Protein Cell Classification

Multi-label classification of protein types present in biological microscopy images of blood cells.

Overview

This project applies deep learning to identify and classify co-occurring protein types within individual blood cell images. The core challenge is a multi-label classification problem — multiple protein types can be present simultaneously in a single image — which requires models and evaluation strategies that go beyond standard single-label approaches.

Problem

Given a fluorescence microscopy image of a blood cell, predict which protein types are present. The dataset exhibits class imbalance and label co-occurrence, both common in real biological imaging tasks.

Approach

• Architecture: CNN-based classification models trained on biological cell imagery
• Multi-label handling: Binary cross-entropy loss with per-class thresholding
• Evaluation: Decomposed per-class metrics (precision, recall, F1) alongside aggregate scores to isolate model performance across individual protein types
• Data challenges addressed: Class imbalance via weighted sampling and augmentation strategies

Key Challenges

• Class co-occurrence: Multiple protein types present in a single image require independent per-label prediction rather than mutually exclusive classification
• Class imbalance: Rare protein types underrepresented in training data — addressed through sampling strategies and evaluation-aware thresholding
• Evaluation design: Standard accuracy is misleading for multi-label tasks; decomposed per-class metrics were essential for meaningful model assessment

Tech Stack

• Python, PyTorch
• torchvision for image preprocessing and augmentation
• scikit-learn for evaluation metrics
• Jupyter Notebook

Repository Structure

Protein-Classification/
├── Protein Multi-label classification.ipynb   # Full training and evaluation pipeline
├── README.md
└── LICENSE

Results

Per-class F1 scores evaluated on held-out test set. Decomposed metrics revealed meaningful variation across protein types, informing targeted data augmentation for underperforming classes.

Relevance

This work directly addresses the challenge of predicting multiple biological properties from imaging data — a problem structurally analogous to multi-property biosensor performance prediction in protein engineering pipelines.

Author

Haider Ali github.com/AliHaider20 | linkedin.com/in/haiderzali