Adithya V. Madduri

Identifying small, interpretable gene sets that robustly capture disease-associated variation in singlecell transcriptomic data remains a central challenge for biological interpretation and experimental followup. In practice, commonly used differential expression and sparsity-based approaches often produce large, unstable gene lists that fail to generalize across patients due to strong donor-specific confounding. We study sparse gene selection for reconstructing donor-robust, disease-aligned cellular trajectories in real single-cell RNA-seq datasets. We introduce Sparse Linear Manifold Control (SLMC), a practical workflow that defines a disease-aligned score after removing donor-associated variation and selects minimal gene programs whose expression reconstructs this score. We focus on diagnosing the structure of the resulting reconstruction objective and evaluating selection strategies under realistic health data conditions. Across five human single-cell datasets spanning oncology and neurodegeneration, we find that the reconstruction objective exhibits strong diminishing returns, explaining why simple greedy selection methods perform well in practice. Under strict donor-heldout evaluation, greedy methods consistently outperform LASSO at small gene budgets and achieve accurate reconstruction with as few as 25 genes. Together, these results highlight how careful objective design and empirical evaluation enable robust and interpretable gene selection for disease-aligned representation learning in single-cell health data. Data and Code Availability: All code used for data preprocessing, model training, evaluation, and feature importance analyses is available at: https://github.com/AdiVM/SLMC_single-cell . The singlecell data used in this study are publicly available human transcriptomic datasets generated by prior studies and accessible through the Gene Expression Omnibus (GEO). Analyses were performed using renal cell carcinoma single-cell RNA-seq data (GEO accession: GSE314072 ) and Alzheimer’s disease single-nucleus RNA-seq data from human cortex (GEO accession: GSE138852 ), with additional publicly available datasets used for cross-context diagnostic evaluation (GEO accessions: GSM8652069 , GSE308624 , and GSE227734 ). All datasets contain de-identified human samples and are available under standard publicuse terms via GEO. Institutional Review Board (IRB): This study analyzes de-identified, publicly available human transcriptomic data obtained from previously published studies. No new data were collected, and no identifiable private information was accessed. In accordance with institutional policy, this work was determined to constitute non-human subjects research and did not require additional IRB approval.

Forelimb reaching tasks are widely used in neuroscience across species. In mice, pellet reaching is especially useful because it requires fine motor control. However, few automated pellet-delivery systems are well suited for head-restrained reaching, where pellet position must be highly consistent. Here we present a complete pipeline for head-restrained, or head-fixed, forelimb reaching in mice. The pipeline includes a simple customizable pellet-presentation rig, an effective training protocol, and analysis code to detect reaches and classify their outcomes. To demonstrate pipeline utility, we quantified interruption of an ongoing reach. The temporal requirement for initiating cancellation was dependent on the phase of the reach, lengthening significantly as the reach neared completion. These results show that the pipeline can resolve fine phase-dependent features of skilled behavior and provide a practical framework for studying motor control in head-restrained mice.

Plasma proteomics captures a functional snapshot of human physiology; yet, most machine learning models treat protein abundances as independent variables, ignoring the fact that biological systems and proteomic measurements are inherently compositional. Many molecular processes depend not on absolute concentrations but on relative balances: receptor–ligand stoichiometry, enzyme–substrate ratios, and homeostatic feedbacks that govern signaling and metabolism. We propose that these relationships are best captured through pairwise protein ratios, which more faithfully reflect underlying biochemical constraints than raw expression values. We evaluate a machine learning framework that models pairwise log-ratios of proteins ( log ( A ) − log ( B )) as features, thereby encoding compositional structure directly into the learning space. Applied to the ROSMAP plasma proteomics cohort (n = 871), this approach substantially improved the classification of Alzheimer’s subtypes (NCI, MCI, AD, AD+) with an average AUROC gain of +0.1274 over a strong baseline that incorporated raw proteomics and demographics. The top-ranked ratios (e.g., SEMA3C:TMEM70, IDUA:NPTXR) captured converging pathogenic pillars of Alzheimer’s disease, including microglial activation, proteostasis dysregulation, and lipid-clearance imbalance, highlighting that ratio-based features recover biologically coherent axes of disease. To assess generality, we scaled the method to the UK Biobank proteomic dataset (n > 53,000; 587 phenotypes). The ratio-based model outperformed raw-level models in 95.1 % of diseases, with statistically significant (FDR < 0.05) gains in 56.7 %. Together, these results suggest that proteomic data should be viewed and modeled as compositional systems, where relative protein abundances carry the accurate functional signal. This insight supports the broader utility of ratio-based representations for disease prediction and biomarker discovery.