Fine-Tuning a Protein Language Model for Epitope Library Generation and Sequence Filtering

Linear peptide epitopes play a key role in vaccines, immunotherapies, and diagnostics, yet discovering them remains difficult because peptide sequence space grows exponentially and experimentally confirmed epitopes are relatively few. This chapter presents a computational framework for the generation of synthetic linear epitope libraries that reproduce key statistical and sequence-level properties of naturally occurring epitopes. A large language model pre-trained on protein sequences and subsequently fine-tuned on curated epitope datasets is employed to generate biologically plausible peptide candidates beyond those observed experimentally. To further enrich functional relevance of the generated sequences, they are filtered using transformer-based classification models trained on viral and bacterial epitope datasets. These models are applied as discriminative tools to distinguish likely epitopes from non-epitopes rather than to infer pathogen origin. The resulting libraries are designed to be compatible with high-throughput experimental screening platforms, such as droplet microfluidics, enabling efficient integration of generative artificial intelligence with experimental immunology workflows.