Artificial intelligence models have been widely used for predicting antimicrobial peptides (AMPs). Previous researchers (Gao et al., 2024) utilized a BERT-based Transformer architecture to predict AMPs. However, these predictions were mainly focused on antibacterial peptides. Therefore, we plan to use DeBERTa to build a model for predicting antifungal peptides (AFP).

DeBERTa (Decoding-enhanced BERT with Disentangled Attention) is a pre-trained language model based on BERT, with two major improvements:

• Disentangled Attention Mechanism: DeBERTa represents each word with two separate vectors for content and position, allowing for more accurate calculation of attention weights between words. This helps the model better understand word relationships, especially in complex sentences.
• Enhanced Masked Decoder: DeBERTa introduces absolute position embeddings in the decoding layer, improving the model’s ability to predict masked words by considering their absolute positions within the context, thus enhancing overall performance.

In this project, we utilize pre-trained BERT models and fine-tune them for specific tasks by adding different output layers. We constructed three sub-models based on DeBERTa: the AFP model, the MIC classification model (MIC cls model), and the MIC regression model (MIC model), to screen for antifungal peptides and predict the minimum inhibitory concentration (MIC).

We collected AFP and MIC data from databases such as DBAASP, dbAMP, and drAMP, and non-AFP data from UniProt. The data were randomly divided into training, testing, and validation sets, labeled accordingly, and then used to train the models. Subsequently, the models were evaluated.

Considering Candida albicans is one of the most common fungal pathogens, studying antifungal peptides against it holds significant clinical relevance. Therefore, we chose C. albicans as our target organism.

Several peptides with the lowest predicted MIC values were selected for experimental testing against the following three C. albicans strains:

• C. albicans SC5314
• C. albicans F5
• C. albicans G5

In the experiments, each peptide was diluted in a gradient using a microdilution method, and its MIC was determined in the above strains. The MIC is defined as the lowest concentration of peptide required to inhibit growth. A screening threshold of MIC values less than or equal to 16 μg/mL was used for selecting antifungal peptides.

To further validate the potential application of AI-predicted AFPs in treating actual infections, we designed a hemolysis experiment and constructed a mouse skin infection model to evaluate their therapeutic effects.

The mouse experiment consisted of three groups: positive control, negative control, and experimental. First, wounds were created on the back skin of each mouse, and C. albicans SC5314 fungal suspension was applied to establish the skin infection model.

Gao, W., Zhao, J., & Yue, Z. (2024). Antimicrobial peptides prediction based on BERT and ensemble learning. Journal of University of Electronic Science and Technology of China, 53(4), 635-640. https://doi.org/10.12178/1001-0548.2023295

Li, T., Ren, X., Luo, X., Wang, Z., Li, Z., Luo, X., Shen, J., Li, Y., Yuan, D., Nussinov, R., Zeng, X., Shi, J., & Cheng, F. (2024). A Foundation Model Identifies Broad-Spectrum Antimicrobial Peptides against Drug-Resistant Bacterial Infection. Nature Communications, 15(1), 7538. https://doi.org/10.1038/s41467-024-51933-2

Yang, M., Zhang, C., Hansen, S. A., Mitchell, W. J., Zhang, M. Z., & Zhang, S. (2019). Antimicrobial efficacy and toxicity of novel CAMPs against P. aeruginosa infection in a murine skin wound infection model. BMC Microbiology, 19(1), 293. https://doi.org/10.1186/s12866-019-1657-6