CBDPF

A phosphosite is a switch with two directions, and the two are not equally documented. Kinases add the phosphate and phosphatases strip it off, but mass spectrometry reports phosphorylated residues in bulk while the residues a phosphatase actually acts on are annotated far more sparsely. Sequence-based predictors of dephosphorylation therefore work from a much thinner labelled corpus than their phosphorylation counterparts, and the classical response — hand-built composition and motif descriptors fed to a shallow classifier — captures little of the sequence context that governs phosphatase specificity.

CBDPF, the Convolutional Bobcat Dephosphorylation Prediction Framework, answers that scarcity by moving the representational burden off the task corpus. Rather than learning what a protein looks like from a sparse labelled corpus, it borrows a general-purpose protein language model that already encodes that, and spends the labelled data only on a small adapter and a classification head. It was built by Anurag Singh, P. K. Singh and Rohit Kumar Tiwari in the Department of Computer Science and Engineering at Madan Mohan Malaviya University of Technology in Gorakhpur, India, and published in Journal of Biomolecular Structure and Dynamics in April 2026.

Architecturally it belongs to the same class as PhosF3C, which applies Low-Rank Adaptation to an ESM-2 encoder for the forward, phosphorylation half of the cycle. CBDPF takes the equivalent recipe to ProtT5 — the T5-family encoder from the ProtTrans suite — and points it at the reverse reaction, which has attracted far less modelling attention.

#Key Features

• Adapted encoder, not frozen embeddings: LoRA trains a small set of low-rank matrices inside the ProtT5 encoder while the pretrained weights stay frozen, so the representation shifts toward dephosphorylation discrimination instead of being read off unchanged.
• Residue-centred sliding windows: Each candidate serine, threonine or tyrosine becomes its own segment, centred on the residue under test with flanking sequence on both sides, giving the encoder local context without requiring the whole protein per call.
• Metaheuristic-tuned convolutional head: The convolutional network reading the contextual embeddings has its configuration searched by a bobcat-hunting population metaheuristic during training — the "Bobcat" in the name is a training-time optimiser, not an inference-time step.
• Fixed checkpoint over arbitrary sequences: Training happens once against a dephosphorylation-site corpus; scoring a new protein means running the sliding window through the adapted encoder, with no per-protein or per-dataset refitting.
• Binary site calls at phosphorylatable residues: The output is a site / non-site decision at S, T and Y positions, the formulation this literature has standardised on.

#Technical Details

The representation layer is ProtT5, a roughly 3-billion-parameter T5 encoder pretrained on unaligned protein sequence at UniRef scale, which supplies a contextual vector for every residue in a window. Parameter-efficient fine-tuning through LoRA leaves those pretrained weights untouched and learns only the injected low-rank updates, which makes adapting an encoder of this size tractable on commodity hardware and keeps the small task corpus from overwriting general protein knowledge. On top of the embeddings sits the convolutional classifier, whose architecture and hyperparameters are selected by the bobcat metaheuristic rather than by hand or grid search. The authors report that CBDPF improves on existing dephosphorylation-site methods in accuracy, stability across runs, and generalisation to held-out sequence. Neither source code, trained weights, nor the assembled training and evaluation data have been released. The article's data availability statement reads, in full, "Data sharing not applicable to this article as no datasets were generated or analysed during the current study" — the only availability signal the article gives, and one hard to reconcile with a paper that trains and benchmarks a classifier on a dephosphorylation-site corpus. The window size, the LoRA rank, the source database for the sequences and the reported accuracy, MCC and AUC figures all sit behind the paywall and are not restated here.

#Applications

The practical use is proteome-scale triage. Phosphatase substrate specificity is poorly mapped relative to kinase specificity, and a sequence-only scorer lets a signalling group rank which annotated phosphosites in a protein of interest are plausible phosphatase targets before committing to mutagenesis or targeted mass spectrometry — most usefully where phosphoproteomics has established that a residue is phosphorylated but nothing establishes what removes the mark. Groups reconstructing signalling cascades and annotating post-translational modification sites across a newly assembled proteome are the intended beneficiaries.

#Impact

CBDPF's contribution is less a new architecture than a demonstration of economics: a task with a sparse labelled corpus can now draw on a multi-billion-parameter protein encoder, because parameter-efficient fine-tuning has made the adaptation cheap. That pattern is spreading quickly across post-translational modification prediction, and the reverse reaction tests whether it holds where data is thinnest. Two caveats bound how far the result travels. The model is single-task and binary, with no zero-shot or cross-task transfer demonstrated, so it is a narrow probe on a general backbone rather than a foundation model in its own right. And with no code, weights or data published and the article behind a subscription paywall, the reported gains cannot currently be reproduced outside the group. The same three authors published a second, architecturally distinct dephosphorylation model — TLCBM-LoRA, which pairs LoRA with a termite-life-cycle Boltzmann machine rather than a convolutional head — in the Journal of Molecular Modeling within the same window; the two report separate benchmark results, and no figure quoted for one belongs to the other.