🧬 The Genomic Oracle 🧬

University of Maryland Global Campus (UMGC) | Bioinformatics Capstone Project

Welcome to the official repository for The Genomic Oracle, a cascaded machine learning pipeline designed for high-precision DNA sequence classification and phenotypic prediction.

Our Mission

As genomic datasets grow exponentially, the need for rapid, automated sequence annotation is critical. The Genomic Oracle serves as an intelligent routing network, evaluating raw DNA sequences through a multi-stage gauntlet to classify their biological function, structural feature type, and associated phenotypic risks.

The 4-Stage Cascading Architecture

Our platform utilizes a highly specialized, branching AI architecture hosted on Hugging Face ZeroGPU infrastructure:

1. Level 1: The Gene Finder (Coding vs. Non-Coding): A logistic regression machine learning classifier that rapidly screens raw k-mer vectors to identify protein-coding potential.
2. Level 2: The Multi-Feature Classifier (LightGBM): Sequences flagged as coding are passed through a LightGBM gradient boosting model to classify them into one of 7 highly specific structural features.
3. Level 3: Phenotype Classification (Lean vs. Obese): A custom ALiBi-configured BERT architecture evaluates specific coding regions to predict downstream phenotypic associations.
4. Level 4: The Promoter Network (DNABERT-2): Sequences flagged as non-coding are routed to a neural network transformer model that analyzes spatial attention tensors to identify regulatory promoter regions.

Integration: The pipeline concludes with an automated API routing to NCBI and Ensembl databases for real-world chromosomal coordinate mapping and validation.

The Team

We are a team of graduate researchers specializing in data science, machine learning, and bioinformatics.

Created by: Kadir Galindo, Duncan Hall, Rebecca Mellinger, George Paccione