#6 · ML / AI methods
Multimodal fusion of omics, structural, and histology inputs for biological prediction
The common thread is the conversion of heterogeneous biological inputs into learned representations that support downstream prediction. Explicit multimodal fusion appears in SynLeaF, which combines VAE cross-encoders with relational features to predict synthetic lethality across pan- and single-cancer contexts, and in Triad-LMF, a hierarchical low-rank fusion framework over multi-omics data for cancer subtype classification. Other items recast non-standard data for established architectures: Dynomap learns task-optimized spatial feature maps from unordered tabular data so vision models can operate on it, while FEAST and D2Bio link whole slide images to spatial gene expression and genetic biomarkers, respectively. What is new is the focus on how the representation is constructed, through hierarchical fusion, learned spatial cartography, and attention over tissue regions. This pattern is dominant but not universal across the set.
Published papers
Triad-LMF: a hierarchical low-rank multimodal fusion framework for robust cancer subtype classification using multi-omics data.
Triad-LMF, a hierarchical low-rank multimodal fusion framework integrating multi-omics data for cancer molecular subtype classification, reports improved…
Preprints
SynLeaF: A Dual-Stage Multimodal Fusion Framework for Synthetic Lethality Prediction Across Pan- and Single-Cancer Contexts
SynLeaF is a dual-stage multimodal fusion framework predicting synthetic lethality across pan- and single-cancer contexts, using VAE cross-encoders, relational…
Vision-based Deep Learning Analysis of Unordered Biomedical Tabular Datasets via Optimal Spatial Cartography
Dynomap, an end-to-end deep learning method that learns task-optimized spatial feature maps from unordered biomedical tabular data to enable vision models on…
i-IF-Learn: Iterative Feature Selection and Unsupervised Learning for High-Dimensional Complex Data
Proposes i-IF-Learn, an iterative unsupervised method jointly performing adaptive feature selection and clustering for high-dimensional data, tested on gene…
FEAST: Fully Connected Expressive Attention for Spatial Transcriptomics
FEAST predicts spatial gene expression from whole slide images using fully connected negative-aware attention and off-grid sampling, outperforming prior…
Dictionary-based Pathology Mining with Hard-instance-assisted Classifier Debiasing for Genetic Biomarker Prediction from WSIs
D2Bio: a dictionary-based hierarchical pathology mining and hard-instance-assisted debiasing framework for predicting genetic biomarkers such as MSI from WSIs,…
Topic history
| Date | Rank |
|---|---|
| #6 | |
| #2 | |
| #4 | |
| #16 | |
| #19 | |
| #28 | |
| #31 | |
| #30 | |
| #40 | |
| #38 | |
| #38 | |
| #40 | |
| #37 |