A positive-unlabeled random forest (PURF) framework trained on 5393 Plasmodium falciparum proteins and 272 protein variables prioritized a top-200 list of previously untested vaccine candidate antigens, but these predictions remain unvalidated computational hypotheses. The open problem is to convert this candidate list into experimentally supported vaccine antigen nominations: establish which prioritized candidates can be expressed and correctly folded in heterologous systems and elicit functional antibody or T-cell responses against the parasite; integrate developability and manufacturability criteria into the computational prioritization; and place the prioritization on robust methodological footing through principled reference-antigen selection. To date no candidate on this list has been experimentally validated, and the validation core and the supporting prioritization-methodology gaps remain open.

Background

A highly efficacious, long-lasting vaccine against Plasmodium falciparum malaria remains an unmet need: the two WHO-recommended vaccines, RTS,S/AS01 and R21/Matrix-M, both target a single antigen (circumsporozoite protein, CSP) and show partial efficacy, and vaccine design is hampered by the parasite's multiple life stages (sporozoite, liver, blood, sexual), extensive antigenic variation, and incomplete knowledge of immune correlates of protection. Reverse vaccinology — genome-wide computational screening for candidate antigens rather than laboratory serendipity — has been proposed to widen the antigen pipeline, but classical sequential-filtration approaches lack standardized thresholds, and supervised machine learning is limited by the scarcity of labeled examples: only 52 high-quality known P. falciparum antigens can be assembled (intersection of literature curation and Immune Epitope Database epitope mappings), while reliable non-antigens cannot be identified at all. Positive-unlabeled (PU) learning is a semi-supervised setting suited to exactly this situation: a small set of positives (known antigens) is labeled, all remaining proteins are treated as unlabeled rather than negative, and the model learns to rank the unlabeled set by similarity to the positives. Chou, Ouattara, Adams, Berry, Takala-Harrison and Cummings (npj Systems Biology and Applications 10:44, 2024) built a PU random forest (PURF) over 5393 proteins of the P. falciparum 3D7 reference strain with 272 proteomic, structural, functional, immunological, genomic, and transcriptomic variables, using a PU prior ('positive level', selected as 0.5 by estimated-ROC sweep) and ensembles filtered by validation performance on four reference antigens chosen to span genetic diversity (CSP, RH5, MSP5, P230). Candidates were ranked by probability scores (proportion of trees predicting antigenicity), filtered by gene essentiality from piggyBac transposon-insertion mutagenesis screening data (a genome-scale perturbation assay in which transposon insertion disrupts each gene, so that insertion intolerance marks genes the parasite cannot lose; keeping essential antigens means vaccine-induced selection against the antigen threatens parasite viability and may reduce immune escape), and characterized by gene ontology and Malaria Cell Atlas single-cell expression across life stages. Robustness to reference-antigen choice was probed only by a leave-one-known-antigen-out adversarial control. The output is a top-200 candidate list (interactive report at https://mrp-bioinformatics.github.io/malaria_antigen_candidates/; data at https://doi.org/10.13016/me1l-1ahr; PURF code at https://github.com/RTChou/purf), most of them never previously proposed as vaccine antigens. The authors themselves flag what remains unresolved: the predictions are computational hypotheses requiring experimental validation (heterologous expression, correct folding, functional antibody or T-cell responses); practical developability filters (isoelectric point, molecular weight, folding propensity) were not integrated into the prioritization; and candidates could alternatively be prioritized with a different set of reference antigens selected using other criteria, leaving the reference-antigen selection strategy open. The same group subsequently extended the framework to P. vivax (npj Systems Biology and Applications 10:133, 2024), showing that including known antigens from P. falciparum is critical for model performance while unlabeled-only cross-species data can misdirect the model toward species classification — but that study too defers experimental validation of its candidates. A 2025 scoping review of machine learning in malaria vaccine research identifies both Chou papers as the key antigen-discovery studies and lists experimental validation, data quality, and generalizability as persistent field-wide gaps.

Problem Statement

Determine whether the vaccine candidate antigens that positive-unlabeled learning prioritized for Plasmodium falciparum are genuine vaccine targets, and put the prioritization itself on defensible footing. The objective is to convert the published top-200 computational candidate list into experimentally supported vaccine antigen nominations: (1) experimental validation — a ranked, actionable validation plan that selects the top-prioritized candidates (including the candidates already carried through the gene-essentiality filter and the life-stage expression characterization) and specifies their heterologous expression, the establishment of correct folding, and the testing of whether they elicit immune responses that act on the parasite, i.e., functional antibodies (for example growth/invasion inhibition of blood-stage parasites or transmission blocking in mosquito-feed assays) or protective T-cell responses, in vitro and in vivo; (2) developability-integrated prioritization — incorporate practical vaccine-design criteria such as isoelectric point, molecular weight, folding propensity, and manufacturability into the PURF (positive-unlabeled random forest) prioritization so that the candidates advanced to expression and testing are those one can actually produce and formulate; (3) methodological robustness — establish principled strategies for reference-antigen selection, which the source notes could alternatively use different sets of reference antigens selected by other criteria, and characterize how adding newly validated antigens to the known-antigen set changes feature importances and the robustness of the ranking. (2) and (3) are addressed by computational analysis, while (1) is addressed by the ranked validation plan above rather than by wet-lab data; and (3) holds independently of whether (1) yields a newly validated antigen.

References

  1. Chou RT, Ouattara A, Adams M, Berry AA, Takala-Harrison S, Cummings MP (2024). Positive-unlabeled learning identifies vaccine candidate antigens in the malaria parasite Plasmodium falciparum. npj Systems Biology and Applications 10:44. doi:10.1038/s41540-024-00365-1, https://doi.org/10.1038/s41540-024-00365-1
  2. Chou RT, Ouattara A, Takala-Harrison S, Cummings MP (2024). Plasmodium vivax antigen candidate prediction improves with the addition of Plasmodium falciparum data. npj Systems Biology and Applications 10:133. doi:10.1038/s41540-024-00465-y, https://doi.org/10.1038/s41540-024-00465-y
  3. Balmer AJ, Lee C, Pearson RD, Ariani CV, Almagro Garcia J (2025). Pf-PeptideFilter: An Interactive Catalogue of Peptide Vaccine Candidates for Plasmodium falciparum. bioRxiv preprint (posted 17 December 2025). doi:10.64898/2025.12.15.694343, https://doi.org/10.64898/2025.12.15.694343
  4. Khanday S, Sayeed M, Jafri N, Jafri IF, Jafri RF, Ashraf G, Safwat S, Nasr DS (2025). From data to immunity: the role of machine learning in advancing malaria vaccine research: a scoping review. Tropical Diseases, Travel Medicine and Vaccines 11:38. doi:10.1186/s40794-025-00271-2, https://doi.org/10.1186/s40794-025-00271-2
  5. Aamir M, Rezaee K, Saberi Anari M (2026). Intelligent in-silico prioritization of antimalarial peptide candidates under explicit physicochemical windows via de novo CTCM-Neo generation and conformal-gated calibrated classification. Frontiers in Cellular and Infection Microbiology 16:1707267. doi:10.3389/fcimb.2026.1707267, https://doi.org/10.3389/fcimb.2026.1707267
  6. Tang Q, Li K, Zhang L, Cheng F, Qiu J (2026). An evolutionary multitasking method for positive and unlabeled learning. Knowledge-Based Systems 331:114673. doi:10.1016/j.knosys.2025.114673, https://doi.org/10.1016/j.knosys.2025.114673
  7. Li L, Wu Z (2026). Learning Optimal Classification Trees from Positive and Unlabeled Data. In: 38th Chinese Control and Decision Conference (CCDC). doi:10.1109/CCDC69976.2026.11560433, https://doi.org/10.1109/CCDC69976.2026.11560433
  8. Colstrup E, Nakajima R, Krol JMM, Lamers OAC, de Assis RR, Jain A, Jasinskas A, Iliopoulou E, de Bes-Roeleveld HM, Franke-Fayard BMD, Roestenberg M, Felgner PL, Murugan R (2025). Correlative humoral and cellular immunity to genetically attenuated malaria parasites in humans. iScience 28:112589. doi:10.1016/j.isci.2025.112589, https://doi.org/10.1016/j.isci.2025.112589