- 31 March 2026
- Published by Oxford University Press (OUP) in Bioinformatics
- Vol. 42 (5)
Open AccessLocal transcriptional covariation produces accurate estimates of cell phenotype
Summary: The utility of single-cell RNA sequencing (scRNA-seq) is premised on the notion that transcriptional state can faithfully reflect cell phenotype. However, scRNA-seq measurements are noisy and sparse, with individual transcript counts showing limited correlation with cell phenotype markers such as protein expression. To better characterize cell states from scRNA-seq data, researchers analyze gene programs—sets of covarying genes—rather than individual transcripts. We hypothesized that more accurate estimation of gene covariation, especially at a local (i.e. cell-state) rather than global (i.e. experimental) scale, could better capture cell phenotypes. However, the field lacks appropriate mathematical frameworks for analyzing gene covariation: coexpression is quantified as a symmetric positive-definite matrix, where even basic operations like arithmetic differences lack biological interpretability. Here, we introduce Sceodesic, which exploits the Riemannian manifold structure of gene coexpression matrices to quantify cell state-specific coexpression patterns using the log-Euclidean metric from differential geometry. Unlike principal components analysis and non-negative matrix factorization, which infer only global covariation, Sceodesic efficiently discovers local covariation patterns and organizes them into interpretable, linear gene programs. Sceodesic outperforms existing approaches in predicting protein expression levels, distinguishing transcriptional responses to gene perturbations, and identifying biologically meaningful programs in fetal development. By respecting the mathematical structure of gene coexpression, Sceodesic bridges the gap between biological variability and statistical analysis of scRNA-seq data, enabling more accurate characterization of cell phenotypes. Availability and implementation: https://singhlab.net/Sceodesic.
- Open Access
EvoPool: Evolution-Guided Pooling of Protein Language Model Embeddings
Protein language models (PLMs) encode amino acid sequences into residue-level embeddings that must be pooled into fixed-size representations for downstream protein-level prediction tasks. Although these embeddings implicitly reflect evolutionary constraints, existing pooling strategies operate on single sequences and do not explicitly leverage information from homologous sequences or multiple sequence alignments. We introduce EvoPool, a self-supervised pooling framework that integrates evolutionary information from homologs directly into aggregated PLM representations using optimal transport. Our method constructs a fixed-size evolutionary anchor from an arbitrary number of homologous sequences and uses sliced Wasserstein distances to derive query protein embeddings that are geometrically informed by homologous sequence embeddings. Experiments across multiple state-of-the-art PLM families on the ProteinGym benchmark show that EvoPool consistently outperforms standard pooling baselines for variant effect prediction, demonstrating that explicit evolutionary guidance substantially enhances the functional utility of PLM representations. Our implementation code is available at https://github.com/navid-naderi/EvoPool .
Open AccessEvoPool: Evolution-Guided Pooling of Protein Language Model Embeddings
Protein language models (PLMs) encode amino acid sequences into residue-level embeddings that must be pooled into fixed-size representations for downstream protein-level prediction tasks. Although these embeddings implicitly reflect evolutionary constraints, existing pooling strategies operate on single sequences and do not explicitly leverage information from homologous sequences or multiple sequence alignments. We introduce EvoPool, a self-supervised pooling framework that integrates evolutionary information from homologs directly into aggregated PLM representations using optimal transport. Our method constructs a fixed-size evolutionary anchor from an arbitrary number of homologous sequences and uses sliced Wasserstein distances to derive query protein embeddings that are geometrically informed by homologous sequence embeddings. Experiments across multiple state-of-the-art PLM families on the ProteinGym benchmark show that EvoPool consistently outperforms standard pooling baselines for variant effect prediction, demonstrating that explicit evolutionary guidance substantially enhances the functional utility of PLM representations. Our implementation code is available at https://github.com/navid-naderi/EvoPool .
- Open Access
DeltaNMF: A Two-Stage Neural NMF for Differential Gene Program Discovery
Non-negative matrix factorization (NMF) is a foundational dimensionality-reduction method in single-cell transcriptomics, valued for its interpretable gene programs. However, in case-control settings common in perturbation and disease research, standard NMF conflates quantitative shifts in program usage with qualitative emergence of novel programs, obscuring whether observed differences reflect altered activity of shared programs or genuinely new cellular states. Here, we present DeltaNMF, a neural network-based reformulation of NMF that enables flexible biological priors and structured multi-stage fitting. We first demonstrate how our framework incorporates foundation model-derived gene similarities through graph-Laplacian regularization, improving program coherence by 2.4-fold on protein interaction networks. Building on this flexibility, we introduce a two-stage architecture that explicitly separates baseline programs (learned from control cells) from case-specific programs, disambiguating altered usage from novel program emergence. Our GPU-accelerated implementation achieves 20× speedup over consensus NMF while maintaining comparable accuracy. On synthetic data with known ground truth, DeltaNMF correctly identifies which programs are new versus shifted. Applied to coronary artery disease (CAD) Perturb-seq data, DeltaNMF isolates disease programs including the cerebral cavernous malformations pathway identified by Schnitzler et al., while clearly distinguishing them from baseline endothelial programs with altered activity. Our neural network formulation of NMF opens new directions for incorporating diverse biological priors into interpretable single-cell analysis, providing a principled framework for differential program discovery in case-control studies.
Open AccessDeltaNMF: A Two-Stage Neural NMF for Differential Gene Program Discovery
Non-negative matrix factorization (NMF) is a foundational dimensionality-reduction method in single-cell transcriptomics, valued for its interpretable gene programs. However, in case-control settings common in perturbation and disease research, standard NMF conflates quantitative shifts in program usage with qualitative emergence of novel programs, obscuring whether observed differences reflect altered activity of shared programs or genuinely new cellular states. Here, we present DeltaNMF, a neural network-based reformulation of NMF that enables flexible biological priors and structured multi-stage fitting. We first demonstrate how our framework incorporates foundation model-derived gene similarities through graph-Laplacian regularization, improving program coherence by 2.4-fold on protein interaction networks. Building on this flexibility, we introduce a two-stage architecture that explicitly separates baseline programs (learned from control cells) from case-specific programs, disambiguating altered usage from novel program emergence. Our GPU-accelerated implementation achieves 20× speedup over consensus NMF while maintaining comparable accuracy. On synthetic data with known ground truth, DeltaNMF correctly identifies which programs are new versus shifted. Applied to coronary artery disease (CAD) Perturb-seq data, DeltaNMF isolates disease programs including the cerebral cavernous malformations pathway identified by Schnitzler et al., while clearly distinguishing them from baseline endothelial programs with altered activity. Our neural network formulation of NMF opens new directions for incorporating diverse biological priors into interpretable single-cell analysis, providing a principled framework for differential program discovery in case-control studies.
- Open Access
Learning a PRECISE language for small-molecule binding
Virtual screening of billion-scale compound libraries has become feasible through machine learning approaches. In particular, CoNCISE (RECOMB 2025) introduced drug quantization via codebooks, achieving highly scalable and accurate binary predictions. However, drug discovery requires understanding not just whether molecules bind, but where they bind and how to target specific sites. Here, we present PRECISE which leverages CoNCISE's quantized small-molecule representations while operating on the target's 3D structure as its input. The key innovation of PRECISE is reimagining drug-target interaction as compatibility between quantized drug embeddings and a latent representation of the target's surface mesh, enriched with electrostatic and geometric features. PRECISE designs a novel surface representation, interpreted through a geometric deep learning architecture, enabling it to identify binding sites more accurately than state-of-the-art methods (DiffDock-L, Chai, and Boltz-2) while the codebook ensures billion-scale screening capability. Our formulation unlocks zero-shot generalization to complex targets such as metalloproteins and multi-chain complexes. To enable efficient integration with downstream docking workflows, we introduce PRECISE-MCTS, which combines fast PRECISE-based screening with selective Vina docking through an iterative Monte Carlo Tree Search approach. By providing both mechanistic understanding and massive scalability, PRECISE delivers capabilities that were previously mutually exclusive in virtual screening.
Open AccessLearning a PRECISE language for small-molecule binding
Virtual screening of billion-scale compound libraries has become feasible through machine learning approaches. In particular, CoNCISE (RECOMB 2025) introduced drug quantization via codebooks, achieving highly scalable and accurate binary predictions. However, drug discovery requires understanding not just whether molecules bind, but where they bind and how to target specific sites. Here, we present PRECISE which leverages CoNCISE's quantized small-molecule representations while operating on the target's 3D structure as its input. The key innovation of PRECISE is reimagining drug-target interaction as compatibility between quantized drug embeddings and a latent representation of the target's surface mesh, enriched with electrostatic and geometric features. PRECISE designs a novel surface representation, interpreted through a geometric deep learning architecture, enabling it to identify binding sites more accurately than state-of-the-art methods (DiffDock-L, Chai, and Boltz-2) while the codebook ensures billion-scale screening capability. Our formulation unlocks zero-shot generalization to complex targets such as metalloproteins and multi-chain complexes. To enable efficient integration with downstream docking workflows, we introduce PRECISE-MCTS, which combines fast PRECISE-based screening with selective Vina docking through an iterative Monte Carlo Tree Search approach. By providing both mechanistic understanding and massive scalability, PRECISE delivers capabilities that were previously mutually exclusive in virtual screening.
- Open Access
USHER: Guiding Foundation Model Representations through Distribution Shifts
Foundation models pre-trained on certain biological data modalities exhibit systematic representational biases when encountering out-of-distribution (OOD) data from new assays. The embedding drift largely arises from instrumentation and protocol-related artifacts rather than true biological variation in cell states or tissue morphology. These drifts are distinct from conventional batch effects and cannot be remedied by retraining as sample sizes are often insufficient, and modifying existing embeddings breaks downstream tools that depend on stable representations. We introduce USHER, an adaptable framework to learn simple transforms that return OOD embeddings to a foundation model’s reference space. USHER enables embedding transformation via an expectation maximization-style procedure. Given a reference in-distribution sample, USHER first estimates a Fused Gromov-Wasserstein coupling that aligns unpaired OOD (source) and reference (target) embeddings by minimizing transport distance while preserving local structure. To make optimal transport couplings more useful for down-stream tasks, we introduce the concept of entropic filtering to retain only high-confidence correspondences. In the second step, USHER learns a low-complexity transformation that reliably restores the model’s representation space for OOD data. We demonstrate this learned transformation generalizes to other OOD data from similar experimental conditions. We applied USHER to correct platform-specific biases seen when running scGPT on Xenium transcript counts: USHER maps Xenium embeddings back to the native scRNA-seq representation space, improving cell type clustering and cross-platform integration. Histopathology foundation models trained on H&E images fail on MALDI metabolite-profiled tissue images due to data-acquisition artifacts. USHER corrects these, enabling cell-type classification and protein abundance imputation. USHER offers a generalizable framework to make biological foundation models portable across a rapidly-evolving experimental landscape.
Open AccessUSHER: Guiding Foundation Model Representations through Distribution Shifts
Foundation models pre-trained on certain biological data modalities exhibit systematic representational biases when encountering out-of-distribution (OOD) data from new assays. The embedding drift largely arises from instrumentation and protocol-related artifacts rather than true biological variation in cell states or tissue morphology. These drifts are distinct from conventional batch effects and cannot be remedied by retraining as sample sizes are often insufficient, and modifying existing embeddings breaks downstream tools that depend on stable representations. We introduce USHER, an adaptable framework to learn simple transforms that return OOD embeddings to a foundation model’s reference space. USHER enables embedding transformation via an expectation maximization-style procedure. Given a reference in-distribution sample, USHER first estimates a Fused Gromov-Wasserstein coupling that aligns unpaired OOD (source) and reference (target) embeddings by minimizing transport distance while preserving local structure. To make optimal transport couplings more useful for down-stream tasks, we introduce the concept of entropic filtering to retain only high-confidence correspondences. In the second step, USHER learns a low-complexity transformation that reliably restores the model’s representation space for OOD data. We demonstrate this learned transformation generalizes to other OOD data from similar experimental conditions. We applied USHER to correct platform-specific biases seen when running scGPT on Xenium transcript counts: USHER maps Xenium embeddings back to the native scRNA-seq representation space, improving cell type clustering and cross-platform integration. Histopathology foundation models trained on H&E images fail on MALDI metabolite-profiled tissue images due to data-acquisition artifacts. USHER corrects these, enabling cell-type classification and protein abundance imputation. USHER offers a generalizable framework to make biological foundation models portable across a rapidly-evolving experimental landscape.