Engineering

Based on Python's selenium library, we simulated Chrome browser requests (the code is shown below) to build a crawler. And based on the URL structure of the database, we determined the crawling target.

service = Service(executable_path=CHROME_DRIVER_PATH)
options = webdriver.ChromeOptions()
options.add_argument("--headless")
options.add_argument("--disable-blink-features=AutomationControlled")
options.add_experimental_option("excludeSwitches", ["enable-automation"])
options.add_experimental_option("useAutomationExtension", False)

driver = webdriver.Chrome(service=service, options=options)

After crawling the remaining databases, we successfully crawled data from APD3 and DRAMP. We then converted the obtained HTML files into JSON files. Because of the huge amount of data in Uniprot, we could not use search filters to ensure that all the downloaded data were antimicrobial peptides. Therefore, we had to abandon the Uniprot database. At the same time, since the crawler cannot effectively access the dbAMP database, we also discarded it.

Since these data come from different databases, their data structures are different. We unified their data structures. Finally, we removed duplicates based on the sequence attribute and finally obtained more than 39,000 data items. When merging data from different databases, we found that some properties of the same antimicrobial peptides had different values.

We first screened out antimicrobial peptides with a large amount of missing data. Based on the Python Biopython library and the KD scale ^[1], we built an algorithm that can calculate molecular weight, isoelectric point, network charge, and hydrophobicity (the code is shown below) to calculate partial properties of antimicrobial peptides with missing references when data conflict.

from Bio.SeqUtils.ProtParam import ProteinAnalysis
from Bio.SeqUtils import molecular_weight
import json 
import numpy as np

# 定义氨基酸分类
AMINO_ACIDS = 'ACDEFGHIKLMNPQRSTVWY'
HYDROPHOBIC_AAS = {'A', 'V', 'I', 'L', 'M', 'F', 'C'}  # 疏水氨基酸
POLAR_AAS = {'S', 'T', 'N', 'Q', 'C', 'H'}                  # 极性氨基酸
BASIC_AAS = {'K', 'R', 'H'}                                 # 碱性氨基酸
ACIDIC_AAS = {'D', 'E'}                                     # 酸性氨基酸

# KD疏水性量表 (单位: kcal/mol)
KD_HYDROPATHY = {
    'A': 1.8, 'R': -4.5, 'N': -3.5, 'D': -3.5, 'C': 2.5,
    'Q': -3.5, 'E': -3.5, 'G': -0.4, 'H': -3.2, 'I': 4.5,
    'L': 3.80, 'K': -3.9, 'M': 1.9, 'F': 2.8, 'P': -1.6,
    'S': -0.80, 'T': -0.70, 'W': -0.9, 'Y': -1.3, 'V': 4.2
}

def calculate_peptide_properties(sequence):
    """计算抗菌肽的综合性质"""
    # 检查序列有效性
    sequence = sequence.upper()
    if not all(aa in AMINO_ACIDS for aa in sequence):
        invalid_aas = set(sequence) - set(AMINO_ACIDS)
        raise ValueError(f"Invalid amino acids found: {invalid_aas}")
    
    analysis = ProteinAnalysis(sequence)
    
    
    # 3. Molecular Mass
    mass = molecular_weight(sequence, seq_type='protein')
    
    # 4. Isoelectric Point (pI)
    pI = analysis.isoelectric_point()    
    # 6. Net Charge at pH 7
    net_charge = 0
    for aa in sequence:
        if aa in BASIC_AAS:  # 碱性氨基酸 (pHpKa时带负电)
            net_charge -= 1
    # N端氨基 (+1) 和 C端羧基 (-1) 在pH=7时的贡献
    net_charge += 1 - 1
    
    # 7. Hydrophobicity (KD)
    ww_scores = [KD_HYDROPATHY.get(aa, 0) for aa in sequence]
    mean_hydrophobicity = sum(ww_scores) / len(sequence)
    
    return {
        'Sequence': sequence,
        'Length': len(sequence),
        'Mass': round(mass, 2),
        'PI': round(pI, 2),
        'Net Charge': net_charge,
        'Hydrophobicity': round(mean_hydrophobicity, 2),
    }

By comparing with known data, our algorithm shows extremely high accuracy.

After data unification, our data quality has been significantly improved. In addition, the team also designed a unique identifier system to generate a unique SPADE ID for each peptide in the database (using U and UN suffixes to distinguish natural and non-natural antimicrobial peptides) to ensure the uniqueness and traceability of the data. However, the format of the json file makes the data observation less intuitive and user-friendly. Therefore, we also need to use some visualization techniques to present the data in a more intuitive way.

To meet the diverse needs of antimicrobial peptide research, the SPADE platform has designed powerful multi-dimensional information retrieval and display capabilities. Users are no longer limited to single sequence or name searches; instead, they can combine multiple criteria for precise screening, accurately locating antimicrobial peptides that meet specific research objectives. These screening criteria cover various aspects of antimicrobial peptide research, including but not limited to:

Search Criteria Features

Language Support

The database supports five languages:

Database Homepage

The homepage of the database is shown below:

Fig.1. Database homepage display

Peptide Entry Display

The first antimicrobial peptide in the database is shown below:

Fig.2. Antimicrobial peptide entry display

The SPADE database platform now has the powerful search function mentioned above. However, when encountering complex semantic search tasks, common search functions seem to be inadequate.

We retained all antimicrobial peptides with lengths between 1 and 100 in the SPADE database. We matched the non-antimicrobial peptides from the Uniprot database using keywords and used them together with the antimicrobial peptides as data for the AMP classification task. We extracted and cleaned the bioactivity label data from SPADE database for the bioactivity classification task. For task half-life regression, we selected in vitro experimental data as labels (unit: min) and the hosts were all mammalian cells. Because both lethality and concentration are important for analyzing the hemolytic activity of antimicrobial peptides, we fused these two indicators into a hemolytic activity score as the regression target of the hemolytic activity regression task. For task MIC regression, There are 5824 AMPs containing mic_regression labels. An antimicrobial peptide may have multiple target organisms and their corresponding MIC values. We set the high threshold 100000 μg/ml and the low threshold 0.001 μg/ml to clip the original MIC values. We retained microorganisms that appeared more than 100 times.

At the beginning, we used the original data to train the model directly, and the accuracy of the model was very low.

The range of the original data is too large, so we should normalize the data before training the model. After establishing a high-quality training dataset, we need a suitable neural network architecture as the target network for training. The neural network need to complete 5 tasks：1. AMP classification 2. bioactivity classification 3. MIC regression 4. half-life regression 5. hemolysis regression

We built a model architecture with ESM2 150M as encoder and simple MLP (three linear layers in average and elu function as activation function) as decoder. And we set the output dimensions of decoder for each task as follows: amp_classification: 1, bioactivity_classification: 13, mic_regression: 1, half_life_regression: 1 and hemolysis_regression: 1. Finally, the neural network was trained in the order of the tasks above. In order to make a model compatible with all tasks, we use the EWC(Elastic Weight Consolidation) mechanism as a mechanism to prevent the model from forgetting.

We finish the training process and test the model. The output value of the model on each task fluctuates around a certain number, which is obviously wrong. We then tested the ESM2 650M's ability to characterize antimicrobial peptides on the ESM 650M and found that the ESM2's ability to characterize antimicrobial peptides was limited. An example is shown in the figure below. The amino acid masked by the mask token should be V. However, V was not among the top five amino acids predicted by the model.

Fig. 4. Example of the ESM2's ability to characterize antimicrobial peptides

We are going to rebuild the encoder using the Bert architecture and pre-train it using our dataset. At the same time, increase the depth of the decoder's neural network. On reflection, we found that it was unrealistic to integrate all tasks into a single set of parameters.

We selected the mask training task as the pre-training task and then performed the same training task as the previous DBTL cycle. In order to pursue a more stable training process, we set the output dimension corresponding to the task AMP classification to 2 and use cross entropy as the loss function. We abandoned the EWC mechanism and froze the first 5 layers of the encoder after pre-training as shared parameters for all tasks.

In the masked_lm task, the model significantly outperforms the ESM2 150M baseline, achieving a top-1 accuracy of 0.8736 and a top-5 accuracy of 0.9407, compared to ESM2's 0.3300 and 0.6392, respectively.

For AMP classification, the model attains an outstanding AUC score of 0.9951 and an F1 score of 0.9644, indicating strong discriminative ability between antimicrobial and non-antimicrobial peptides.

In the MIC regression task, the model achieves an overall Pearson correlation coefficient of 0.7000 across 18,668 samples. It also shows robust organism-specific predictive performance, with particularly high correlations for Bacillus subtilis (0.8109), Klebsiella pneumoniae (0.7687), and Enterococcus faecalis (0.7450).

For hemolysis regression, the model yields a very low mean absolute error (MAE) of 0.1061, reflecting accurate prediction of hemolytic activity.

Finally, in half-life regression, the model demonstrates nearly perfect predictive capability with a Pearson correlation of 0.9851, underscoring its strength in estimating peptide stability in mammalian systems.

We used early stopping during training and automatically saved the best model for each task. Finally, the best models for all tasks were combined into a single model file (suffixed with pth) with a size of 1.4GB. The model file was uploaded to the Hugging Face model hub and can be accessed via the link: https://huggingface.co/muskwff/amp4multitask_124M.