PlasmoFP

PlasmoFP (Plasmodium Function Predictor) is a set of deep learning models that assign Gene Ontology (GO) terms to proteins from the malaria parasite genus Plasmodium. More than two decades after the first Plasmodium falciparum genome was published, a large share of parasite gene products remain labelled "proteins of unknown function" (PUFs) or carry annotation in only one or two GO subontologies — in some species roughly half of all proteins have no GO term at all. Standard annotation transfer struggles here because apicomplexan proteins share little sequence identity with the model-organism proteins that anchor curated databases.

The central design choice is phylogenetic: rather than training on a broad, SwissProt-wide protein set, PlasmoFP trains on proteins from the SAR supergroup (Stramenopiles, Alveolates, Rhizarians), the eukaryotic clade that contains Plasmodium. It pairs that training universe with structure-imbued protein representations instead of raw sequence features, on the premise that structural signal survives the sequence divergence that defeats homology search. Ablations show both choices help independently.

PlasmoFP was developed by Srivastava, Berenberg, Qassab, Wang, Bonneau, and Carlton at the Johns Hopkins Malaria Research Institute, New York University, and Prescient Design, and posted to bioRxiv in September 2025 (revised October 2025). It remains a preprint awaiting peer review.

#Key Features

• Phylogenetically scoped training: Models learn from SAR supergroup proteins rather than a generic reference set, targeting the evolutionary neighbourhood of Plasmodium instead of distant model organisms.
• Structure-imbued representations: Proteins are encoded as fixed-dimensional TM-Vec embeddings, which fine-tune a ProtT5 sequence encoder against a structural-similarity objective, so predictions draw on structural rather than purely sequence-level similarity.
• Calibrated epistemic uncertainty: Each subontology model is a 20-member deep ensemble, and per-term uncertainty is folded into the score by subtracting the median absolute deviation from the ensemble median.
• Explicit false discovery control: Predictions are reported at expected-FDR (eFDR) thresholds rather than a single Fmax-optimised cutoff, letting users trade coverage against precision for their own downstream use.
• Robust to intrinsic disorder: Performance holds up on proteins with high predicted disorder content, a substantial fraction of the Plasmodium proteome that sequence-based annotation handles poorly.

#Technical Details

PlasmoFP consists of three independently trained models, one per GO subontology. Each is a compact multilayer perceptron with hidden layers of 256 and 128 units that maps a frozen TM-Vec embedding to a multi-label vector of GO terms; no gradients propagate into the encoder. The label spaces cover 744 molecular function, 1,548 biological process, and 255 cellular component terms. Training data was drawn from 907,145 SAR supergroup proteins in UniProt release 2024_04, deduplicated with MMseqs2 at 90% identity and then split into train/validation/test partitions by whole 30%-identity clusters to prevent leakage, yielding 289,930 (MF), 247,286 (BP), and 227,756 (CC) training sequences. Manually asserted Plasmodium GO annotations were withheld entirely to serve as a holdout set.

On that curated Plasmodium holdout set, PlasmoFP outperforms DeepGOPlus and ProteInfer on Fmicro and Smin across all three subontologies, including at the strictest 5% eFDR. For molecular function, PlasmoFP reaches Fmicro of 0.885 at 5% eFDR and 0.902 at 20% eFDR, against 0.639 for ProteInfer and 0.570 for DeepGOPlus; biological process shows a comparable margin. Ensemble size matters: raising k to 20 produced the largest gains in precision and FDR reduction, while Monte Carlo dropout and temporal ensembles failed to yield meaningful epistemic uncertainty.

#Applications

The models were applied without per-species refitting to all 19 Plasmodium proteomes in PlasmoDB. Merging PlasmoFP predictions with existing annotations cut the fraction of proteins lacking any GO term from 15–59% to 3–28%, and raised the fully annotated fraction from 7–42% to 36–68%. Gains are largest for understudied species, where curation has lagged far behind P. falciparum — several reach a roughly 90% relative increase in fully annotated proteins even at a stringent 5% eFDR. Newly predicted terms cluster into biologically coherent groups, expanding the RNA-associated proteome and adding protease and transport annotations relevant to invasion, egress, and nutrient acquisition. Researchers can query predictions per species and per eFDR threshold through a hosted explorer and a Colab notebook, or run the released checkpoints on their own sequences.

#Impact

PlasmoFP demonstrates that narrowing a function predictor's training universe to a phylogenetically relevant clade beats generic breadth for a divergent pathogen — a template applicable to other neglected eukaryotic parasites. Pairing that with uncertainty quantification and FDR control rather than a single hard threshold yields annotations that downstream enrichment analyses can use without the false-positive bias unqualified predictions introduce. The practical output is a more complete and more evenly distributed functional annotation across the genus, supplying hypothesis-generating leads for drug-target discovery in species that have received little experimental attention. The scope is narrow by design: these are taxon-specific models, not a general-purpose protein predictor, and the label space is restricted to GO terms seen at least 50 times in SAR training data. Code, trained weights, and data splits are public, though the repository carries no license file.