SenSeqNet

Cellular senescence — the irreversible arrest of proliferation that follows stress or damage — accumulates with age and drives much of the tissue dysfunction behind age-related disease. Identifying which proteins participate in senescence programs is therefore a prerequisite for finding senolytic targets, but the standard readouts are cell-based assays: slow, labour-intensive, and impossible to run across a proteome. SenSeqNet addresses the gap by predicting senescence association directly from an amino acid sequence.

The model was built by a group spanning Chongqing Medical University and its affiliated hospitals, the University of Toronto, the Ontario Institute for Cancer Research, the University of Wisconsin–Madison, and the University of Illinois Urbana-Champaign. It follows the now-common pattern of a task-specific head on a frozen protein language model: ESM-2 supplies the representation, and a hybrid bidirectional LSTM plus convolutional network does the classification. The design choice the paper argues for is the ordering — sequential modelling first, spatial refinement second — which outperforms the reverse arrangement by a wide margin.

#Key Features

• Sequence-only senescence prediction: A FASTA file in, a per-sequence senescence call with a confidence score out, which makes it usable on newly discovered or poorly characterized proteins that have no experimental annotation.
• Language model representations over handcrafted features: t-SNE projections of the embeddings form clean class-aligned clusters, whereas composition-based descriptors (ACC, CKSAAP, DDE) leave the two classes heavily overlapped.
• Hybrid recurrent-convolutional head: A four-layer bidirectional LSTM captures long-range dependencies along the sequence, and three convolutional blocks then sharpen the spatial structure of those features before a fully connected output layer.
• Biologically coherent predictions: Genes the model classifies with high confidence are significantly enriched for senescence and autophagy in cancer, negative regulation of proliferation, and response to oxygen levels — evidence it has learned program-level signal rather than label artefacts.
• Packaged for reuse: The trained checkpoint is distributed inside a PyPI package with a senseqnet-predict command-line entry point, alongside a Colab notebook.

#Technical Details

Positive examples derive from 210 senescence-associated genes curated from CellAge, restricted to those experimentally shown to induce senescence, plus in vivo signature genes; negatives come from 148 genes covering cytoskeletal, housekeeping, ribosomal, proliferative, apoptotic, metabolic and receptor-signalling roles. All UniProtKB isoforms were collected, giving 292,759 positive and 373,098 negative sequences, then reduced with MMseqs2 at a 0.3 identity threshold to 43,381 and 32,673 representative sequences split 80/20. A model-selection sweep across ESM-2 sizes settled on the 650M-parameter esm2_t33_650M_UR50D variant (1,280-dimensional embeddings, 86.43% accuracy) over the 8M, 35M and 150M variants at 79.15%, 82.26% and 84.16%. The classifier is a four-layer bidirectional LSTM with 181 hidden units feeding three convolutional blocks of 76, 111 and 487 channels with max pooling. On the independent test set SenSeqNet reaches 86.43% accuracy, 90.51% sensitivity, 81.01% specificity, 88.38% F1, MCC 0.7221 and AUC 0.94, ahead of standalone CNN, RNN, LSTM and BiLSTM heads (78–82%) and of random forest, XGBoost, SVM and logistic regression baselines (all under 80%). Swapping the backbone for ESM-1b or ESM-1v costs a few points; ProtTrans T5 and ProteinBERT collapse to near-random.

#Applications

The practical use is triage. A researcher studying an aging phenotype can score every protein in a candidate list, or an entire proteome, and rank sequences for follow-up in culture — the same job the model was tested on when it classified 52,227 sequences from 26 held-out senescence genes with a 77.55% true-positive rate. Because the checkpoint is fixed and the interface is a single command, it slots into an existing annotation pipeline without training infrastructure.

#Impact

SenSeqNet, published in Aging Cell after a bioRxiv preprint, gives the senescence field a concrete sequence-level classifier where previously there was largely curated gene-list lookup. Its scope is genuinely narrow: it is a binary classifier over a small, curated gene universe, and the paper's own per-gene analysis shows systematic failures on ribosomal proteins such as RPL21, RPL26 and RPL28, whose sequence features confound the senescence signal. Predictions are best read as hypotheses for experimental follow-up rather than annotations. The repository ships code, data and weights but carries no license file, which limits reuse in downstream software.