SeqProp

SeqProp optimizes DNA and protein sequences using activation maximization with continuous one-hot representations and gradient ascent against a predictor oracle to identify sequences with improved predicted fitness.

Key Features:

• Improved Activation Maximization: Refines activation maximization for differentiable models by optimizing continuous relaxations of one-hot sequence encodings.
• Continuous One-Hot Representation: Transforms discrete one-hot coded sequences into a continuous representation suitable for gradient-based optimization.
• Gradient Ascent with Predictor Oracle: Performs iterative gradient ascent optimization of the continuous sequence representation guided by a predictor oracle.
• Straight-Through Approximation and Normalization: Employs straight-through approximation combined with normalization across input sequence distribution parameters to mitigate vanishing gradients and skewed parameter distributions.
• Enhanced Convergence Speed: Achieves up to 100-fold faster convergence compared to previous techniques.
• Optimized Fitness Optima: Consistently finds improved fitness optima across tested applications.
• Regularization Methods: Supports various regularization methods to maintain confidence, biological relevance, and stability of designed sequences.

Scientific Applications:

• Sequence Design: Designs DNA and protein sequences tailored to properties predicted by deep learning models.
• Protein Structure Prediction: Optimizes sequences to influence protein folding and stability for structure-related objectives.
• Drug Development: Enables creation or modification of sequences to improve therapeutic efficacy as predicted by computational models.
• Vaccine Design: Optimizes antigen sequences to elicit predicted immune responses for vaccine development.

Methodology:

Transforms one-hot coded sequences into continuous representations, applies iterative gradient ascent guided by a predictor oracle, uses normalization and straight-through estimation to stabilize optimization, and applies regularization techniques.