CRISPRL and: Interpretable large-scale inference of DNA repair landscape based on a spectral approach

Abstract

Summary We propose a new spectral framework for reliable training, scalable inference and interpretable explanation of the DNA repair outcome following a Cas9 cutting. Our framework, dubbed CRISPRL and, relies on an unexploited observation about the nature of the repair process: the landscape of the DNA repair is highly sparse in the (Walsh–Hadamard) spectral domain. This observation enables our framework to address key shortcomings that limit the interpretability and scaling of current deep-learning-based DNA repair models. In particular, CRISPRL and reduces the time to compute the full DNA repair landscape from a striking 5230 years to 1 week and the sampling complexity from 1012 to 3 million guide RNAs with only a small loss in accuracy (R2R2 ∼ 0.9). Our proposed framework is based on a divide-and-conquer strategy that uses a fast peeling algorithm to learn the DNA repair models. CRISPRL and captures lower-degree features around the cut site, which enrich for short insertions and deletions as well as higher-degree microhomology patterns that enrich for longer deletions. Availability and implementation The CRISPRL and software is publicly available at https://github.com/UCBASiCS/CRISPRLand.

Sparse Epistatic Regularization of Deep Neural Networks for Inferring Fitness Functions

“…The WH loss penalizes the distance between DNN and a function constructed using the top- k WH coefficients recovered in the first subproblem. In order to solve the first subproblem, we design a careful subsampling of the input sequence space [22] that induces a linear mixing of the WH coefficients such that a greedy belief propagation algorithm (peeling-decoding) over a sparse-graph code recovers the noisy DNN landscape in sublinear sample ( i.e ., ) and time ( i.e ., ) complexity in p (with high probability) [13, 22, 24, 25]. Briefly, the peeling-decoding algorithm identifies the nodes on the induced sparse-graph code that are connected to only a single WH coefficient and peels off the edges connected to those nodes and their contributions on the overall graph.…”

“…We denote the rows corresponding to these subsampled sequences as X T , where | T | ∼ 𝒪 ( k log 2 p ). The subsampling induces a linear mixing of WH coefficients such that a belief propagation algorithm (peeling-decoding) over a sparse graph code recovers a p -dimensional noisy landscape with k non-zero WH coefficients in sublinear sample ( i.e ., 𝒪 ( k log 2 p )) and time complexity ( i.e ., ( k log 3 p )) with high probability [13, 22, 24, 25] (see Supplementary Materials for a full discussion). This fully addresses both the time and space scalability issues in solving the u -minimization problem.…”

“…The dimension of the layers were set to d × fd, fd × fd, fd × d , and the dimension of the final layer was d × 1, where f is an expansion factor. We searched for a value of f that resulted in best generalization accuracy on an independent data set—a prediction task on DNA repair landscapes [13] which we did not use for evaluation in this paper. DNN prediction performance was stable around f = 10 with highest validation accuracy on the independent data set.…”

“…Recent advances in next-generation sequencing have enabled the design of high-throughput combinatorial mutagenesis assays that measure molecular functionality for tens of thousands to millions of sequences simultaneously. These assays have been applied to many different sequences in biology, including protein-coding sequences [1–3], RNAs [4–6], bacterial genes [7–10], and the SpCas9 target sites [11–13]. The labeled sequences collected from these assays have been used to train supervised machine learning (ML) models to predict functions ( e.g ., fluorescence, binding, repair outcome, etc .)…”

“…Recent studies in biological landscapes [3, 13, 21] have reported that a large fraction of the variance in many fitness functions can be explained by only a few number of (high-order) interactions between the mutational sites. The epistatic interactions in these functions are a mixture of a small number of (high-order) interactions with large coefficients, and a larger number of interactions with small coefficients; in other words, their epistatic interactions are highly sparse.…”