Is-Pero Dataset: Training and Validation Data for Peroxisomal Protein Prediction

Description This dataset contains all training, validation, and test data used in the development and evaluation of Is-Pero, a comprehensive web server for end-to-end prediction of peroxisomal protein localization from raw FASTA sequences. The dataset supports reproducibility of results presented in the accompanying manuscript and includes data for three integrated prediction modules: Is-Pero (Binary Classification): Distinguishes peroxisomal from non-peroxisomal proteins using logistic regression trained on 1,674 experimentally validated peroxisomal proteins and 133,761 non-peroxisomal proteins from SwissProt (release 2024-01). Includes 10-fold cross-validation (CV) splits and independent test sets for comparison with DeepLoc2.0. In-Pero (Sub-compartment Prediction): Predicts localization within peroxisomes (matrix vs membrane) using a multi-layer perceptron trained on 160 peroxisomal proteins with experimentally validated sub-compartment annotations. Includes 5-fold cross-validation splits and a blind test set of 143 proteins. Is-mPTS (PEX19 Binding Site Prediction): Identifies membrane-targeting signals recognized by the PEX19 chaperone using logistic regression trained on 82 balanced peptide sequences from human and yeast proteins. Includes 5-fold cross-validation and external validation on experimentally validated PEX19 binding sites from organisms not represented in training data. Contents Dataset is-pero-seqs.tar.gz contains FASTA files with protein sequences or peptide windows for each module and benchmarking. Cross-validation fold assignments or test set designation are available in .tsv files.Is-Pero and Benchmarking with DeepLoc 2.0: CV DeepLoc 2.0 - deeploc-cv.tsv CV Is-Pero - is-pero-cv.tsv Blind Test - deeploc-test.tsv In-Pero: CV - in-pero-cv.tsv Test - in-pero-blind.tsv Is-mPTS: CV - is-mpts-cv.tsv Test - is-mpts-blind.tsv Prediction Dataset Test Positives Negatives Total Peroxisomal vs not-Peroxisomal is-pero-cv cross-validation 1,674 133,761 135,435 deeploc-cv cross-validation 304 27,666 27,970 deeploc-blind blind test 216 2,160 2,376 Matrix vs Membrane in-pero-cv cross-validation 28 132 160 in-pero-blind blind test 36 22 58 mPTS vs not-mPTS is-mpts-cv cross-validation 41(23) 41 82 is-mpts-test blind test 13(9) 13 26 Table 1: Dataset composition and validation strategy for Is-Pero modules. Is-Pero uses ESM-2 1280-dimensional full-sequence embeddings; In-Pero and Is-Pero use the same embeddings. Is-mPTS extracts and averages local 15-amino acid window representations from per-residue embeddings. LR: logistic regression; MLP: multilayer perceptron; CV: cross-validation. ∗Experimentally validated. For mPTS datasets, the value outside parentheses refers to the number of mPTS-binding sequence segments, while the value in parentheses indicates the number of original protein sequences from which those segments were extracted. Usage These data enable:1. Complete reproducibility of Is-Pero model training and evaluation metrics2. Benchmarking of alternative methods on the same datasets3. Extension of models with additional sequences or features4. Development of complementary peroxisomal protein prediction tools Redundancy Control All datasets were constructed with explicit redundancy control by clustering sequences with MMseqs at 20% sequence identity, thereby reducing redundancy, limiting information leakage during cross-validation, and yielding more robust estimates of generalization performance. Related Resources - Is-Pero Web Server: https://structure.biofold.org/is-pero- Source Code: https://github.com/biofold/is-pero- Docker Container: https://hub.docker.com/r/biofold/is-pero Citation If you use this dataset, please cite the accompanying manuscript:[] License [CC-BY-4.0]