FURI | Spring 2022

Boosting TCR and Epitope Binding Affinity via Supervised Contrastive Loss

Health icon, disabled. A red heart with a cardiac rhythm running through it.

In this paper, the research team investigates using Supervised Contrastive Loss (SCL) instead of Cross Entropy Loss (CEL) as the loss optimization function in existing machine learning models that are used to predict T-cell receptor (TCR) and Epitope binding affinity. The goal of SCL is to pull positive examples away from negative ones in latent space, before classification. The resulting process is one that trains a projection network using SCL, freezes the gradients of that network, and finally trains two dense layers on top of it using CEL for final classification.

Student researcher

Nicholas A. Moran

Computer science

Hometown: Glendale, Arizona, United States

Graduation date: Spring 2022