#2 · ML / AI methods
Molecular property models are being tested across inputs, data splits, and design tasks
The dominant thread is how sequence- and structure-based models are evaluated for molecular properties, with input modality and data regime treated as explicit variables. PCIM-DTA reports drug-target affinity under warm and cold-start settings on Davis and BindingDB-Kd, and the sparse-autoencoder study of ESM-2 probes causal latent features behind zero-shot mutation predictions. Other items extend the same representations to geometry and retrieval: OmniScore encodes proteins, nucleic acids, and small molecules, and PocketScope retrieves binding sites from frozen ESM-C embeddings using late-interaction MaxSim. NCFlow and the peptide design work address generative placement and design rather than evaluation, so this thread is not shared by every item. What is new is treating split design and causal interpretability alongside predictive accuracy.
Published papers
Robust and comprehensive meta-path modeling with multi-graph learning for miRNA-disease association prediction.
SHiMDA, a robust multi-graph learning framework with segment-wise meta-path message passing and robustness-oriented losses, predicts miRNA-disease associations…
RGCNMDA: pathway-bridged relational graph learning with adaptive multi-view fusion for human miRNA-disease association prediction.
RGCNMDA: leakage-controlled multi-view relational GCN with pathway graph and adaptive gated fusion for miRNA-disease association prediction; cold-start AUCs…
PCIM-DTA: pairwise conditional interaction modeling for drug-target affinity prediction under cold-start scenarios.
PCIM-DTA is a pair-conditional interaction model for drug-target affinity prediction, evaluated on Davis and BindingDB-Kd across warm and cold-start settings…
BriGHT: transcriptome-regularized multimodal neuroimaging for brain disorder prediction.
BriGHT is a transcriptome-regularized hypergraph framework for multimodal neuroimaging brain disorder prediction, with reliability-aware fusion, evaluated on…
Preprints
ReTIF: Granularity-Aware Multitask Interaction Routing for RNA-Compound Interaction Prediction and Binding-Site Localization
ReTIF is a multitask model for RNA-compound interaction prediction and nucleotide-level binding-site localization, using separate task-specific cross-modal…
Topic history
| Date | Rank |
|---|---|
| #2 | |
| #2 | |
| #3 | |
| #1 | |
| #1 | |
| #18 | |
| #13 | |
| #15 | |
| #14 | |
| #11 | |
| #17 | |
| #21 | |
| #7 | |
| #12 | |
| #28 | |
| #18 | |
| #19 | |
| #15 | |
| #20 | |
| #19 | |
| #40 | |
| #30 |