LD Block-Based Genotype Embeddings for Asthma Biology and UK Biobank Extension (Ongoing)

Built a two-stage LD-block representation learning framework for asthma genotype data, combining block-wise VAEs with transformer-based contextualization to learn biologically meaningful subject embeddings.

Genomics Representation Learning VAE Transformer LD Blocks Asthma UK Biobank PyTorch
LD block-based genotype embedding framework showing block VAE encoding, transformer contextualization, emergent genomic structure, and biological interpretation in asthma

Two-stage LD block embedding framework showing independent block encoding, transformer contextualization, emergent genomic structure, and biological interpretation including HLA class II dominance and PDE4D emergence.

Key insights

  • Recovered known structure: HLA class II emerged as the dominant unsupervised axis, showing that the model recovers meaningful asthma biology without using phenotype labels.
  • Found secondary biology: After masking HLA effects, PDE4D appeared as an independent signal, suggesting the framework can reveal additional disease-relevant pathways beyond the strongest locus.
  • Transformer adds value: Cross-block contextualization preserved local LD structure while reorganizing subjects into clearer biological patterns than block-level embeddings alone.
  • Scaling next: The same framework is now being extended to UK Biobank cohort.

GitHub

Single-Cell Immune State Modeling and Multi-Modal Integration

Built single-cell analysis workflows spanning RNA-only PBMC benchmarking and ongoing paired RNA+ATAC integration. Combined clustering, cell-state scoring, cell–cell communication, chromatin accessibility, and transcription-factor inference to characterize immune-cell states across stimulated and control conditions.

Single Cell scRNA-seq scATAC-seq Multiome Scanpy Seurat WNN Cell-State Scoring Cell-Cell Communication TF Inference
UMAP projection of PBMC single-cell RNA-seq data showing annotated immune cell populations

UMAP of PBMC single-cell RNA-seq data after clustering and annotation, highlighting major immune-cell populations used for downstream cell-state scoring, differential expression, and multi-modal integration analyses.

Highlights

  • RNA-only benchmarking: Compared multiple cell-state scoring strategies across PBMC3k and interferon-stimulated Kang datasets using clustering, marker genes, and immune program scores.
  • Cross-modal integration: Integrated paired RNA+ATAC data with weighted nearest neighbors (WNN) to evaluate how chromatin accessibility refines immune-cell separation.
  • Regulatory modeling: Inferred transcription-factor activity from RNA and validated it against ATAC motif accessibility, with ongoing work on pseudotime and dynamic regulatory programs.

RNA Benchmark Repository · Multiome / RNA+ATAC Repository

Proteomics Analysis of Long COVID Heterogeneity

Built a proteomics analysis workflow to study neurocognitive Long COVID using Olink NPX data, combining covariate-adjusted modeling in a pediatric cohort with scalable phenotype–proteomics integration in ~50,000 UK Biobank participants.

Proteomics Olink NPX Long COVID UK Biobank PySpark Spark SQL Logistic Regression

Highlights

  • Subtype-focused analysis: Distinguished neurocognitive and non-neurocognitive Long COVID using protein-level signatures.
  • Replication at scale: Extended the workflow to ~50K UK Biobank participants with distributed Spark-based integration.
  • Biological follow-up: Candidate proteins and pathways point to immune and neuroinflammatory mechanisms underlying Long COVID heterogeneity.

GitHub

Endophenotype-driven GWAS Framework for Asthma Heterogeneity

Developed a subtype-aware GWAS framework for asthma by combining PCA-derived clinical endophenotypes with categorical association modeling, enabling genetic analysis across biologically coherent patient subgroups rather than treating asthma as a single uniform phenotype.

PCA GWAS ANCOVA Statistical Genetics Endophenotypes Asthma ICS Response R
Endophenotype-driven GWAS framework illustrating PCA-based stratification and subtype-specific association modeling

PCA-based clinical stratification into endophenotypes integrated with subtype-aware GWAS to identify asthma-associated loci across disease severity groups.

Highlights

  • Subtype-aware association testing: Endophenotype-based modeling improved detection of genetic signals compared with standard case-control GWAS.
  • Replication across cohorts: Multiple loci, including DGKI and MIR99AHG, replicated in independent asthma datasets.
  • Clinical relevance: Endophenotypes aligned with severity, IgE, lung function, and differential ICS treatment response.

Paper 1 · Paper 2 · Paper 3 · GitHub

miRNA Modifiers of Inhaled Corticosteroid Response in Asthma

Identified miR-584-5p as a potential biomarker of inhaled corticosteroid resistance in asthma using interaction modeling across two pediatric cohorts.

miRNA Pharmacogenomics Asthma ICS Response RNA-seq

Key finding: Higher miR-584-5p expression was associated with increased exacerbation risk specifically in ICS-treated patients.
Paper · GitHub

Machine Learning Analysis of Pediatric COVID-19 Clinical Data

Built a Random Forest model integrating chest X-ray impressions, symptoms, and demographics to predict pediatric COVID-19 infection status.

Random Forest Clinical NLP SHAP Chest X-ray EHR

Result: Achieved AUC 0.85 and F1-score 0.79, with radiological findings emerging as the strongest predictors.
Paper · GitHub

Structured Modeling of Context-Aware Systems

PhD research on modeling event-driven mobile applications using probabilistic sequence models, neural representation learning, and combinatorial optimization.

CRF Neural Networks Sequence Modeling Optimization PhD Dissertation

Established the foundations for my later work in biological sequence modeling, latent representations, and scalable search in high-dimensional data.

Publications: Paper 1 · Paper 2 · Paper 3 · GitHub