#21 · ML / AI methods

Self-supervised networks learn representations of transcriptomic and tissue imaging data

The dominant thread across these items is self-supervised representation learning, which replaces labels with structure intrinsic to the data. Two spatial transcriptomics papers, ST-GCP and MaskGraphene, use graph neural networks with contrastive or masked objectives; ST-GCP adds gene-expression permutation, and MaskGraphene jointly embeds multi-slice, multi-condition data. The imaging items extend this logic to tissue and organ scales: SSGAT identifies regions of interest in 3D tissue images, BASSL-MI finds tumor niches in CODEX imaging with reduced batch effects, and masked modeling of sparse cardiac MR planes yields a latent space for function assessment. The brain-connectome preprint applies an unsupervised graph transformer autoencoder with functional gradient geometry. What is new is attention to batch and condition structure in the data, alongside geometry-aware and masked objectives. The items do not benchmark against each other, so no relative performance claims are made.