TrendBio

#4 · ML / AI methods

Inferring unmeasured spatial omics layers from histology, transcripts and host signals

8 preprints · 1 published · ↑ 2.4× vs 4-wk avg
Items per day, last 8 weekstoday: 9

The dominant thread is inferring an unmeasured spatial layer from a measured one. TSNR predicts spatial protein abundance from spatial transcriptomics using protein-specific center signals, compact broad molecular context and a Huber loss. A separate benchmark of cross-slice imputation in spatial single-cell proteomics reports that learning-based methods outperform alignment-based ones. PatHoST uses weak supervision to reconstruct pathogen spatial distributions from host spatial transcriptomes, benchmarked on M. abscessus. On the histology side, a multi-scale graph neural network predicts bulk, spatial and single-cell transcriptomics from whole-slide images. HistoTx instead early-fuses H&E patches with transcript tokens at arbitrary resolution under self-supervised training. These two are described mainly through architecture rather than recovery testing. A secondary strand targets cancer survival modeling: CMVAE learns contrastive single-omic latents, GOTSurv applies unbalanced optimal transport to TCGA cross-sectional data, and DMG-PoE handles missing modalities with product-of-experts fusion. What is new is the explicit evaluation of whether imputed spatial layers are faithfully recovered.

Published papers

Preprints

Topic history

DateRank
#4
#21
#31
#29
#24
#2
#40
#19
#8
#8
#8
#36
#5
#10
#3
#5
#3
#21
#20
#20
#8
#13
#19
#11
#11
#13
#23
#22
#37
#6
#3
#4
#2
#2
#1
#8
#9
#19
#13
#23
#33
#35
#32
#37
#40

Sponsors