PlantPTM

Post-translational modifications are how a plant cell reconfigures a protein it has already made: phosphorylation switching a kinase cascade on within seconds of a drought signal, ubiquitination marking a transcription factor for degradation, N-glycosylation deciding whether a receptor ever reaches the cell surface. Mapping these sites experimentally means a separate enrichment chemistry and mass-spectrometry run for each modification type, which is why the plant PTM record is both small and lopsided. Arabidopsis and rice dominate it, and the predictors available to a wheat or soybean researcher were mostly trained on human and yeast proteomes, one modification type at a time, inside a short fixed window around the candidate residue.

PlantPTM, from the rice group at Henan Agricultural University, addresses that generalization problem by declining to commit to a single view of the sequence. Four encoders run in parallel over the same input: a CNN for local sequence motifs, a bidirectional GRU for longer-range dependencies, the ProtBert protein language model from ProtTrans for the evolutionary signal held in pretrained representations, and a handcrafted branch fusing a PSSM profile with one-hot encoding. A self-attention feature-fusion module weighs those four streams against one another, and a three-layer fully connected decoder emits a per-residue modification probability. The point of the redundancy is that the views degrade differently: where a species is thinly represented in training, the frozen language model and the alignment-derived profile still carry signal a purely supervised motif detector never had the data to learn. An ablation confirms each branch contributes substantively.

The work was published in Molecular Plant in August 2026, with Gensheng Dou, Kaiyuan Wang and Zhaohui Qin as co-first authors. It sits alongside broader residue-level predictors such as AstraPTM, but is built for the plant kingdom, where training data is scarcer and cross-species transfer is the binding constraint.

#Key Features

• Nine modification types: N-glycosylation, S-acylation, 2-hydroxyisobutyrylation, crotonylation, succinylation, malonylation, acetylation, ubiquitination and phosphorylation, each served by its own trained checkpoint.
• Multi-view encoder fusion: four parallel encoders feed a self-attention fusion layer, so sequence motifs, long-range context, pretrained embeddings and evolutionary profiles are combined rather than chosen between.
• Transfer to unseen species: leave-one-species-out validation that removes data-rich Arabidopsis or rice from training entirely still holds AUROC above 0.8 on the withheld species.
• Wet-lab confirmation: independent in-house rice mass spectrometry, deposited as ProteomeXchange PXD078192, rather than benchmark splits alone.
• Structure handoff: five of the nine types can be passed to AlphaFold 3 to generate PTM-modified structures, demonstrated on the Arabidopsis G-protein α subunit GPA1.

#Technical Details

Training data was pooled from CPLM, EPSD 2.0, PhosPhAt 4.0, Plant PTM Viewer and qPTMplants plus the literature, giving 657,623 PTM sites on 211,535 proteins across six species — Arabidopsis, rice, maize, wheat, soybean and tomato — then reduced for redundancy at 30% sequence identity. PSSM features come from three PSI-BLAST iterations against UniRef50; language-model embeddings come from the Rostlab/prot_bert checkpoint. On the independent test set the mean AUROC across the nine types is 0.8640, peaking at 0.9699 for N-glycosylation. Against 25 existing general-purpose and plant-specific tools, the margin over the best competitor for each type ranges from 2.90% to 19.08%, averaging 15.46%. The in-house rice mass-spectrometry validation returns AUROC 0.8089 for ubiquitination, 0.8150 for acetylation and 0.9073 for N-glycosylation, with motifs in the MS-identified peptides matching those in the predicted set.

#Applications

The immediate use is proteome-scale PTM annotation in crops where the experimental record is thin: running a wheat or soybean proteome through the code and getting confidence-ranked candidate sites back, either to annotate function directly or to prioritise which proteins are worth the cost of an enrichment-and-MS experiment. That scale means a local install; the free web server is sized for smaller queries, capping a submission at 100 sequences for a single modification type and 50 when several are selected, for pasted FASTA and uploaded files alike. Because the model scores individual residues, it also supports variant work — comparing a wild-type and mutant sequence to see whether a substitution creates or destroys a modification site is a cheap first screen for functional alleles in a breeding program. The GPA1 case study shows the downstream path, with predicted phosphorylation sites feeding a structure predictor.

#Impact

PlantPTM fills a real gap for plant proteomics, where researchers have largely had to borrow predictors calibrated on mammalian data. Its release profile is unusually complete: the Apache-2.0 GitHub repository ships the source code, all nine per-type checkpoints and the redundancy-reduced sequence and PSSM datasets together under one licence, with the project server mirroring the same downloads and a free web service on top. The constraints are worth stating plainly. Coverage stops at nine modification types and six species, with the authors describing further types, further species and the integration of structural features as future work. Each PTM type has its own model rather than one multi-task network, and generating PSSM features requires a local BLAST installation and a UniRef50 database — the practical bottleneck for anyone running the code rather than the web server. Apart from the rice mass-spectrometry experiment, the evaluation is computational.