DLFea4AMPGen

August 1, 2025 · View on GitHub

By fine-tuning the pre-trained model MP-BERT, we constructed ABP-MPB model for identifying antibacterial peptides, AFP-MPB model for identifying antifungal peptides, and AOP-MPB model for identifying antioxidant peptides.

Based on this, a de novo design strategy for multifunctional antimicrobial peptides was developed.

Installation

Huawei Atlas Server (Linux, with Huawei Ascend 910 NPU)



The model construction process uses the MindSpore framework. Installation guide for installing Ascend MindSpore on different hardware platforms can be found at: https://www.mindspore.cn/install/.

Here, we recommend to use Docker, an open source application container engine that allows developers to package their applications, as well as dependency packages, into a lightweight, portable container.

By using Docker, rapid deployment of MindSpore can be achieved and isolated from the system environment.

Finetune Task

The overall architecture of the model is shown in the figure below：

Model Structure

The pre-trained model was published in our previous study, and it was trained using publicly available unlabeled pure sequence protein sequences from UniProt.

Therefore, we can directly fine-tune this pre-trained model to construct the desired prediction model.

We have fine-tuned the framework for classification to create three bioactive models: ABP-MPB, AFP-MPB, and AOP-MPB. The details are as follows:

modelfunction
ABP-MPBAntibacterial activity identification
AFP-MPBAntifungal activity identification
AOP-MPBAntioxidant activity identification

Prediction Task

If you want to identify bioactive peptides using an existing model, you can directly use one of our fine-tuned models without the need to construct a new prediction model from scratch.

Data Preparation

A total of two columns of data are required in CSV file. The first column is the ID number, with the column name "id"; the second column is the sequence information, with the column name "seq". As shown in the table below：

idseq
id1seq1
id2seq2
id3seq3

Sequence Generation

Once the pre-built model is ready and the environment is properly configured, follow these steps sequentially to generate the sequences from scratch.

Important Environment Recommendation

To maximize performance for both model training and SHAP value extraction, we recommend a two-environment setup:

  1. Train and fine-tune your model inside a Docker environment with Ascend NPU support — ideal for efficient model training.
  2. Sequence Generation in a separate environment outside the Docker container — starting from SHAP value Extraction — using either GPU or CPU resources (not NPU-based).

SHAP value extraction

Run the following command to extract SHAP values. A total of three files are generated: seq.txt, base_value.txt, and SHAP_value.txt. The SHAP_value.txt file contains the SHAP values extracted by the model for each amino acid, with each amino acid assigned a corresponding SHAP value. This file is the primary focus of our analysis.

python SHAP_value.py --model <Finetuned model> --input_file <sequence.csv> --output_path <output_path>

Parameters:

  • --model: Path to the fine-tuned BERT model checkpoint (e.g., .ckpt file).
  • --input_file: A CSV file with a column named seq, where each row is an amino acid sequence.
  • --output_path: Directory where output files (SHAP_value.txt, base_value.txt, seq.txt) will be saved.

Sliding window

The sliding window process is illustrated in the figure below：

Model Structure

If you are generating single-activity peptides, you can skip this normalization step. Normalize SHAP values across multiple models to mitigate differences in magnitude, enabling fair comparison or integration:

python shap_value_normalization.py --input_files <file1.txt> <file2.txt> <file3.txt> --output_path <output_path>

Parameters:

  • --input_files: A list of one or more text files containing SHAP values. Each line in a file should represent the SHAP values for a single sequence, with one SHAP value per amino acid.
  • --output_path: Directory where the results will be saved.

Set a sliding window of 13 amino acids and retain the window with the highest average SHAP value, provided that it is greater than 0, for each sequence：

python sliding_windows.py --SHAP_file <SHAP_value.txt> --seq_file <sequence.csv> --length <length> --output_path <output_path>

Parameters:

  • --SHAP_file: Path to the file containing SHAP values (e.g., SHAP_value.txt) where each row corresponds to one sequence, and each value corresponds to one amino acid.
  • --seq_file: The CSV file containing the original sequences (seq column required), which should match the SHAP values by order.
  • --length: Length of the sliding window to apply (default is 13 amino acids).
  • --output_path: Directory where the extracted motifs or windowed sequences will be saved.

Phylogenetic tree

Build the phylogenetic tree with the following command:

bash tree.sh <motif.fasta> <tree.newick>

Parameters:

  • <motif.fasta>: Input FASTA file containing sequences to be used for tree construction.
  • <tree.newick>: Output file name for the generated phylogenetic tree in Newick format.

Systematic shuffle

Systematically generate all possible combinations of key high-frequency amino acids for each group to obtain a set of candidate sequences:

python systematic_shuffle.py --input_file <amino_acids.csv> --length <motif_length> --output_path <output_path>

Parameters:

  • --input_file <amino_acids.csv>: CSV file listing key amino acids for each position to be shuffled and combined.
  • --length <motif_length>: The length of the motif or peptide to be generated.
  • --output_path <output_path>: Directory where the generated candidate sequences in FASTA format will be saved.

Representative sequence

Use the iFeature software to cluster all sequences from each group. After configuring the environment, execute the following code to complete this process:

python iFeature.py <shuffle.fasta>

Systematically generate all possible combinations of key high-frequency amino acids for each group to obtain a set of candidate sequences:

python centroid.py --input_file <example_data.csv> --output_path <output_path> --random_seed 1 --n_clusters <n_clusters>

Parameters:

  • --input_file <example_data.csv>: CSV file containing PCA clustering results generated by iFeature.
  • --output_path <output_path>: Directory to save clustering results.
  • --random_seed 1: (Optional) Random seed for reproducibility.
  • --n_clusters <n_clusters>: Number of clusters to generate based on the features extracted by iFeature.py.