Application of Machine Learning Models in Predicting Plasmodium falciparum Antigens for Malaria Vaccine Development

Background: Due to the time- and resource-intensive nature of traditional vaccine design, vaccine development has evolved from purely laboratory-based methods to a hybrid approach that combines computational and laboratory evaluations. These computational methods can be applied to identify novel vaccine targets, such as protein antigens, and machine learning (ML) has demonstrated the ability to analyze large datasets and predict protein antigenicity by recognizing complex patterns that might elude human researchers. This study aimed to compare various ML models for classifying P. falciparum protein antigen candidates and to apply the best-performing model to screen the P. falciparum reference proteome. Methodology: Thirty-two (32) ML models grouped into traditional ML, deep learning, and protein language models (PLMs) were trained and tested on the PlasmoFAB dataset, using a 70/20/10 split and a homology-aware cross-validation (MMseqs2 clustering at ≤30% identity), with the PlasmoFAB MMseqs2 test split (60 proteins) held out as a locked test set. Models were evaluated on their performance metrics and compared using the paired DeLong test with Holm correction, McNemar’s test, and an exploratory Friedman/Nemenyi rank comparison. Calibration was assessed via Brier score, ECE, calibration slope/intercept. A proteome-wide screen of the P. falciparum 3D7 proteome was then performed using the best model’s 10-fold ensemble to identify new potential protein antigen candidates. Results: Under the 70/20/10 split, the PLM ProtT5-XL outperformed all models with an ROC-AUC of 0.995, and in homology-aware cross-validation, ProtT5-XL still achieved the highest pooled ROC-AUC (0.944), with no statistically significant difference from ESM-1b or ESM-2-650M, but significantly exceeding all other models. In the proteome-wide screening, the ProtT5-XL ensemble prioritized 1,615 candidate-positive proteins, of which 1,171 were not present in the PlasmoFAB dataset. After homology de-replication at ≥90%, 1,166 candidates remained with 512 uncharacterized proteins and 22 proteins with previously defined functions in UniProtKB. Conclusion: Pretrained protein language models, particularly ProtT5-XL, can predict P. falciparum protein antigens more effectively than traditional ML and deep learning models. The resulting 1,166 proteins provide a ranked list of candidates for experimental validation.