AMPlify is an attentive deep learning model for antimicrobial peptide prediction.

• Python 3.6
• Keras 2.2.4
• Tensorflow 1.12
• Numpy <1.17
• Pandas
• Scikit-learn
• Biopython
• h5py <3

1. Create a new conda environment:

conda create -n amplify python=3.6

2. Activate the environment:

conda activate amplify

3. Install AMPlify in the environment:

conda install -c bioconda amplify

train_amplify and AMPlify can now be run. See usage information below.

4. To deactivate an active environment, use:

conda deactivate

Datasets for training and testing are stored in the data folder. Please specify the directory if you would like to use those datasets for training or testing the model.

• Balanced training set: AMPlify_AMP_train_common.fa + AMPlify_non_AMP_train_balanced.fa
• Balanced test set: AMPlify_AMP_test_common.fa + AMPlify_non_AMP_test_balanced.fa
• Imbalanced training set: AMPlify_AMP_train_common.fa + AMPlify_non_AMP_train_imbalanced.fa
• Imbalanced test set: AMPlify_AMP_test_common.fa + AMPlify_non_AMP_test_imbalanced.fa

AMPlify only accepts the 20 standard amino acids. However, if the user specifies a stop codon by adding an asterisk character to the end of a peptide sequence, AMPlify will ignore the last character and allow the peptide to participate in the training or inference stage.

Weights for the pre-trained balanced/imbalanced sub-models are stored in the models folder.

• Balanced model: The one proposed in the AMPlify publication; Trained on the balanced training set; Suitable to use for predicting putative AMPs from a relatively curated candidate set (e.g., sequences already filtered by length/charge and through sequence homology search).
• Imbalanced model: Trained on the imbalanced training set; Suitable for situations where the number of non-AMPs in the input sequence set is far greater than that of AMPs (e.g., a large transcriptomics sequence database).

Usage: train_amplify [-h] -amp_tr AMP_TR -non_amp_tr NON_AMP_TR [-amp_te AMP_TE] [-non_amp_te NON_AMP_TE] [-sample_ratio {balanced,imbalanced}] -out_dir OUT_DIR -model_name MODEL_NAME

optional arguments:
  -h, --help            Show this help message and exit
  -amp_tr AMP_TR        Training AMP set, fasta file
  -non_amp_tr NON_AMP_TR
                        Training non-AMP set, fasta file
  -amp_te AMP_TE        Test AMP set, fasta file (optional)
  -non_amp_te NON_AMP_TE
                        Test non-AMP set, fasta file (optional)
  -sample_ratio {balanced,imbalanced}
                        Whether the training set is balanced
                        or not (balanced by default, optional)
  -out_dir OUT_DIR      Output directory
  -model_name MODEL_NAME
                        File name of trained model weights

Example: train_amplify -amp_tr ../data/AMPlify_AMP_train_common.fa -non_amp_tr ../data/AMPlify_non_AMP_train_balanced.fa -amp_te ../data/AMPlify_AMP_test_common.fa -non_amp_te ../data/AMPlify_non_AMP_test_balanced.fa -out_dir ../models/ -model_name model

Expected output: 1) The model weights trained using the specified data; 2) Test set performance, if test sequences have been specified.

Usage: AMPlify [-h] [-m {balanced,imbalanced}] -s SEQS [-od OUT_DIR] [-of {txt,tsv}] [-sub {on,off}] [-att {on,off}]

optional arguments:
  -h, --help            Show this help message and exit
  -m {balanced,imbalanced}, --model {balanced,imbalanced}
                        Balanced or imbalanced model (balanced by default, optional)
  -s SEQS, --seqs SEQS  Sequences for prediction, fasta file
  -od OUT_DIR, --out_dir OUT_DIR
                        Output directory (optional)
  -of {txt,tsv}, --out_format {txt,tsv}
                        Output format, txt or tsv (tsv by default, optional)
  -sub {on,off}, --sub_model {on,off}
                        Whether to output sub-model results, on or off (off by
                        default, optional)
  -att {on,off}, --attention {on,off}
                        Whether to output attention scores, on or off (off by
                        default, optional)

Example: AMPlify -s ../data/AMPlify_AMP_test_common.fa

Expected output: Predicted probability scores, AMPlify log scaled scores, and classes of the input sequences. The AMPlify log scaled score is calculated as -10*log10(1-Probability_score). Results for invalid sequences will be filled with NA.

Note: In the default setting, sequences with AMPlify scores > 3.01 (i.e., AMPlify probability scores > 0.5) are predicted as AMPs.

• Additional scripts and data for the AMP discovery pipeline that has been utilized for bullfrog genome mining are provided in auxiliary/amp_discovery_pipeline (added in v1.0.1). Parameters for GMAP and MAKER2 are described in the Methods section of the manuscript.

• Additional scripts and data for the AMP mining workflow that has been utilized to mine novel AMPs from the UniProtKB/Swiss-Prot database are provided in auxiliary/amp_mining_workflow (added in v2.0.0). Detailed methods can be found in the manuscript. The input data and result files are publicly accessible through a Zenodo repository.

Chenkai Li (chenkai.li@ubc.ca)

If you have any questions, comments, or would like to report a bug, please file a Github issue or contact us.

If you use AMPlify in your work, please cite our publications:

1. The research article for AMPlify and the discovery of RaCa series peptides:

Li, C., Sutherland, D., Hammond, S.A. et al. AMPlify: attentive deep learning model for discovery of novel antimicrobial peptides effective against WHO priority pathogens. BMC Genomics 23, 77 (2022). https://doi.org/10.1186/s12864-022-08310-4

2. The data note paper introducing the imbalanced model in addition to the original balanced model:

Li, C., Warren, R.L. & Birol, I. Models and data of AMPlify: a deep learning tool for antimicrobial peptide prediction. BMC Research Notes 16, 11 (2023). https://doi.org/10.1186/s13104-023-06279-1

3. The research article for mining the UniProtKB/Swiss-Prot database for novel AMPs:

Li, C., Sutherland, D., Salehi, A. et al. Mining the UniProtKB/Swiss-Prot database for antimicrobial peptides. Protein Science 34, e70083 (2025). https://doi.org/10.1002/pro.70083