Extended Data Fig. 3: Effect of morphological constraints on count-based deconvolution of ST data with NMF.

(a) Reconstruction of expression by NMF⁴⁸ is unaffected by constrained weight parameter \((\alpha )\). Distribution (y axis) of mean squared error (MSE, x axis) between observed (N = 11761 genes) and SPOTlight-reconstructed expression (Methods) for N = 69,721 spots at different values of \(\alpha\) (color code). (b) Cell compositions proposed by NMF decrease in entropy with increased \(\alpha\). Distribution (y axis) of entropy of cell composition vectors predicted by constrained NMF as in (a). Entropy (x axis) is calculated across all 17 snRNA-seq cell types \(({k}_{{sn}}=17\)), and thus ranges between 0 (one cell type present) and 1 (all cell types present in equal proportion). (c) Reconstruction of morphological cell compositions is greatly enhanced with increased \(\alpha\). Distribution of residual MSE between predicted and observed morphological cell type (“superclass”) composition vectors from SPOTlight as in (a). Morphological cell type vectors are calculated from the output of SPOTlight by pooling snRNA-seq cell types from the same morphological class (\({k}_{{morph}}\), as in Supplementary Table 3), reducing the space from \({{R}_{0\le x\le 1}}^{{k}_{{sn}}=17}\) to \({{R}_{0\le x\le 1}}^{{k}_{{morph}}=5}\).