FastISM

FastISM accelerates in-silico saturation mutagenesis (ISM) for convolutional neural networks to efficiently assess the effects of all possible input-sequence mutations on model outputs.

Key Features:

• Performance acceleration: Provides over tenfold acceleration of ISM on commonly used CNN architectures by reducing redundant computation.
• Convolutional-layer locality: Leverages the observation that a single mutation typically affects only a limited region of intermediate convolutional layers.
• Redundant computation elimination: Identifies and avoids recomputing unchanged activations during ISM to reduce forward propagations through the trained model.
• Comparison with attribution methods: Narrows the performance gap with backpropagation-based feature attribution methods and surpasses their runtime efficiency on multi-output architectures.
• Scalability: Reduces runtime sufficiently to enable ISM across large numbers of sequences.
• Implementation: Implemented in Keras/TensorFlow 2.

Scientific Applications:

• Variant effect assessment: Predicts the effects of all possible mutations within input biological sequences on model outputs.
• Sequence-to-function mapping: Facilitates mapping biological sequences to functional readouts and properties using deep learning models.
• Feature attribution for sequences: Infers contributions of individual characters within input sequences to model predictions to enhance interpretability.

Methodology:

Performs in-silico saturation mutagenesis by executing forward propagations through trained CNNs, identifies the limited regions of convolutional layers affected by single mutations, and eliminates redundant computations to reduce the number of forward propagations; implemented in Keras/TensorFlow 2.