Gunnar Ratsch, Multi-omics Analyses

Machine Learning in biomedical research

Gunnar Ratsch, ETH Zurich

Tumor Profiler Study

240 patients, melanoma. Health records, pathology, NGS, cfDNA, bulk RNA, drug testing, single cell genomics and transcriptomics, cfDNA, etc \(\rightarrow\) data integration \(\rightarrow\) molecular research report \(\rightarrow\) summary \(\rightarrow\) decision \(\rightarrow\) treatment

Multi-omics treatment suggestions improved overall treatment outcome (?) No randomized trial.

Challenges:

  1. Predict protein abundance from cell morphology.
  2. Single-cell integration via deep learning and matching.
  3. Modeling in-vitro single-cell drug responses. They measure in vitro treatment response on many cells. Can they predict their response in advance for new cells, new patients, new drugs?
    Measuring a cell destroys it, cannot measure same cell before and after treatment. Can you predict cells after perturbation? Primal Optimal transport problem. Dual formulation. Brenier’s Theorem. CellOT. Bunne, Stark et al, Nature Methods, 2023.

Sequence reading

Metagraph framework

  • given raw sequencing data from thousands of experiments

  • assemble clean graphs for each experiment

  • integrate into joint graph index

  • link to informative metadata (experiment, organism, clinical data, etc.)

  • Binary relation wavelet trees (BRWT) Use \(k = 1\) to obtain.