#24 · ML / AI methods

Latent diffusion models extend to histology staining and drug-response prediction

The common thread is generative latent modeling in which biological structure is imposed as explicit conditioning rather than left to generic embeddings. HistDiT uses a latent diffusion transformer with dual-stream conditioning to virtually stain HER2 IHC from histopathology images, claiming gains over GAN-based approaches. The drug-response model pairs a VAE with a diffusion process to predict condition-aware drug-induced gene expression, with improved generalization to unseen compounds. The retina study is less closely aligned: it couples a gap-junction filter with a DNN as a biological hybrid model, reporting improved adversarial robustness and altered representational manifold geometry along the visual hierarchy. What is new across these items is diffusion-based latent generation with explicit conditioning signals, and, in the retina case, a biologically grounded front-end that constrains the learned representation.