Deep Learning & Bio-AI
Research, model architecture design, and mathematical visualization for predicting three-dimensional RNA tertiary folding structures directly from nucleotide sequences.
Python
PyTorch
Deep Learning
Bioinformatics
Manim
manim-slides
NumPy
SciPy
1. Problem Statement & Biological Context
Determining the tertiary (3D) conformation of ribonucleic acid (RNA) molecules is critical for understanding cellular mechanisms and accelerating targeted therapeutics. Traditional wet-lab methods (such as X-ray crystallography or cryo-EM) remain time-intensive and costly. This project explores modern deep learning architectures to accurately infer spatial 3D coordinates and base-pairing interaction patterns directly from primary sequence inputs.
2. Technical Architecture & Methodology
- Sequence Representation & Feature Pipelines: Engineered numerical representations that encode nucleotide sequences, secondary structure priors, and contact probability maps into dense embedding spaces.
- PyTorch Model Development: Implemented and trained deep neural network architectures tailored to capture long-range base-pairing dependencies and geometric coordinate transformations.
- Loss Formulations & Validation Metrics: Benchmarked structural alignment and accuracy using domain-standard metrics including Root Mean Square Deviation (RMSD), TM-score, and lDDT.
- Mathematical Visualizations (Manim): Designed programmatic, high-resolution animations using
Manimandmanim-slidesto break down high-dimensional vector spaces, folding mechanics, and network operations into intuitive visual presentations.
Focus: Explainability and biological plausibility — combining mathematically sound structural predictions with clear, intuitive visualizations of complex molecular dynamics.