Protein Structure Prediction with CNNs

This notebook, CNN_Randy.ipynb, provides a comprehensive workflow for predicting protein secondary structure from amino acid sequences using Convolutional Neural Networks (CNNs).

Overview

Protein structure prediction is a key challenge in bioinformatics, as a protein's function is determined by its 3D structure. Experimental methods to determine structure are expensive and time-consuming; deep learning, particularly CNNs, offers a scalable computational alternative.

This notebook focuses on predicting two levels of secondary structure:

• sst3: Three-state classification (helix, strand, coil)
• sst8: Eight-state classification (more granular)

Workflow

The notebook follows these main steps:

1. Introduction & Setup

• Explains the challenge and project goals.
• Installs required Python packages, including PyTorch, pandas, scikit-learn, matplotlib, seaborn, etc.
• Sets random seeds for reproducibility.

2. Data Loading & Analysis

• Loads the protein sequence dataset (CSV format, e.g. from Google Drive).
• Inspects the dataset and its columns (e.g., pdb_id, chain_code, seq, sst3, sst8, etc.).
• Performs descriptive statistics and visualizations:
  • Sequence length distribution
  • Amino acid frequency distribution
  • Secondary structure class imbalance for sst3 and sst8

3. Data Preprocessing

• User-configurable preprocessing parameters:
  • Cutoff length for filtering sequences
  • Padding/cropping length
  • Option to filter non-standard amino acids
  • Embedding method for encoding protein sequences (e.g., one-hot, learnable, ProtBERT, TAPE)
• Functions for filtering, cropping, padding, and encoding sequences and secondary structure labels.

4. Exploratory Data Analysis (EDA)

• Analysis of class imbalances
• Strategies for dealing with long sequences and class imbalance (e.g., dropping long sequences, padding/cropping, etc.)

Usage

To use this notebook:

1. Upload your protein sequence CSV to your storage (e.g., Google Drive).
2. Update the file path in the notebook to point to your data.
3. Run all cells sequentially to:
   • Install dependencies
   • Load and preprocess the data
   • Visualize class distributions
   • Prepare the data for CNN input

Requirements

• Python 3.10+
• PyTorch
• Pandas
• Matplotlib
• Seaborn
• scikit-learn

You can install dependencies via:

pip install torch torchvision torchaudio pandas matplotlib numpy scikit-learn seaborn

File Structure

• Introduction & Setup: Project description and environment setup.
• Data Loading & Analysis: Loads and analyzes the protein dataset.
• Data Preprocessing: Filters, crops, and encodes sequences for CNN input.
• EDA & Visualization: Shows class and amino acid distributions.

Notes

• The notebook is designed for use in Google Colab but can be adapted to other environments.
• The dataset path must be updated to reflect your local or cloud storage.
• Preprocessing parameters are configurable for experimentation.

References

• PyTorch Documentation
• Protein Structure Prediction
• ProtBERT