This project investigates how data split size and sampling methods affect TCR-epitope binding affinity prediction performance using catELMo embeddings and a multilayer perceptron (MLP) model.
We analyzed various data selection strategies and found that the catELMo Entropy Low strategy achieved the best performance:
- AUC: 0.9717
- Accuracy: 91.94%
- F1 Score (micro): 0.9194
This demonstrates that training on sequences with lower entropy (higher model confidence) yields better generalization and more consistent predictions with reduced noise.
Model weights and a copy of these files can be found at:
https://drive.google.com/drive/folders/1vDSaw27REq1wLWhkIeEjuQaAWB9fVHCP?usp=drive_link
- Loveneet Singh - lpsingh@asu.edu
- Seonho Yeom - syeom3@asu.edu
- Adrian Zhang - awzhang1@asu.edu
- Rohit Khoja - rkhoja2@asu.edu
- Prasham Titiya - ptitiya@asu.edu