File size: 2,092 Bytes
034e679
 
 
6022975
034e679
 
 
6022975
034e679
9086f1c
 
 
 
 
 
 
 
 
 
 
 
ddd27c9
9086f1c
 
ddd27c9
9086f1c
 
 
034e679
9086f1c
95c36ab
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
---
license: mit
title: The Genomic Oracle
sdk: static
emoji: 🔥
colorFrom: indigo
colorTo: purple
short_description: UMGC Bioinformatics Capstone
---
# 🧬 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