scaLR — Low-Resource Deep Learning Platform for Single-Cell Analysis & Biomarker Discovery
Single-cell RNA sequencing (scRNA-seq) produces vast amounts of individual-cell profiling data, and accurately annotating cell types and identifying associated biomarkers from this data is a significant computational challenge. Deep neural network pipelines extract meaningful signal from that noisy, high-dimensional data, but existing tools typically require substantial compute to process large-scale datasets — putting rigorous single-cell analysis out of reach for teams without access to high-memory infrastructure.
Platform & feature-extraction approach
We led the development of scaLR, an end-to-end deep learning platform that processes data into feature subsets and batches samples to significantly reduce the memory footprint required for large-dataset analysis, while running efficiently across multiple CPUs. The platform covers the full analysis workflow — data processing, feature extraction, model training, evaluation, and downstream analysis — and includes a novel feature-extraction algorithm that trains on feature subsets, scores feature importance within each subset, and then trains a final model on the aggregated top-K most important features. I guided the project's overall direction, contributing to batch optimization strategy, design of experiments, and technical review throughout development.
Outcomes & impact
scaLR achieves prediction accuracy comparable to existing Python-based single-cell analysis pipelines while requiring significantly less training time and computational resources — making rigorous biomarker discovery accessible on standard hardware rather than requiring specialized infrastructure. The batch-wise processing approach was central to that efficiency gain, letting the pipeline scale to large sample sizes without a proportional increase in memory requirements.
Downstream analyses — differential gene expression, gene recall curves, and ROC-AUC evaluation — surface significant, statistically grounded biomarkers for different cell types and disease states. The approach was validated across real-world use cases including cell-type classification and clinical condition-specific biomarker discovery (demonstrated on COVID-19 vs. normal patient data). scaLR was published in Briefings in Bioinformatics (2025) and released as open-source software.