Koushik Chowdhury

Profile

Computer science graduate specializing in machine learning, deep learning and graph-based modeling applied to high-dimensional biological data. Experienced in single-cell transcriptomics, immune phenotype modeling, and robust predictive system development.

Research Experience

• Designed and published bioconverge, a Python library (MIT license, pip-installable, 20-test pytest suite) that takes per-patient multi-omics scores and runs them through four analytical layers: concordance analysis, gradient-based fragility scoring, automated hypothesis generation via Reactome, Enrichr, STRING, and GWAS Catalog APIs, and multi-source confidence tier assignment; applied to 116 TNBC patients, producing 6 Tier A, 9 Tier B, and 3 exploratory findings
• Designed adversarial gradient probes (FGSM, PGD) as synthetic biological measurement tools for TNBC characterisation; validated signal reproducibility across two independent neural architectures (Spearman r = 0.96) and two independent hospital cohorts; recovered 11 known TNBC-associated genes in the top 20 most sensitive features without biological supervision
• Built and evaluated a multimodal fusion model combining histopathology image sensitivity and gene expression features under leave-one-out cross-validation (AUC 0.9693); implemented shuffled-label permutation controls (n = 20) as a rigorous signal-to-noise evaluation framework for small-sample medical datasets
• Conducted scRNA-seq atlas construction and T cell immune phenotyping on 100,064 cells from 26 TNBC patients (GSE176078); implemented Leiden clustering with quantitative annotation validation (NMI 0.616–0.671) and per-cell PDCD1/CD2 ratio computation as a quantitative parameter for synthetic ligand surface design
• Translated single-cell immune phenotypes into patient-stratified therapeutic design recommendations via a bootstrap-stable DesignPriorityScore (Spearman ρ > 0.85 across 200 resamples; top-quartile retention > 90%)
• Performed longitudinal survival modelling (Kaplan–Meier, Cox proportional hazards regression) across multiple TCGA cohorts; identified adversarial sensitivity score as an independent prognostic signal for progression-free interval (log-rank p = 0.0072, HR = 0.54)
• Applied ComBat batch correction for cross-platform harmonisation of TCGA and GEO transcriptomic datasets; validated classifiers externally on METABRIC (Illumina microarray) demonstrating cross-platform generalisation
• Built deep learning architectures (CNN, BiLSTM, Graph Neural Networks) for biological sequence modelling; designed end-to-end preprocessing pipelines including normalisation, batch correction, imputation, augmentation, and cross-validation
• Published all analysis pipelines as open-source code on GitHub, archived processed datasets on Harvard Dataverse, and released one project as a standalone pip package; all work is independently reproducible

Selected Publications and Preprint

Work Experience

Part-Time Programmer

Part-Time IT Application

Student Assistant (HiWi)

Software Engineer

Education

Master of Science (MSc.) in Computer Science

Bachelor of Science (BSc.) in Computer Science & Engineering

Certification

Machine Learning With Big Data – University of California, San Diego (Coursera, 2018)

Skills

References