Yuzu

Yuzu applies compressed sensing to accelerate in-silico saturation mutagenesis (ISM) for genomic sequence models by reducing the number of forward passes required during convolutional operations.

Key Features:

• Compressed sensing acceleration: Incorporates principles of compressed sensing to reduce the number of forward passes required for ISM.
• Constant forward-pass complexity: Performs ISM with a constant number of forward passes independent of sequence length when model operations have limited receptive fields.
• Receptive-field exploitation: Leverages operations with limited receptive fields, such as convolutional layers, to enable compression.
• Convolutional speedup: Dramatically accelerates ISM during convolution operations, yielding orders-of-magnitude reductions in computation time.
• Scalability with sequence length and complexity: Efficiency gains increase with longer input sequences and greater convolutional complexity.
• Feature attribution support: Enables faster calculation of feature attributions on biological sequences.

Scientific Applications:

• In-silico saturation mutagenesis (ISM): Accelerating ISM analyses for genomic sequence models, especially those containing convolutional layers.
• Feature attribution on biological sequences: Rapid computation of variant effects and attribution scores across sequences.
• Genomics model evaluation: Efficiently evaluating convolutional models used in genomics on longer sequences and complex convolutional operations.

Methodology:

Applies compressed sensing to reduce ISM forward passes to a constant independent of sequence length by exploiting limited receptive fields in operations such as convolutional layers.