AFP-LSE

AFP-LSE predicts antifreeze proteins (AFPs) from protein sequences using latent space encoding and deep neural network classification to identify AFPs for biodiversity, biotechnological, and pharmaceutical research.

Key Features:

• Latent Space Encoding: Uses a deep auto-encoder to transform protein sequences into lower-dimensional latent representations that retain essential information.
• k-Spaced Amino Acid Pairs Composition: Computes k-spaced amino acid pair composition to capture sequence patterns and structural signals useful for distinguishing AFPs from non-AFPs.
• Deep Neural Network Classifier: Trains a deep neural network on auto-encoder–pruned latent representations to map reduced-dimensional data to AFP/non-AFP class labels.
• Performance Metrics: Evaluated on independent test datasets with a Matthews correlation coefficient of 0.52, an F-score of 0.49, and Youden’s index of 0.81.

Scientific Applications:

• Biodiversity Research: Facilitates identification of AFPs across species to study evolutionary adaptations to cold environments.
• Biotechnological Innovations: Enables discovery of novel AFP sequences applicable to cryopreservation and frost protection technologies.
• Pharmaceutical Development: Supports exploration of AFP properties relevant to designing agents that modulate ice formation in medical contexts.

Methodology:

Latent space encoding via a deep auto-encoder to obtain reduced-dimensional representations, extraction of k-spaced amino acid pair composition features, pruning of the latent space, training a deep neural network classifier on the reduced representations, and evaluation on independent test datasets reporting MCC, F-score, and Youden’s index.