Smita Krishnaswamy

GeoSinkhorn Flow converts static phenoscapes into dynamic ones by learning perturbation-conditioned transport trajectories that respect the phenoscape's geometry and generalize to unseen conditions.

This study shows human claustrum neurons encode uncertainty and prediction errors during aversive learning, with distinct temporal profiles compared to anterior cingulate cortex.

Cell type-specific logistic regression estimators trained on 427 ROSMAP prefrontal donors project Braak- and β-amyloid-like transcriptomic signatures onto the 108-region Siletti atlas, flagging brainstem, hypothalamic and rostral hippocampal regions.

This study identifies that epithelial stem cell fate, not DNA damage, determines chemoradiotherapy response in rectal cancer by analyzing patient-derived organoids and using SPARTA to validate findings.

This study identifies stromal prostaglandin E2 as a key regulator of cell-fate plasticity in colorectal cancer, showing it converts chemosensitive stem cells into chemorefractory, prometastatic cells.

This study shows that beta cells drive pancreatic ductal adenocarcinoma development through endocrine-exocrine signaling, with obesity enhancing this process via CCK+ beta cell adaptation.

ImmunoStruct improves immunogenicity prediction by integrating peptide-MHC sequence, structure, and biochemical data, showing better performance than existing methods.

CellSpliceNet predicts alternative splicing outcomes in C. elegans neurons by integrating multiple data modalities and achieving high predictive accuracy.

This study shows that leveraging the intrinsic manifold geometry of brain activity can accelerate learning of a noninvasive BCI by aligning brain activity within the manifold during task performance.

The paper develops a deep learning framework to classify first-episode psychosis using multimodal data from neuroimaging, EEG, and facial features.

This paper introduces GSPA, a method to map gene space at single-cell resolution, enabling gene characterization and analysis of gene expression patterns across biological tasks.

The paper introduces NeuroSC, a platform that integrates graph analytics with connectome data to explore neurodevelopmental changes in C. elegans.

The paper presents an atlas of alternative splicing in C. elegans neurons, identifies novel splicing events, and describes regulatory patterns and networks.

The paper investigates the developmental fate of zebrafish embryos lacking the reprogramming factors Nanog, Pou5f3, and Sox19b, finding that cells can acquire novel cell states or primordial germ cell-like states.

AAnet identifies archetypal cell states in single-cell data to reveal tumor heterogeneity and spatial organization.

This paper introduces GEFMAP, a geometric deep learning method to predict metabolic flux from scRNA-seq data, estimating reaction rates in metabolic networks.

The paper identifies two micropeptides encoded by lincRNAs that regulate chromatin accessibility in vertebrate-specific neural cells and show their role in brain development.

This paper introduces T-PHATE, a nonlinear manifold learning algorithm for time-series data like fMRI, showing improvements in visualization, classification, and segmentation.

Neurospectrum extracts interpretable spatiotemporal features from neural activity using a geometric and topological deep learning framework, outperforming traditional methods in simulations and fMRI data.

scMMGAN integrates multi-omic data to reveal spatial patterns in triple negative breast cancer, showing improved alignment and novel gene associations.

The paper introduces spARC, a method to recover spatial signaling networks in gliomas, revealing an Osteopontin-CD44 interaction enriched in grade IV astrocytomas.

A NeurIPS 2021 benchmarking competition evaluated over 2600 submissions from 280 competitors across three multimodal single-cell integration tasks using a purpose-built benchmarking dataset and cloud evaluation pipeline.

The paper examines the role of the human-specific enhancer HACNS1 in human evolution by analyzing its effects on gene expression in humanized mice.

The paper presents computational tools for analyzing SARS-CoV-2 genetic data, showing that haplotype networks and unsupervised learning can reveal lineage emergence and population dynamics.

The paper introduces FMGAN, a conditional GAN that predicts hard-to-obtain biological information from easy-to-obtain data, demonstrating applications in drug discovery and clinical outcomes.

The paper uses diffusion condensation to analyze the C. elegans nerve ring neuropil, revealing its stratified structure and developmental principles.

This paper introduces a method to quantify perturbation effects at single-cell resolution using manifold-based likelihood estimation and vertex frequency clustering, achieving higher accuracy than existing methods.

A topological machine learning approach applied to single-nucleus RNA sequencing identifies shared glial activation states in AMD, AD, and MS.

The paper introduces Multiscale PHATE, a method for exploring high-dimensional biomedical data, and applies it to identify immune cell signatures associated with severe COVID-19 outcomes.

This paper introduces PhEMD, a method to embed drug perturbation effects in single-cell experiments, revealing low-dimensional manifold structures and enabling efficient drug effect inference.