#24 · ML / AI methods

Machine learning and inference applied directly to raw single-trace and image data

The common thread is inference applied to primary measurements rather than heavily summarized features. mCOAST applies cumulants to a single DNA-PAINT intensity trace to count molecular targets and estimate kinetics without calibration or adjustable parameters. A nanopore study compares threshold-based MSD and spectral entropy descriptors with a CNN for classifying nanopore current delays in a DNA-barcoded miRNA assay, making it the one item that directly contrasts hand-defined descriptors with a learned classifier. Piscis trains a deep spot detector for RNA FISH images using a differentiable SmoothF1 loss that approximates F1 score. The DLD study uses ML regressors to predict particle trajectories for designing microfluidic separation of cancer cells. What is new is reliance on near-raw data, though the extent to which engineered features are displaced varies across these methods.