Key result
BeatVecNet deep learning improves detection of drug-induced cardiomyocyte contractile changes with ~0.92 macro-F1 score.
Why the study?
Scalable assays for detecting drug-induced contractile dysfunction in human induced pluripotent stem cell-derived cardiomyocytes remain limited despite advances in electrophysiological cardiotoxicity assessment.
Absolute Event Rate: 0.916% vs 0.706%
The BeatVecNet deep-learning framework outperforms conventional scalar analysis in classifying subtle drug-induced contractile phenotypes in hiPSC-CMs, providing an interpretable platform for scalable in vitro cardiotoxicity assessment.
No takes yet. Share an insight, caveat, or question.
BeatVecNet improves contractile phenotype classification in hiPSC-CMs; extends in vitro screening tools but leaves clinical translation open.
Background and Purpose Assessment of electrophysiological cardiotoxicity in vitro has advanced, yet scalable assays for contractile dysfunction remain limited. BeatVecNet, an interpretable deep‐learning framework, classifies drug‐induced contractile phenotypes from beat‐gated motion vectors in human induced pluripotent stem cell‐derived cardiomyocytes (hiPSC‐CMs). Experimental Approach BeatVecNet was developed and evaluated using beat‐gated motion‐vector tensors derived from hiPSC‐CMs recorded by label‐free phase‐contrast imaging at 150 fps. hiPSC‐CMs were cultured on micropatterned substrates to provide a consistent contraction axis for directional motion analysis. Dense optical flow was used to derive motion vectors, which were retained as beat‐gated spatiotemporal tensors rather than reduced to scalar traces. The model uses two convolutional branches to process horizontal and vertical velocity components separately and was evaluated on independent assays containing control, doxorubicin, trastuzumab and sunitinib. Key Results BeatVecNet distinguished drug‐induced contractile phenotypes with a macro‐F1 of 0.916, outperforming conventional kinematic classifiers (best baseline 0.706), with the largest advantage at low doses. Branch‐replacement ablation indicated non‐redundant contributions of horizontal and vertical motion components, with trastuzumab classification showing greater sensitivity to horizontal‐component replacement and sunitinib classification to vertical‐component replacement. Phase‐resolved attribution further identified temporal patterns within the beat cycle not explained by peak velocity alone. Micropatterning improved signal consistency and contractile coherence, enabling robust directional motion analysis. Conclusion and Implications BeatVecNet extracts drug‐associated contractile signatures not captured by conventional scalar analysis while providing direction‐ and phase‐resolved interpretability at the motion‐representation level. This platform complements electrophysiology‐centred safety assessment and provides proof of concept for classification of subtle drug‐induced contractile phenotypes.
Magi et al. (2026) studied Drug-induced contractile dysfunction. BeatVecNet deep-learning framework vs. Conventional kinematic classifiers was evaluated on Distinction of drug-induced contractile phenotypes (macro-F1 score). The BeatVecNet deep-learning framework distinguished drug-induced contractile phenotypes in hiPSC-CMs with a macro-F1 score of 0.916, outperforming conventional kinematic classifiers (macro-F1 0.706).