Loading…
Plant DNA-binding protein prediction that averages a ProtT5 sequence-embedding classifier with a SaProt structure-aware one at the score level.
No providers recorded yet. Browse all providers
Whether a protein binds DNA is often a question about geometry: a recognition helix inserted into the major groove, a zinc finger holding a coordinated fold, a leucine zipper dimerizing before it ever touches a promoter. DNA-binding protein (DBP) predictors have historically had to infer that geometry indirectly, from position-specific scoring matrices and physicochemical descriptors computed on the sequence. PLM-DBPs instead pairs a sequence-trained protein language model with a structure-aware one and lets each cast an independent vote.
The model is built for plants specifically, where expanded, lineage-restricted transcription factor families are poorly served by predictors trained on general cross-species datasets. PLM-DBPs runs a protein through ProtT5-XL-U50, from the ProtTrans suite, to get a 1,024-dimensional mean-pooled representation, and in parallel through SaProt 650M, whose vocabulary pairs each amino acid with a Foldseek 3Di structural token, to get a 1,280-dimensional one. Each vector feeds its own small multi-layer perceptron head, and the two sigmoid probabilities are averaged with equal weights and thresholded at 0.5. Both backbones stay frozen; only the heads are trained.
It was developed by Suresh Pokharel, Kepha Barasa, Pawel Pratyush, and Dukka B. KC at the Rochester Institute of Technology and Michigan Technological University, and published in Briefings in Bioinformatics in 2025. The published model adds the structure-aware branch that the earlier preprint's ProtT5-only classifier lacked.
Training used the PlDBPred plant dataset — 849 DNA-binding proteins balanced against an equal number of non-binders — with an independent test set of 997 sequences held out. The ProtT5 head stacks hidden layers of 512, 512, 128, and 16 units; the SaProt head uses 512, 256, and 16. Both apply ReLU activations, dropout between 0.2 and 0.4, and a single sigmoid output. Backbone selection was itself an ablation: Ankh in Base and Large sizes, ESM-2 at 650M, 3B, and 15B, and ProtT5 were each paired with SVM, random forest, feed-forward, and 1D-CNN heads under five-fold cross-validation, and ProtT5 with a feed-forward head gave the best balance.
On the independent test set the fused model reached 84.7 MCC, 92.3% accuracy, 89.8% sensitivity, 95.1% specificity, 97.5 AUROC, and 97.3 AUPR, against 76.4 MCC for PlDBPred, the SVM-on-PSSM tool that supplied the dataset, and 81.7 for ESM-DBP on a leakage-corrected subset. A second test set of 40 plant proteins annotated in UniProt after the training data was compiled gave 85.9 MCC. Inference on an eight-core Xeon takes roughly 89 to 218 seconds per ten sequences as length grows from 200 to 1,000 residues, dominated by embedding extraction.
The practical use is annotating plant proteomes where experimental DNA-binding evidence is thin. A plant-tuned classifier gives crop and model-species groups a way to shortlist candidate regulators for stress-response, development, or trait-engineering studies before committing to ChIP or EMSA work. Because the pipeline needs only a FASTA file and runs on commodity CPUs, it fits into ordinary genome-annotation workflows rather than requiring accelerator time.
PLM-DBPs is a narrow binary classifier, not a general-purpose protein model, and its lasting contribution is the measurement it reports: on a small functional classification task, a structure-aware protein language model carries information a sequence-only one does not, and the cheapest possible fusion extracts it better than learned cross-modal attention does. That result generalizes past DNA binding to any setting where a few thousand labeled proteins must be classified. The code is MIT licensed and the plant training and test FASTA files ship in the repository, but the trained classifier heads, the SaProt configuration, and the 3Di prediction checkpoint are distributed through a Google Drive folder rather than the git tree or a model hub, the full training script has not been released, and the announced web server is not yet live.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.