Harnessing Preference Optimisation in Protein LMs for Hit Maturation in Cell Therapy • This study explores the use of Protein Language Models (PLMs), specifically ProGen2, fine-tuned via preference optimization for Chimeric Antigen Receptors (CARs) design in cell therapy. The approach aims to enhance the maturation of therapeutic hits by leveraging high-throughput experimental data. • A core innovation: Direct Preference Optimization (DPO) enables fine-tuning by assigning preferences between CAR designs, avoiding the need for extensive labeled datasets. This method aligns PLMs with biological assay outcomes, correlating model loss with T-cell activation performance. • Results show that fine-tuned models achieve a significant correlation (Pearson r > -0.57) between model loss and CAR performance in biological assays, with mutants outperforming baseline CARs by over 2x in certain cases. • The study emphasizes the advantage of preference-driven PLMs in guided exploration of large design spaces, showcasing strong performance in identifying high-activation CAR mutants through exhaustive and greedy search strategies. • Limitations include computational intensity for exhaustive searches and the lack of integration of spatial structure data. Future efforts will explore graph-based learning techniques to address these gaps. • This work sets a new direction for using PLMs in immunotherapy, offering scalable solutions for optimizing CAR designs and paving the way for broader applications in therapeutic development. @nhammerla @coding_bio @SimonBornschein 📜Paper: arxiv.org/abs/2412.01388 #ProteinLanguageModels #Immunotherapy #MachineLearning #CellTherapy #ChimericAntigenReceptors