DFN-Kcr

Lysine crotonylation attaches a four-carbon crotonyl group to a lysine side chain — bulkier and more rigid than an acetyl group, and enough to alter the residue's charge and geometry. Most of what is known about the mark comes from histones, where it is tied to spermatogenesis, tissue repair, cancer and neurodegeneration; the non-histone crotonylome is far less charted, and mapping it experimentally means mass spectrometry.

The computational alternative has an awkward statistical problem at its center. In the human non-histone benchmark used here, unmodified lysines outnumber crotonylated ones roughly five to one, and the paper's own t-SNE projections show that in the raw encodings — integer-coded residues and ProteinBERT features alike — the two classes overlap completely, with no visible boundary. Whatever signal exists has to be manufactured above the encoding, so what sits on top of the encoder matters as much as the encoder itself.

DFN-Kcr, from Jiangxi Institute of Fashion Technology, answers that with two networks that read the same 29-residue window differently and a small gate that decides, sample by sample, which one to believe. One branch runs a frozen slice of ProteinBERT through a residual convolutional stack; the other learns its own embedding from integer-coded residues and passes it through a transformer block. Each emits a crotonylation probability, and the gate returns their weighted sum. It sits alongside PhosF3C and StackGlyEmbed as a sequence model for modification sites, on a mark neither covers.

#Key Features

• Frozen pretrained features, not a fine-tune: the convolutional branch consumes a fixed slice of the released ProteinBERT checkpoint — the summed outputs of ProteinBERT's first narrow and wide convolution blocks over its token embedding — computed once and never updated in training.
• A second branch that starts from nothing: the same window is integer-coded, given learnable token and positional embeddings, and passed through one transformer block with two attention heads, yielding a representation independent of the pretrained encoder.
• Branch-level gating, not feature concatenation: fusion acts on the two branch probabilities rather than on token representations, with a tanh layer and a softmax producing the weights.
• Fixed checkpoint, served over the web: trained once on a public benchmark and applied to new sequences without retraining, through a server that scores pasted sequences in the page and mails back the results of a FASTA upload.
• Data shipped with the code: the training and independent-test splits sit in the repository beside the notebook, so the evaluation set is reproducible.

#Technical Details

The benchmark is the human non-histone crotonylation set released by Gao et al. in 2024, built from UniProt, redundancy-reduced with CD-HIT at 30% identity, and cut into 29-residue windows centered on each lysine. Training holds 12,262 positive and 60,101 negative sites; the independent test set holds 3,343 and 15,010. The convolutional branch applies a 1D convolution and max pooling, two residual blocks of 64 filters with kernel size 3, a second pooling step, and a four-layer fully connected head; the transformer branch pairs one block with a 128/64/128 feed-forward stack. Training used AdamW and binary cross-entropy for 150 epochs on one Tesla P100. On the test set DFN-Kcr reports sensitivity 0.8244, specificity 0.7679, accuracy 0.7961 and MCC 0.5932, against 0.5203, 0.5179, 0.4879 and 0.3257 for MVNN-HNHC, nh-Kcr, Adapt-Kcr and Deep-Kcr. Under ten-fold cross-validation the fused model reaches MCC 0.5908 versus 0.5828 for the ProteinBERT branch alone and 0.5526 for the transformer branch alone, so the gain from fusion is real but modest. Fine-tuned ESM-2 at 35M, 150M and 650M parameters all land near MCC 0.45 on the same data. Applying the design to serine/threonine phosphorylation required training a separate model, so what transfers is the architecture, not the checkpoint.

#Applications

The model suits proteomics groups working on non-histone crotonylation who want to rank the lysines in a candidate protein before committing to targeted mass spectrometry. Inference needs only a sequence window, so a whole protein can be scanned residue by residue — on the hosted server with no local install, when that server is reachable — useful for labs studying crotonylation in metabolic enzymes and signaling, or shortlisting sites for mutagenesis.

#Impact

The scope is narrow by design: human non-histone lysines, a fixed 29-residue window, and negatives drawn at roughly five to one from the same protein set, so the model should not be pointed at plant or histone crotonylomes without re-fitting. The trained weights are not distributed: the repository ships data, a notebook and a pointer to ProteinBERT's own weights, so reproducing the model means re-running training. The hosted server is the only route to the authors' checkpoint, and it was answering with a server error when this entry was written, which leaves no working route to the trained model. The repository carries no license, though the article is CC BY 4.0. The durable contribution is methodological — a shallow frozen slice of a protein language model, paired with a branch trained from scratch and combined by a two-weight gate, beats fine-tuning a much larger encoder on this task.