HydrAMP: a deep generative model for antimicrobial peptide discovery

HydrAMP is a machine learning model for generating new antimicrobial peptides. There are essentially two ways you can work with the model through this web service.

Analogue generation: you can provide a list of sequences that will be treated as starting points for analogue discovery. The model will aim at improving on the starting sequences. While the resulting sequences will be similar to the input, they might show higher antimicrobial activity.

Unconstrained generation: you can freely sample 100 active or non-active peptides, without predefined starting sequences. As our model has knowledge of typical AMP structure and discriminates between active and non-active peptides, it is capable of suggesting novel antimicrobial peptides.


Filter out:








Filter out:



HydrAMP is a deep learning model for generating antimicrobial peptides. The goal of this website was to provide a method for AMP generation in silico as a convenient web service. During training, HydrAMP acquires knowledge about existing antimicrobial peptides and their MIC measurements and therefore has the capability of suggesting promising candidates.

Currently, the HydrAMP model offers two generation modes: analogue and unconstrained:

  • The analogue generation enables targeted search of analogues of existing peptides. The creativity of the model with respect to the input peptide can be controlled via the temperature parameter. The higher the temperature parameter is, the more likely HydrAMP is to introduce modifications (additions, substitutions, deletions) and hence the output peptides will be more diverse. This comes at the cost of creating increasingly dissimilar peptides from their analogue counterpart.
  • The unconstrained generation allows sampling of novel, diverse peptides from the latent space, without predefined analogues.

HydrAMP model was developed as a modification of conditional variational autoencoder, where two additional classifiers were introduced to help the decoder generate peptides with target characteristics. We are interested in peptides with high antimicrobial activity, therefore we focus on P(AMP) - the probability of a peptide being AMP and P(low MIC) - the probability of the peptide being highly active. Apart from extending the architecture, gradient optimizations were applied to ensure a disentanglement of predicted class c’ from latent vector z. For further information, please refer to the figure and to the original paper.

The model was trained on sequences with length up to 25 amino acids, therefore it is presently capable of accepting and generating sequences no longer than 25 amino acids.

Setting the seed guarantees reproducible generation results with the same seed value.

The output table for each mode comprises of multiple physicochemical features calculated for each generated peptide, as well as P(AMP) and P(low MIC) . We enable downloading the results in the .csv format for your convenience.

HydraAMP was developed at the Faculty of Mathematics, Informatics and Mechanics at the University of Warsaw in Dr. Ewa Szczurek's lab. You can visit the group here:


If you have any questions regarding this web service or you need assistance setting up the inference engine locally, do not hesitate to email us at:


you can also create a new issue on GitHub, whichever feels more appropriate

Web development by: Jacek Sroka

While best efforts have been made to ensure the integrity of this service, we take no reponsibility for damages that may result from its use.
Please do not upload sensitive material.