UniRep

UniRep generates unified statistical representations of protein amino acid sequences using multi-layer Long Short-Term Memory (mLSTM) deep learning to enable prediction of protein stability and function for rational protein engineering.

Key Features:

• Deep Learning on Unlabeled Sequences: Applies unsupervised deep learning to unlabeled amino acid sequences to extract representations that reflect structural, evolutionary, and biophysical contexts.
• Multi-layer Long Short-Term Memory (mLSTM): Utilizes a multi-layer LSTM (mLSTM) network architecture to learn sequence-dependent features.
• Unified Representation (UniRep): Produces a unified statistical representation that encapsulates fundamental protein features for downstream analyses.
• Generalization to Unseen Sequences: Models demonstrate the ability to generalize to unseen regions of sequence space.
• Predictive Capabilities: Predicts stability of natural and de novo designed proteins and quantitatively assesses function of molecularly diverse mutants, performing comparably with state-of-the-art methods.
• Efficiency Improvement: Enhances efficiency in protein engineering tasks by up to two orders of magnitude relative to traditional approaches.

Scientific Applications:

• Protein Stability Prediction: Predicts protein stability to inform design and optimization of proteins.
• Functional Analysis of Mutants: Quantitatively evaluates the function of mutants across molecularly diverse variants.
• De Novo Protein Design: Supports de novo protein design by predicting stability and functionality from learned sequence representations.

Methodology:

Trains a multi-layer Long Short-Term Memory (mLSTM) network via unsupervised deep representation learning on unlabeled amino acid sequences to produce unified statistical representations that capture structural, evolutionary, and biophysical signals.