#17 · ML / AI methods
Explicit prior structure built into omics and clinical prediction models
The common thread is that prior structure is encoded into the model itself rather than recovered afterward. MT-LLE builds locally linear embeddings of longitudinal COPD omics under five jointly optimized objectives, with an RL agent scheduling their weights. DeMoP recasts clinical and copy-number data as text for DeBERTa-based mixture-of-experts prognosis, reporting AUROC 0.939 on GENIE and 0.805 on TCGA. MVGCL fuses noisy multisource drug and disease similarities with type-aware dual-pathway contrastive learning for drug-disease association prediction. InsilicoCell, a pretrained multimodal transformer, departs from this pattern by targeting cellular functional profiles and multi-objective virtual drug screening. The thread is dominant rather than uniform. What is new is explicit encoding of longitudinal, modality, and similarity relations into the learning objective or representation.
Published papers
Contrastive learning of adverse events to provide effective and interpretable vector representations for machine-assisted pharmacovigilance.
Adapts contrastive learning to create interpretable adverse event vector representations from spontaneous reports, outperforming reporting odds ratio and LLM…
Preprints
A Correspondence Score for Validating Learned Patient Representations in Clinical Tabular Data
Proposes a label-free correspondence score, with Bayesian-optimized weights, to test whether learned patient representations of clinical blood-count data…
Interpretable Symptom Vectors for Depression in a Large Language Model
Gemma-3-27B layer-21 activations encode clinician-aligned depression symptom vectors; depression vector gives AUC 0.789 on held-out text.