Researchers have developed a new antibacterial peptide finder based on artificial intelligence.

Edit | Chinese cabbage leaves

Antibiotic resistance is an increasingly serious health problem. Antimicrobial peptides (AMP) destroy harmful microorganisms through nonspecific mechanisms, making it difficult for microorganisms to develop drug resistance. Therefore, they are expected to be substitutes for traditional antibacterial drugs.

In this study, researchers from Gwangju Institute of Science and Technology (GIST) in Korea developed an improved AMP classification model called AMP-BERT.

The team proposed a deep learning model with a fine-tuned bidirectional encoder representation from Transformers(BERT) architecture, aiming at extracting structural/functional information from input peptides and identifying each input as AMP or non-AMP. The researchers compared the performance of this model with other models based on machine learning. AMP-BERT produced the best prediction results among all the models evaluated using external data sets selected by researchers.

In addition, researchers use the attention mechanism in BERT to perform interpretable feature analysis and identify specific residues in AMP that are known to contribute to peptide structure and antibacterial function. The results show that AMP-BERT can capture the structural characteristics of peptides used for model learning, so that AMP or non-AMP can be predicted from the input sequence. AMP-BERT is expected to help identify candidate AMPs for functional verification and drug development.

The study is based on "AMP-BERT: Prediction of antimicrobial peptide function based on a BERT modelThe title was published on December 3, 2022 in "Protein Science》。

In the past decades, antibiotic resistance has become a major public health problem in the world. This leads to the search for alternative methods to treat microbial infections.

One of the innovations is the discovery of antibacterial properties of some peptides. Antimicrobial peptide (AMP) is a short peptide found in most animals, plants and microorganisms, which can be used as a natural defense against infection. AMP fights against harmful bacteria through non-specific mechanism and prevents them from producing antibiotic resistance.

Despite these extraordinary abilities, the research on AMP is still hindered because the existing system for identifying candidate AMPs is like a black box, and the output is not easy to interpret for further analysis.

Now, in a breakthrough recently published by protein Science, a group of researchers from Gwangju Institute of Science and Technology, including Professor Hojung Nam and Mr Hansol Lee, put forward an AMP-BERT classification system, which uses AI-based bidirectional encoder representation from Transformers (BERT) architecture to improve the existing AMP classification model.

When asked about the motivation behind the development of the classification system, Professor Nam explained: "The abuse and overuse of antibiotics have led to the emergence of bacteria that cannot be effectively treated by these antibiotics. This not only increases human health risks, but also increases the health risks of agriculture. Therefore, we hope to develop an AMP pre-screening platform, which is not the black box of the algorithm, but can be easily explained for further research. 」

The team integrated a deep neural network based on natural language processing (NLP), which was pre-trained with billions of protein sequences and then fine-tuned with thousands of peptide sequences in the benchmark AMP database. This makes the AMP-BERT model not only extract structural and functional information from the input peptide sequence, but also distinguish AMP from non-AMP. This enhances the prediction ability and enables the model to make better classification even with external data.

The team also designed a model to assign a separate attention score to each amino acid in the input peptide sequence. Then, the attention characteristics reveal the important sub-regions of AMP, which play an important role in determining whether the peptide has antibacterial properties. In addition, the prediction results show that the applicability of AMP-BERT model even extends to invisible peptide data, and it can learn meaningful functional and structural information from these peptides.

The new AMP-BERT peptide pre-screening model can open a new door for discovering and developing AMP-based candidate drugs for the treatment of drug-resistant diseases. The important peptide subregion information provided by the prediction platform can also be used to optimize the antibiotic efficiency of peptides.

Professor Nam concluded: "As more and more AMPs are verified by experiments and new structural information is discovered by computational methods, we will be able to manufacture more effective antibiotic drugs and may prevent new pandemics from spreading around the world in the near future. 」

Data set:

Paper link:

Related reports: