About AMPidentifier

Antimicrobial resistance kills people now and will kill more. The sequence space where new antimicrobial peptides might be found is far larger than any laboratory can screen, which makes computational triage part of the experimental pipeline rather than an accessory to it. A tool that performs that triage is only useful to the extent that other people can run it, check it and disagree with it.

That is why every part of this project is in the open. The training code, the datasets, the feature extraction, the model files and the web server are in one public repository under a permissive licence. The benchmark figures on this page come from a held-out set that ships with the repository, so anyone can reproduce them or show that they do not hold. A prediction you cannot audit is an opinion with a decimal point.

The FAIR principles are the practical form of that commitment. Findable: the software is deposited on PyPI and GitHub, carries a citable identifier and a machine-readable CITATION.cff. Accessible: the web tool needs no account, no institutional login and no payment, and the package installs with one command. Interoperable: input is FASTA and output is CSV, two formats every other tool in the pipeline already reads. Reusable: the licence permits reuse and modification, the descriptors are documented, and the training procedure is a script rather than a paragraph in a methods section.

Two commitments follow from this and constrain how the tool is built. Sequences submitted here are processed in memory and are never stored: unpublished sequence data is the most valuable thing a research group holds, and asking someone to upload it to a server that keeps it is asking too much. And the tool stays free to run, because a triage step behind a paywall stops being triage for exactly the laboratories that most need it.

Open development is not a licence file. It is answering issues, publishing the failure cases along with the benchmark, and treating a report that the model is wrong as the most useful message the project receives.

Bioinformatics web tools stop working. Kern, Fehlmann and Keller (2020, doi:10.1093/nar/gkaa1125) monitored 2,396 tools published from 2010 onward over 133 days and found 25.7% unreachable at first access. Availability tracks age almost linearly: tools published in 2019 and 2020 were around 90% available, those from 2010 around 50%. When the authors of 47 broken recent tools were contacted, 51.1% were restored, which means half of those failures were not technical problems but abandoned maintenance.

The same pattern appeared inside this project. Building the external benchmark, 20 published antimicrobial peptide predictors were identified as candidates. Nine of them, 45%, could not be evaluated at all: web servers unreachable, DNS resolution failing, or no public code release to run locally. Their papers span 2012 to 2023, and the check was made in March 2026. The remaining 11 configurations were benchmarked against the same independent set of 4,736 sequences.

A tool that exists only as a web server dies with the server. This one is distributed three ways on purpose: the web page, a command line program and a Python package on PyPI, with the training code, the datasets and the model files in the repository. If this page goes down, pip install ampidentifier still reproduces every number printed here, and the repository can be forked by anyone who wants to keep it alive.

Reference: Kern, F., Fehlmann, T. & Keller, A. (2020). On the lifetime of bioinformatics web services. Nucleic Acids Research 48(22), 12523–12533. doi:10.1093/nar/gkaa1125

The default mode is a soft-voting ensemble of five classifiers: random forest, support vector machine, gradient boosting, XGBoost and LightGBM. Each sequence is reduced to 22 physicochemical and compositional descriptors, among them net charge, hydrophobicity, hydrophobic moment, isoelectric point, aliphatic index, instability index and amino acid composition.

On the independent benchmark set of 4,736 sequences the ensemble reaches AUC-ROC 0.950, MCC 0.742, sensitivity 94.9% and specificity 78.4%. Single-model modes are available and score lower on every one of those four.

Training data comes from experimentally validated antimicrobial peptides in public databases, against non-antimicrobial sequences. The coverage includes antibacterial, antifungal, antiviral and other host defence peptides. A protein whose primary function is not antimicrobial can still carry antimicrobial features in specific regions of its sequence, so a positive call is a hypothesis for testing, not a measurement.

AMPidentifier began as a minimum viable product written by João Pacifico Bezerra Neto in the 2020s. That prototype is where the tool comes from: it set the problem, showed that sequence-derived descriptors carried enough signal to classify antimicrobial peptides, and made the case for building the rest.

What is distributed today was rebuilt from that starting point. The descriptor set, the training pipeline, the ensemble, the external benchmark and the three distribution channels are a later and separate implementation, but the idea is his.

Python package: pip install ampidentifier

Source and command line: github.com/madsondeluna/AMPidentifier

Web: the predictor and what is coming next

Sequences are processed in memory and are not stored. Transfer is encrypted with HTTPS/TLS. No account is required and no sequence data is shared with third parties.

de Luna-Aragão, M. A., da Silva, R. L., Pacifico Bezerra Neto, J., dos Santos-Silva, C. A., da Silva Santos, D. E. & Benko‑Iseppon, A. M. (2026). AMPidentifier: A Cross-Platform Ensemble Toolkit for Antimicrobial Peptide Prediction.

Registered with the INPI – Instituto Nacional da Propriedade Industrial under No. BR 51 2025 005859-4, property of the Universidade Federal de Pernambuco (UFPE) and the Laboratório de Genética e Biotecnologia Vegetal (LGBV).