Theoretical modeling on CRISPR-coded cell lineages: efficient encoding and optimal reconstruction

Abstract

Delineating cell lineages is a prerequisite for understanding the genesis of cell types. Recent studies have demonstrated the feasibility of generating and reconstructing CRISPR/Cas9-coded cell lineages. However, these works have not investigated the limitations or optimality of the encoding or reconstruction processes.Here, we surveyed a multitude of reconstruction algorithms and found hierarchical clustering, with a metric based on the number of shared Cas9 edits, provides the best reconstruction. As to the eciency, the simple encoding method, with constant Cas9/gRNA edit rate, produces exponential reduction in available coding units and severely limits the trackable depth of lineages. To overcome this, we propose alternative encoding methods, one based on parallel gRNA cascades enabled by CLADES, and another based on variable Cas9 editing rate. Both signicantly increase the trackable depth. In summary, we provide a theoretical basis in understanding, designing and analyzing ecient and robust CRISPR-based cell lineage tracking system.

Robust Reconstruction of CRISPR and Tumor Lineage Using Depth Metrics

“…To overcome this, instead of using linkage method, we assign the shared edits of the nodes being merged as the code of the newly created node (Fig.2C) and directly calculate SharedEdit distance between the new node and other nodes. When benchmarked against the set of input trees used in the previous study [4], we found ancestral code assignment resulted in better error rates compared to complete linkage method ( Fig.2D).…”

“…When calculating metric between merged nodes (n 1 ) and other node (n 2 ), assigning shared edits of n 3 and n 4 to n 1 (denoted as n 1 ) can reduce the overestimation compared to complete linkage (which takes minimum SharedEdits of members). (D) Performances of code assignment and complete linkage are compared using the benchmark datasets used in [4]. Code assignment performs signicantly better (paired t-test).…”

“…Russell-Rao metric (or SharedEdits metric as we expanded the denition to non-binary case and renamed in [4]) calculate one minus the fraction of common characters, or in the case of CRISPR-encoding, one minus the fraction of shred edits. We note that the number of remaining units/targets (number of total units minus number of edited units) is a depth of the lineage tree if every cell cycle (tree branch) has at least one edit, as this number reduces along the depth.…”

“…By investigating why this is the case, we found Q-matrix of NJ approximates depth matrix, which consists of the depth of least common ancestors, of the input tree when the input consists of multiple independent trees. Analogously, Russell-Rao metric [3] or SharedEdits metric [4] calculates approximate depth matrix when mutations are irreversible and the state of the root node (i.e. unedited for CRISPR-barcodes) is known.…”

“…Based on these ndings, we created a better bottom-up reconstruction method than the complete linkage method which was found to be best previously [4] by directly assigning inferred code to the new node representing parent of nodes being merged. Further, we calculated probabilistic structure when SharedEdits metric fails to faithfully approximate depth matrix and created an improved metric with less failure in the case of CRISPR-barcodes.…”