Deep learning on motion vectors flags drug-induced contractile damage in stem-cell heart cells
- Label-free video of micropatterned hiPSC-cardiomyocytes, which contract along a single axis, serves as a scalable contractility assay.
- Optical-flow motion vectors, gated per beat, feed a two-branch convolutional network that reads horizontal and vertical velocity separately.
- On independent assays with doxorubicin, trastuzumab and sunitinib, it beat conventional kinematic classifiers, especially at low doses; ablations showed both motion axes contribute.
💡 Points
hiPSC-cardiomyocytes are still immature and a classifier inherits whatever its labels get wrong, so the open question is how these calls track human clinical cardiotoxicity. Three reference drugs with distinct mechanisms is a start, not a validation set. Low-dose sensitivity is attractive, but without exposure matched to free plasma levels it says little about safety margin.