The Plant Cell Atlas Computation, Modeling, & AI Two-Part Speaker Series
|
|
|
This series highlights recent advances in AI-driven tools for single-cell datasets to bring awareness of new methodologies and will feature Dong Xu, from the University of South Florida on May 4, 2026 and Romain Lopez, from New York University on May 11, 2026.
Register for this series here!!
Register for this series here!!
"Deep-Learning Methods for Single-cell Data Analyses"
Dong Xu, Health Informatics Institute, Morsani College of Medicine, University of South Florida
May 4, 2026, 12:00pm ET / 11:00am CT / 9:00am PT
Abstract: Single-cell data analysis plays a crucial role in uncovering the heterogeneity and dynamics of tissues, organisms, and complex diseases. Deep learning and statistical methods offer powerful tools for identifying intricate biological patterns within large-scale, noisy datasets. We developed scGNN, a hypothesis-free graph neural network framework for single-cell RNA-Seq analysis, which integrates three iterative multi-modal autoencoders and outperforms existing tools in gene imputation and cell clustering. We also introduced TrimNN, a neural network-based approach for detecting network motifs within triangulated cell graphs from spatial omics data. We applied prompt-based learning on the large single-cell RNA-seq models for data analysis. These methods and tools provide critical insights into the mechanisms underlying the complex tissue heterogeneities in various species.
Dong Xu, Health Informatics Institute, Morsani College of Medicine, University of South Florida
May 4, 2026, 12:00pm ET / 11:00am CT / 9:00am PT
Abstract: Single-cell data analysis plays a crucial role in uncovering the heterogeneity and dynamics of tissues, organisms, and complex diseases. Deep learning and statistical methods offer powerful tools for identifying intricate biological patterns within large-scale, noisy datasets. We developed scGNN, a hypothesis-free graph neural network framework for single-cell RNA-Seq analysis, which integrates three iterative multi-modal autoencoders and outperforms existing tools in gene imputation and cell clustering. We also introduced TrimNN, a neural network-based approach for detecting network motifs within triangulated cell graphs from spatial omics data. We applied prompt-based learning on the large single-cell RNA-seq models for data analysis. These methods and tools provide critical insights into the mechanisms underlying the complex tissue heterogeneities in various species.
"Multimodal pooled genetic screens integrating transcriptomics and image-based phenotypes"
Romain Lopez, Assistant Professor of Computer Science and Biology at New York University.
May 11, 2026, 12:00pm ET / 11:00am CT / 9:00am PT
Abstract: Understanding how genes regulate cellular function benefits from various experimental approaches, with perturbation-based methods offering particularly powerful insights. We developed PerturbPAIR, integrating Perturb-seq and Optical Pooled Screens (OPS) to combine transcriptomic and imaging-based phenotypes at scale. We profiled LPS-stimulated bone marrow-derived macrophages, performing ~1,000 gene perturbations in Perturb-seq and ~3,000 in OPS. Perturb-seq identified co-functional modules encompassing interferon signaling, cytoskeletal regulation, and metabolic adaptation. OPS demonstrated superior sensitivity and revealed regulatory clusters based on protein localization patterns for inflammatory markers (Nos2, Hif1a, p65). Both modalities showed high concordance, with similar perturbations clustering together across transcriptomic and imaging spaces. Some perturbations with distinct transcriptomes converged to similar optical phenotypes, highlighting complementary information captured by each modality.
We therefore proposed a new deep generative model of transcriptional perturbation outcomes that effectively employs the imaging data for imputing the effect of unseen perturbations. We demonstrate that our novel method strongly outperforms existing methods for this task. Additionally, the model can be used for denoising Perturb-seq data when sample sizes are insufficient, a novel and important use-case in the field. Our integrated approach identified regulatory networks spanning cellular stress responses, metabolic control, and immune signaling pathways. We successfully imputed gene expression signatures for ~600 perturbations using only OPS data, enabling direct mapping onto these networks. This multimodal framework enhances our ability to dissect complex biological mechanisms and develop predictive models bridging molecular states with phenotypic outcomes in immune cell activation.
Romain Lopez, Assistant Professor of Computer Science and Biology at New York University.
May 11, 2026, 12:00pm ET / 11:00am CT / 9:00am PT
Abstract: Understanding how genes regulate cellular function benefits from various experimental approaches, with perturbation-based methods offering particularly powerful insights. We developed PerturbPAIR, integrating Perturb-seq and Optical Pooled Screens (OPS) to combine transcriptomic and imaging-based phenotypes at scale. We profiled LPS-stimulated bone marrow-derived macrophages, performing ~1,000 gene perturbations in Perturb-seq and ~3,000 in OPS. Perturb-seq identified co-functional modules encompassing interferon signaling, cytoskeletal regulation, and metabolic adaptation. OPS demonstrated superior sensitivity and revealed regulatory clusters based on protein localization patterns for inflammatory markers (Nos2, Hif1a, p65). Both modalities showed high concordance, with similar perturbations clustering together across transcriptomic and imaging spaces. Some perturbations with distinct transcriptomes converged to similar optical phenotypes, highlighting complementary information captured by each modality.
We therefore proposed a new deep generative model of transcriptional perturbation outcomes that effectively employs the imaging data for imputing the effect of unseen perturbations. We demonstrate that our novel method strongly outperforms existing methods for this task. Additionally, the model can be used for denoising Perturb-seq data when sample sizes are insufficient, a novel and important use-case in the field. Our integrated approach identified regulatory networks spanning cellular stress responses, metabolic control, and immune signaling pathways. We successfully imputed gene expression signatures for ~600 perturbations using only OPS data, enabling direct mapping onto these networks. This multimodal framework enhances our ability to dissect complex biological mechanisms and develop predictive models bridging molecular states with phenotypic outcomes in immune cell activation.
