[Paper Review] Harnessing Preference Optimisation in Protein LMs for Hit Maturation in Cell Therapy

Katarzyna Janocha, Alan C. H. Ling+4 more|arXiv (Cornell University)|Dec 2, 2024

Viral Infectious Diseases and Gene Expression in Insects4 citations

TL;DR

This paper proposes using preference fine-tuning of protein language models (PLMs) on high-throughput experimental data to accelerate hit maturation in chimeric antigen receptor (CAR) therapy. By training an auto-regressive PLM on a preference dataset derived from thousands of CAR variants tested in cell assays, the model achieves strong correlation (r = -0.574) between predicted loss and biological performance, enabling effective few-shot discovery of improved CAR mutants.

ABSTRACT

Cell and immunotherapy offer transformative potential for treating diseases like cancer and autoimmune disorders by modulating the immune system. The development of these therapies is resource-intensive, with the majority of drug candidates failing to progress beyond laboratory testing. While recent advances in machine learning have revolutionised areas such as protein engineering, applications in immunotherapy remain limited due to the scarcity of large-scale, standardised datasets and the complexity of cellular systems. In this work, we address these challenges by leveraging a high-throughput experimental platform to generate data suitable for fine-tuning protein language models. We demonstrate how models fine-tuned using a preference task show surprising correlations to biological assays, and how they can be leveraged for few-shot hit maturation in CARs. This proof-of-concept presents a novel pathway for applying ML to immunotherapy and could generalise to other therapeutic modalities.

Motivation & Objective

• To address the scarcity of large-scale, standardized datasets in immunotherapy by generating high-throughput experimental data from thousands of CAR variants.
• To explore whether protein language models fine-tuned via preference optimization can predict biologically relevant CAR performance.
• To enable few-shot hit maturation in CAR design by leveraging model-generated predictions to guide exploration of the sequence space.
• To assess the generalization potential of such models to unseen CAR sequences and validate their predictive power on a held-out validation set.

Proposed method

• Fine-tune a pre-trained auto-regressive protein language model using a preference learning objective on experimental data from high-throughput T-cell reporter assays.
• Construct a preference dataset by comparing relative performance (ΔGFP) of single and double mutants against their parent CARs.
• Use greedy search with a diversity-aware generation function to explore the sequence space around parent CARs, generating candidate mutants.
• Evaluate model predictions by correlating model loss with experimental ΔGFP values across training and validation sets.
• Apply normalization to account for plate-specific biases in high-throughput assay data.
• Use a recursive generation function that selects top-k amino acid substitutions at each position in the variable domain, ensuring valid sequences through a validity filter.

Figure 1: CAR structure with scFV and VHH binding domains (see text for details).

Experimental results

Research questions

• RQ1 Can a protein language model fine-tuned via preference learning predict biologically relevant performance improvements in CAR-T cell therapy?
• RQ2 How well does model loss correlate with actual experimental readouts (ΔGFP) in high-throughput assays?
• RQ3 Can the model guide effective few-shot exploration of the sequence space to discover superior CAR mutants?
• RQ4 Does the model generalize to unseen CAR sequences in a validation set?
• RQ5 What factors influence the model’s ability to discover high-performing mutants, and is performance driven by biological structure or dataset artifacts?

Key findings

• The fine-tuned PLM shows a strong negative Pearson correlation (r = -0.574) between model loss and experimental T-cell reporter activation (ΔGFP), indicating predictive power.
• For Candidate 1, the correlation is even stronger (r = -0.699), and fine-tuning significantly improves correlation compared to the pretrained model (from r = 0.218 to r = -0.699).
• Greedy and exhaustive search both identify mutants that outperform the parent CAR, with several mutants showing more than double the performance.
• At least one mutant per parent sequence outperforms its parent in 7 out of 10 tested sequences, including in the validation set (S8–S10), suggesting generalization.
• The model’s predictive performance is higher for Candidate 1 than for Candidate 2, though the underlying cause—biological or data-related—remains unclear.
• Exhaustive search yields better results than greedy search, highlighting the importance of comprehensive exploration despite higher computational cost.

Figure 2: Simplified candidate generation pipeline. Using phage display based on a highly diverse starting library we screen for binders to the target of interest using phage display. Resulting candidates are evaluated in cells in scalable proprietary assays.