Graduate Student Internship at Exact Sciences - Summer 2025

Role: NGS Application Developer
Team: P.O.B.O.T. (Precision Oncology Bioinformatics Operations Team)

Languages: Python, Bash
Infrastructure & DevOps: AWS (EC2, S3, CLI), HPC clusters, CI/CD, GitLab, Pixi
Workflow & HPC Scheduling: Nextflow (DSL2/Groovy), SLURM job scripts
Bioinformatic tools: IGV, BAMSurgeon, seqtk, bcftools, samtools, wgsim, pysam, BioPython, numpy, re, pandas, matplotlib
Project Management: Jira, Confluence, Agile and Waterfall methodologies

On-going. Completion by August 29.

As the NGS Application Developer on the POBOT team, I am developing a suite of CLI tools used for genomic data simulation to help facilitate the development of variant calling algorithms, diagnostic studies, and maintanence of bioinformatic pipelines in the OncoExTra product. My role involves designing and implementing robust, scalable workflows using Nextflow and Groovy, integrating bioinformatics tools and libraries, and optimizing workflow performance on high-performance computing (HPC) clusters, ensuring efficient resource utilization and job scheduling using SLURM. Additionally, I am implementing CI/CD processes for continuous integration and deployment of the software tools I am building, and contribute to the development of automated testing frameworks to ensure the reliability and accuracy of bioinformatics analyses. My work supports the development of precision oncology solutions, enabling personalized treatment strategies for cancer patients.

Machine Learning for Precision Oncology:
Predicting Personalized Anti-Cancer Treatments for 200,000+ Breast Cancer Patients

Languages: Python
Libraries: Keras, Scipy, PyTorch, RDKit, NumPy, Pandas, matplotlib/seaborn

In this project, my team and I built a machine learning pipeline to match breast cancer patients with personalized anti-cancer therapeutics and predicted their potential treatment responses.

Our dataset comprised 204,026 breast cancer patients and incorporated 86 clinical, experimental, and drug-related features aggregated from multiple heterogeneous sources. The raw data required extensive cleaning, normalization, and integration to ensure quality and comparability across studies. This process included resolving missing values, unifying feature definitions, and harmonizing measurement scales to form a single, consistent analytical dataset.

In addition, We utilized machine learning and advanced data processing techniques to analyze complex, multi-source datasets consisting 86 dimensions. This included dimensionality reduction methods (PCA, UMAP), feature attraction of drug characterization metrics (IC50 calculations, PK/PD measures), and a suite of machine learning algorithms — including Gaussian Naive Bayes, Nearest Neighbors, feed-forward neural networks, binary classifiers, and k-fold cross-validation — to build predictive models for personalized anti-cancer therapeutics and patient treatment response.

Natural Language Processing for Chatbot Analysis:
Predicting Winners and Prompt Difficulty Across 20 LLMs

Languages: Python
Libraries: Sklearn, Pandas, NumPy, matplotlib/seaborn

This project applied natural language processing (NLP) and machine learning techniques to evaluate and interpret the performance of 28 distinct chatbots and large language models (LLMs) across a diverse set of questions and responses.

My team's analysis began with exploratory data analysis (EDA) to investigate embedded response vectors, extracting meaningful patterns and relationships within the data. To enhance our feature space and interpretability, we employed one-hot encoding for categorical variables, K-means clustering, calculated ELO ratings, Principal Component Analysis (PCA), and statistical feature extraction. For performance assessment and robustness, we incorporated different types of regression testing (linear, logisitic, lasso, and ridge), confusion matrices, k-fold cross-validation, and various visualization tools (histograms, kernel density estimation (KDE) plots, box-and-whisker plots, and waterfall diagrams).

Our results demonstrated that embedded NLP features combined with statistical models can effectively predict competitive chatbot outcomes and quantify question difficulty, offering a data-driven framework for evaluating the capabilities of conversational AI systems.

Molecular substructure search via Graph based modeling and algorithms

Languages: Python, C++, BASH
Libraries: NetworkX, matplotlib, itertools, Eigen/Eigen, complex, utility, iostream, cmath, vector and other C++ container libraries

Inspired by the RDKit library, I used python to develop my own library and molecular fingerprinting algorithm that enables users to perform substructure searches via the terminal and BASH commands for my final project in the CHEM274A course at UC Berkeley. In addition, this program is capable of generating visuals to obtain graph representation of user-chosen molecules. It also assists the user in identifying other molecular properties such as different functional groups, aromatic structures, and features using graph traversal algorithms.

Included in the repository are additional files used to draft the main program:
- A python file where I imeplement the depth-first-search algorithm to detect unique cycles in a ring, and to gain a better understanding of how libraries like RDKit and NetworkX uses DFS/BFS to detect cycles in aromatic compounds.
- A python parser file (generously provided by UC Berkeley CHEM274A course instructors) used to parse .sdf files.
- A C++ program (.cpp, .hpp, and .tpp files) used to compare graph representations, traversals, and computations of molecules using the Eigen library in C++ vs. NetworkX library in Python

Systemic evaluation and run time analysis of molecular simulations

Languages: Python, C++, Git
Libraries: NumPy, math, random, fstream, array, vector, utility, chrono

In collaboration with 2 other engineers, we've modeled lennard-jones fluidic systems, simulated particles in motion via Monte Carlo methods, and refactored scripts using Python Standard Library into NumPy and C++ for our final project in CHEM280 at UC Berkeley. To evaluate the space and run time complexity of our programs, and to assess the performance of our refactored scripts, BASH commands were used to time and compile our programs that can be used to compute the energy of particles in motion, and to make accurate predictions of larger systems through molecular dynamic simulations.

BANKING SYSTEM APPLICATION GitHub Repo

Languages: Python
Libraries: Python standard library

In collaboration with 3 other engineers, we've simulated a banking system capable of creating user accounts, withdrawals, deposits, money transfers, payments, and 2% cashbacks. The application also keeps a historical record of transactions and prevents users from creating invalid or dangerous executions (i.e. overdrafts, duplicate withdrawals, etc.). This application has gone through a series of tests to ensure its usability by business owners who may want to implement a purchasing or bank-like system.