To this end, SCA first quantifies the importance of each transcript in each cell by converting transcript counts into surprisal scores (Fig. 1b; Additional File 1: Algorithm 1). To determine the score of a given transcript in a given cell, we compare its expression distribution among the cell’s k nearest neighbors to its global expression (i.e., to the expected distribution of the transcript among a set of k cells randomly chosen from the entire dataset) for a user-specified neighborhood size k. A transcript whose local expression deviates strongly from its global expression is more likely to inform the cell’s location in relation to other cells, and therefore its identity. We quantify this deviation through a Wilcoxon rank-sum test, which produces a p-value representing the probability of the observed deviation in a random set of k cells. Following Shannon’s definition [24], the surprisal or self-information of the observed deviation is then defined as the negative logarithm of its probability, i.e., as −log⁡(p)-\log (p). This is a positive number which measures how surprising the transcript’s local expression is, in units of nats when the logarithm is natural (changing the base scales the scores by a constant factor, which does not affect SCA’s output). To distinguish over- from under-expression, we flip the sign for under-expressed transcripts (Methods). The resulting scores are compiled into a surprisal matrix with the same dimensionality as the input data.