Fig. 1: Masked autoencoder for panel reduction and marker imputation.

A Model architecture: CyCIF image-derived single cells undergo channel-wise masking followed by the encoding of unmasked channels using a Vision Transformer (ViT). A distinct mask token represents masked channels. A ViT decoder then reconstructs the masked channels, completing the image reconstruction process. B CyCIF channel-wise masking (left) and reconstruction (right): 25-channel images arranged into a 5 × 5 grid format, facilitating conversion from a patch-wise masking strategy into a channel-wise masking strategy. C Iterative marker selection: leveraging the trained model, an optimal marker order is established by gradually increasing the panel size. Each step selects the next marker based on its ability to maximize the Spearman correlation between actual and predicted mean intensity for masked channels. This refines marker panel ordering, enhancing prediction accuracy. Parts of Fig. 1A were created using BioRender (www.biorender.com).