Predicts serine and threonine phosphorylation sites in virus-infected cells by fusing ProtT5 embeddings with a Siamese contrastive encoder.
No providers recorded yet. Browse all providers
When SARS-CoV-2 infects a cell, it rewires the host's kinase signaling: phosphoproteomics studies have mapped more than 15,000 phosphorylation events on host proteins during infection, alongside about 70 sites on the virus's own proteins. Mass spectrometry can find these sites, but it is slow and expensive, and a sequence-based predictor that can rank candidate sites lets a lab prioritize which ones to test. The difficulty is that a protein language model's embedding of a 33-residue window encodes rich sequence context without being organized around the phosphorylated/non-phosphorylated distinction, so a classifier built on it alone has to find that boundary in an unstructured space.
PhoSARte, from the Manavalan lab in the Department of Integrative Biotechnology at Sungkyunkwan University, addresses this with two parallel branches. One takes frozen embeddings from ProtTrans ProtT5 models. The other is a Siamese network trained with a contrastive loss on pairs of sequence windows, explicitly pulling same-class pairs together and pushing phosphosites and non-sites at least a fixed margin apart. The two representations are concatenated and passed to a classification head, so the language model supplies context and the contrastive branch supplies separation. UMAP projections of each branch's features show this division of labor directly.
The work was published in Briefings in Bioinformatics in 2026. It extends the group's earlier MeL-STPhos predictor, which it is benchmarked against on the same datasets.
Inputs are 33-residue windows centered on a candidate serine or threonine. Tyrosine sites were excluded because they are too sparse in the source phosphoproteomes. Training used three benchmark sets: A549 (5,387 positive and 5,387 negative sites, from Stukalov et al.), Vero E6 (3,435 positives, from Bouhaddou et al.), and a Combined set of the two, each reduced for redundancy with CD-HIT and class-balanced. Of eight ProtTrans variants screened, Prot-T5-XXL-UniRef50 was selected for the A549 and generic models and Prot-T5-XL-BFD for the Vero E6 model. The PLM embeddings pass through their own attention BiGRU. Models were trained with Adam for 30 epochs at batch size 128, dropout 0.3 and a contrastive margin of 2, with hyperparameters chosen by five-fold cross-validation. Each released PhoSARte checkpoint is about 25 MB, excluding the ProtT5 backbone.
On held-out test sets, the A549 model reached an MCC of 0.704 and AUC of 0.927, the Vero E6 model an MCC of 0.741 and AUC of 0.923, and the generic model an MCC of 0.664 and AUC of 0.906. The authors report paired-bootstrap significance (P < .05) over DeepIPs, IPs-GRUAtt, PSPred-ALE, PTransIPs and MeL-STPhos. On the adenovirus type 2 IMR-90 case study (2,344 positive and 10,495 negative sites), the generic model reached an MCC of 0.801 and AUC of 0.967. The ablations show both branches contribute: on A549, the full model's test MCC of 0.704 compares with 0.624 for the PLM branch alone and 0.646 for the contrastive branch alone.
PhoSARte is aimed at virologists and phosphoproteomics groups studying host–pathogen signaling. Its main use is ranking candidate S/T sites on host and viral proteins under infection so that experiments target the likeliest ones, and nominating kinase pathways for host-directed antiviral work. A web server accepts pasted sequences or FASTA uploads and returns labels with probability scores as CSV. The GitHub release provides a standalone prediction script for local batch runs.
PhoSARte is one of a dense cluster of SARS-CoV-2 phosphosite predictors built on the same A549 benchmark, and its contribution is methodological: pairing frozen language-model embeddings with a contrastive objective that shapes the feature space for the task. Like PhosF3C, it shows how a general protein language model can be adapted into a narrow modification predictor. Its scope is deliberately limited to virus-infected cellular contexts, and its evidence is computational: the cross-viral result is a single case study, the benchmarks are retrospective, and no new wet-lab validation accompanies the predictions. Its release of trained checkpoints and benchmark data under the MIT license makes it straightforward to reproduce and compare against.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.