Fig. 1 ∣. Coordinated representations of transcriptomic and electrophysiological profiles with coupled autoencoders.

a, A schematic showing the coupled autoencoder architecture for Patch-seq data. Encoders ($\mathcal{E}$) compress input data (X) into low-dimensional representations (z), whereas decoders ($\mathcal{D}$) reconstruct data ($\tilde{X}$) from representations. The coupling penalty in the loss function encourages representations to be similar across the transcriptomic (t) and electrophysiology (e) modalities. b,c, Three-dimensional coordinated representations of the transcriptomic (b) and electrophysiological (c) datasets. Each point represents a single cell, which is colored by its cell-type membership according to the reference transcriptomic taxonomy. d,e, Performance on supervised cell-type classification tasks at different resolutions of the reference taxonomy. Classification with QDA is performed using three-dimensional representations of the transcriptomic (d) and electrophysiological (e) datasets obtained with coupled autoencoders and with linear methods. f, Performance on within-modality ($X_{\mathrm{e}}\to {\tilde{X}}_{\mathrm{e}}$ and $X_{\mathrm{t}}\to {\tilde{X}}_{\mathrm{t}}$) and cross-modality ($X_{\mathrm{t}}\to {\tilde{X}}_{\mathrm{e}}$ and $X_{\mathrm{e}}\to {\tilde{X}}_{\mathrm{t}}$) reconstruction tasks. Uncoupled representations are not suitable for cross-modal tasks. Error bars show mean ± s.d. over 43-fold cross-validation for panels d–f. Note that there are 1,252 genes versus 68 electrophysiology features in the dataset over which f is calculated.