ABCP_finder

Anticancer peptide predictors are trained on positives pooled across every tumour type, so a peptide validated against a leukaemia line and one validated against a breast tumour sit in the same class, and a group working on breast cancer gets back a list in which the signal for their indication has been averaged with everything else. ABCP_finder narrows the label instead of the architecture: its positive class is the peptides in CancerPPD2 with measured activity against breast cancer, and its negative class is built by subtracting those from the pooled anticancer and non-anticancer peptides of ACP-ML — so a general anticancer peptide with no breast-cancer evidence is an explicit negative, not an unlabelled case.

The model itself is a probe rather than a new protein language model. A peptide sequence is whitespace-tokenized, passed through a frozen pretrained encoder, and mean-pooled across residues into a single vector, which a multilayer perceptron classifies as anti-breast cancer peptide or not. Two encoders were compared: ProtBERT, an alignment-informed BERT encoder from the ProtTrans family, and ESM-2 at 650M parameters, which is alignment-free. ProtBERT led on recall, F1, MCC and AUC, and is the backbone the headline results are reported on. The work comes from Indian Institute of Information Technology Allahabad, with collaborators at IIIT Nagpur and Bose Institute.

It sits beside AntiCP3 without overlapping it: that model fine-tunes its encoder end to end over full-length anticancer proteins of 50 to 1,000 residues, whereas ABCP_finder keeps the encoder frozen, works only on peptides under 50 residues, and asks a breast-cancer-specific question.

#Key Features

• Indication-specific positive class: Training positives are peptides with recorded breast cancer activity, so a general anticancer peptide with no breast-cancer evidence counts against the model rather than for it.
• Frozen encoder, trained head: Only the perceptron is fitted; the encoder is a released general-purpose checkpoint, keeping the trainable parameter count small enough for a 1,353-peptide corpus.
• Homology-aware evaluation: Sequences were clustered with CD-HIT at 30% identity and 80% coverage, and whole clusters — never individual sequences — were assigned to folds, so the held-out set is evolutionarily distant from training.
• Calibrated probabilities, not just labels: Calibration analysis supports a 0.7 probability cut-off for treating a call as high confidence, backed by a Brier score of 0.0633.
• Command-line screening tool: A FASTA file in, a CSV of per-sequence class probabilities out, with either backbone or both, and automatic CPU/GPU selection.

#Technical Details

The curated corpus is 301 anti-breast cancer peptides and 1,052 negatives, all under 50 amino acids, with duplicates and non-standard residues removed. CD-HIT produced five non-redundant clusters, partitioned into five folds: four gave 240 positives against 935 negatives for training and the held-out fold gave 61 against 117, with pairwise train-test identity concentrated in the 0–20% band. Embeddings come from ProtBERT (30 layers, 1024 hidden, roughly 420M parameters) and ESM2 (33 layers, 1280 hidden, roughly 650M), mean-pooled rather than read from a CLS token. The classifier is a 512-unit and then 128-unit hidden stack with ReLU, batch normalization and dropout of 0.1, trained with Adam at a learning rate of 0.001, batch size 16, 15 epochs, and hyperparameters chosen by 10-fold grid-search cross-validation. On the held-out fold the ProtBERT model reaches 93.82% accuracy, 97.43% specificity, 94.64% precision, 86.88% recall, 90.59% F1, 0.8618 MCC and 96.67% AUC; the ESM2 model matches on specificity but drops to 81.96% recall, 87.71% F1 and 0.8241 MCC.

#Applications

The intended use is triage ahead of synthesis. The authors ran 3,022,830 unclassified peptide sequences from PeptideAtlas through the model and kept the 256,889 scoring above 0.70, then passed those to xDeep-AcPEP, an independent predictor of anticancer peptide potency. The predicted IC50 values clustered between roughly 1 × 10⁻⁴ M and 6 × 10⁻⁴ M with a median near 4.4 × 10⁻⁴ M, rather than scattering across the concentration range — an in-silico cross-check that the retained peptides are enriched for activity. That turns an intractable library into a shortlist a peptide therapeutics group can actually order.

#Impact

ABCP_finder shows how far a thin supervised head on a frozen protein language model can go when experimental labels are scarce: a 1,353-sequence corpus is far too small to train a sequence model from scratch, but enough to fit a classifier over representations someone else paid to pretrain. The authors state the limits plainly. The dataset is small and imbalanced, and much of it derives from assays on particular breast cancer cell lines, so the model may be learning line-specific rather than subtype-general signal. The output is binary — no potency, mechanism of action or subtype specificity — and the external validation is itself a prediction, not a wet-lab measurement. Whatever biases the backbones carry are inherited wholesale, and the work is a preprint that has not been peer reviewed.