As part of our effort to develop a robust image-based artificial intelligence (AI) platform for drug discovery and development, we designed a new method based on a concept of Chaos Game Representation (CGR) to characterize peptide sequences. The CGRs were encoded into images for the peptides based on a supervised autoencoder (SAE), and these images were then used to build deep learning convolutional neural network models for prediction of peptide hemolytic properties. The approach can guide lead optimization to derive safe peptidic agents. Concurrently, such CGR-SAE combination has also been devised to model the susceptibility/resistance of AIDS patients with different HIV protease mutations to FDA approved drugs, available here. Additionally, a large language model-based PepHemo-LLM has also been built to tokenize peptide sequences and identify hemolytic peptides, achieving similarly high prediction robustness and accuracy.

Input of Peptide Sequence(s) in FASTA

Output of Predicted Hemolytic Properties