This project explores the use of deep learning to predict functional annotations of proteins using sequence data. It was developed as part of the Deep Learning (GRS34806) course at Wageningen University & Research (WUR). The project aims to bridge the gap between the abundance of large-scale sequencing data and the limited experimental characterization of gene/protein functions.

To build and evaluate neural network models capable of predicting Gene Ontology (GO) terms for proteins, using both simulated and real human sequence datasets. The overarching goal is to determine how effectively deep learning models can infer protein functions based solely on amino acid sequence information.

Project Goals Develop and train a Convolutional Neural Network (CNN) on simulated protein sequence datasets to classify functional vs. non-functional proteins based on embedded sequence motifs. Fine-tune Transformer-based models (such as ProtBERT or ESM-2) to learn higher-level sequence representations for GO term prediction. Evaluate model performance under different conditions: Varying dataset sizes and sequence lengths. Different hyperparameters (e.g., dropout rate, learning rate). Reducing known positive/negative samples to test data dependency. Apply the trained models to real human protein datasets to predict five selected GO annotations: GO:0005739 — Mitochondrion GO:0007165 — Signal transduction GO:0005576 — Extracellular region GO:0055085 — Transmembrane transport GO:0043066 — Negative regulation of apoptotic process

Methods and Tools Frameworks: PyTorch, HuggingFace Transformers Data: Simulated protein sequences and annotated human protein datasets

Models: CNN-based binary classifiers (Conv1D architecture) Fine-tuned Transformer models (ProtBERT, ESM-2 small) Evaluation Metrics: Accuracy, precision, recall, F1-score