Protein Structure Prediction

Attention-Based Protein Structure Prediction Developed a simplified attention-based model to predict protein structures from amino acid sequences, focusing on predicting protein angles. Unlike AlphaFold, our model does not use MSA or ESM embeddings. We trained the model using the SidechainNet dataset, with two approaches: one with PSSM data and another using only amino acid sequences. The model successfully generates 3D protein structure predictions, with performance improving by incorporating secondary structures and PSSM data. The method has potential for further enhancement by using MSA data and predicting coordinates instead of angles.

Key Highlights:

• Implemented a transformer-based attention model for protein angle prediction.
• Trained with amino acid sequences, secondary structure data, and PSSMs.
• Visualized protein structures through interactive 3D plots.
• Achieved promising results compared to actual protein structures.