Estimated limits of organism-specific training for epitope prediction

Background The identification of linear B-cell epitopes remains an important task in the development of vaccines, therapeutic antibodies and several diagnostic tests. Machine learning predictors are trained to flag potential epitope candidates for experimental validation and currently, most predictors are trained as generalist models using large, heterogeneous data sets. Recently, organism-specific training has been shown to improve prediction performance for data-rich organisms. Unfortunately, for most organisms, large volumes of validated epitope data are not yet available. This article investigates the limits of organism-specific training for epitope prediction. It explores the validity of organism-specific training for data-poor organisms by examining how the size of the training data set affects prediction performance. It also compares the performance of organism-specific training under simulated data-poor conditions to that of models trained using traditional large heterogeneous and hybrid data sets. Results This work shows how models trained on small organism-specific data sets can outperform similar models trained on (potentially much larger) heterogeneous and mixed data sets. The results reported indicate that as few as 20 labelled peptides from a given pathogen can be sufficient to generate models that outperform widely-used predictors from the literature, which are trained on heterogeneous data. Models trained using more than about 100 to 150 organism-specific peptides perform consistently better than most generalist models across a wide variety of performance measures, and in some cases can even approach the performance of organism-specific models trained on considerably larger data sets. Conclusions Organism-specific training improves linear B-cell epitope prediction performance even in situations when only small training sets are available, which opens new possibilities for the development of bespoke, high-performance predictive models when studying data-poor organisms such as emerging or neglected pathogens.