AMPTransformer

AMPTransformer classifies peptides as antimicrobial/non-antimicrobial. It uses fine-tuned protein NLP models in combination with physicochemical descriptors to classify a FASTA file of peptide sequences. See example.fasta for an example of this file format. A more detailed description of the model is provided in the Description section, below the usage instructions.

Dependencies

• python >= 3.9
• numpy 1.21.6
• pandas 1.5.0
• torch 1.13.1
• transformers 4.26.1
• tokenizers 0.13.2
• biopython 1.81
• peptides 0.3.1
• propy3 1.1.1
• scikit-learn 1.1.2
• catboost 1.1.1
• lightgbm 3.3.5
• xgboost 1.6.2
• autogluon 0.7.0
• joblib 1.2.0
• tqdm 4.65.0

Download NLP Models

Before installation, it is necessary to download the pretrained ProtBERT (1) and ESM_V2 (2) protein NLP models. They can be downloaded from the following Google Drive:

https://drive.google.com/drive/folders/1HFZvBG0VWW2kO6uJqLSIFCs_TISwjGa-?usp=share_link

You can add a shortcut to the folder in your Google Drive by right-clicking on the "nlp_models" folder title, and selecting "Add Shortcut to Drive".

Note: There are zipped and unzipped versions of the files. The zipped versions will need to be unzipped prior to use.

The files can also be found as a dataset on kaggle.

https://www.kaggle.com/datasets/brendanmoore14/amptransformer-models

Google Colaboratory Notebook

The predictor can be run using the following Google Colaboratory notebook:

https://colab.research.google.com/drive/1HE0n7tk53NVMpdCURF_JFmIsNESn2g-v?usp=sharing

Please make a copy of the notebook so that your changes can be saved.

Installation Guide

1. Clone the GitHub repository.

git clone https://github.com/Brendan-P-Moore/AMPTransformer

2. Set the current working directory to the AMPTransformer folder.

cd AMPTransformer

3. Install the required packages listed in the dependencies section and requirements file.

pip install -r requirements.txt

4. Import the AMPTransformer python file, and the predict function.

import AMPTransformer
from AMPTransformer import predict

Predict

To predict the antimicrobial nature of the peptide sequences contained within a fasta file, we can run the following function. The path to the fasta file and model folders must be the absolute path.

Peptide sequences should be between 5 and 100 amino acids in length. The ES

predict(path_to_the_fasta_file, path_to_the_protbert_file_folder, path_to_the_esm_file_folder)

An example from the Google Colaboratory notebook:

predict('/content/AMPTransformer/test_data/CAMP_test_neg.fasta', '/content/drive/MyDrive/nlp_models/protbert_models', '/content/drive/MyDrive/nlp_models/esm_models')

The output of this prediction is a dataframe with three columns. The first column is the peptide label in the fasta file, the second is the sequence, and the third is the antimicrobial prediction. The prediction is a number between 0 and 1, with values above 0.5 being antimicrobial, and those below 0.5 being non-antimicrobial.

prediction_dataframe = predict('/content/AMPTransformer/test_data/CAMP_test_neg.fasta', '/content/drive/MyDrive/nlp_models/protbert_models', '/content/drive/MyDrive/nlp_models/esm_models')

For prediction of many peptide sequences, it is recommended to use a GPU.

Description and Evaluation

AMPTransformer is an antimicrobial peptide classifier trained on the length-balanced training datasets used by amPEPpy (3,4). These datasets were selected because of the equal peptide sequence length distribution for the antimicrobial and non-antimicrobial peptides (3,4). Because of this, AMPTransformer is more useful at discerning between antimicrobial/non-antimicrobial peptides of equal length. Other predictors should be used if the peptides of interest have large variation in sequence length.

AMPTransformer was evaluated on eight non-redundant test datasets published by Yan et al., and details of the test sets are described therein (5,6). The predictor's performance compared with amPEPpy, a random forest model trained on the same training dataset, is shown in the table below (3). As the test sets are not length-balanced, and are non-redundant only for amPEPpyy, the performance of the AMPTransformer on the independent test sets cannot be compared directly with other methods.

Overall, AMPTransformer has similar accuracy and matthews correlation coefficients to ampPEPpy, better overall area under the curve and precision, and worse F1 score and recall. This varies by test data set, however, with AMPTransformer performing better relative to amPEPpy on test datasets with more length-balanced datasets such as XUAMP, dbAMP, and LAMP. The distribution of lengths for positive and negative samples in the test datasets can be found in the supplementary material of Yan et al. (5).

Test datasets
XUAMP	Accuracy	F1 score	Precision	Recall	MCC	AUC		Number of peptides:	3072
amPEPpy	0.636	0.523	0.760	0.398	0.310	0.714
AMPTransformer	0.635	0.502	0.790	0.368	0.320	0.743
DBAASP								Number of proteins:	356
amPEPpy	0.767	0.743	0.828	0.674	0.543	0.870
AMPTransformer	0.744	0.700	0.848	0.596	0.512	0.858
dbAMP								Number of peptides:	440
amPEPpy	0.802	0.774	0.903	0.677	0.624	0.878
AMPTransformer	0.834	0.820	0.897	0.755	0.677	0.922
LAMP								Number of peptides:	1750
amPEPpy	0.728	0.662	0.874	0.533	0.495	0.826
AMPTransformer	0.750	0.700	0.878	0.582	0.532	0.865
DRAMP								Number of peptides:	2364
amPEPpy	0.690	0.597	0.852	0.459	0.428	0.720
AMPTransformer	0.665	0.537	0.868	0.388	0.395	0.718
CAMP								Number of peptides:	44
amPEPpy	0.841	0.829	0.895	0.773	0.688	0.856
AMPTransformer	0.773	0.737	0.875	0.636	0.567	0.886
APD3								Number of peptides:	354
amPEPpy	0.884	0.881	0.905	0.859	0.769	0.924
AMPTransformer	0.864	0.860	0.886	0.836	0.730	0.927
YADAMP								Number of peptides:	226
amPEPpy	0.894	0.889	0.932	0.850	0.791	0.950
AMPTransformer	0.881	0.873	0.930	0.823	0.766	0.957
Weighted Overall Performance								Total peptides:	8606
amPEPpy	0.702	0.619	0.830	0.501	0.439	0.769
AMPTransformer	0.698	0.602	0.845	0.479	0.440	0.788

References

1. Elnaggar, A. et al., ProtTrans: Toward Understanding the Language of Life Through Self-Supervised Learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 10, pp. 7112-7127, 1 Oct. 2022, doi: 10.1109/TPAMI.2021.3095381.
2. Lin, Z. et al., Evolutionary-scale Prediction of Atomic Level Protein Structure With a Language Model, bioRxiv 2022.07.20.500902; doi: doi: 10.1101/2022.07.20.500902.
3. Lawrence, TJ. et al. amPEPpy 1.0: a Portable and Accurate Antimicrobial Peptide prediction tool. Bioinformatics. 2021 Aug 4;37(14):2058-2060. doi: 10.1093/bioinformatics/btaa917.
4. Bhadra, P. et al. AmPEP: Sequence-based Prediction of Antimicrobial Peptides Using Distribution Patterns of Amino Acid Properties and Random Forest. Sci Rep 8, 1697 (2018). doi: 10.1038/s41598-018-19752-w.
5. Yan, K. et al. sAMPpred-GAT: Prediction of Antimicrobial Peptide by Graph Attention Network and Predicted Peptide Structure. Bioinformatics. 2023 Jan 1;39(1):btac715. doi: 10.1093/bioinformatics/btac715.
6. Xu, J. et al. Comprehensive assessment of machine learning-based methods for predicting antimicrobial peptides, Briefings in Bioinformatics, Volume 22, Issue 5, September 2021, bbab083, doi: 10.1093/bib/bbab083.

FAQ

To be updated as questions arise.