Assessing ion channel blockade and electromechanical biomarkers’ interrelations through a novel Multi-Channel Causal Variational Autoencoder
Résumé
Knowing the impact of causal relationships between ion channel
blockade and electromechanical biomarkers is essential to improve drug-induced torsades de pointes (TdP)-risk assessment. Apart from common
purely electric torsadogenic indices, mechanical biomarkers may provide
additional proarrhythmic information, but the impact and interrelationships
between those variables need to be assessed to guide feature selection for
classification methods. Variational Autoencoders (VAE) offer a reliable
framework for learning disentangled representations from complex data distributions and can handle a variety of data types and structures. Causal discovery strives to reveal causal links between observed variables, thereby
providing a better understanding and insight into the phenomenon under investigation. Nevertheless, establishing causal relationships between heterogeneous multichannel observations is far from straightforward.
We
propose a novel VAE architecture, Multi-Channel Causal Variational Autoencoder (MC$^2$VAE), to identify mutual relationships between ion channel blockades, torsadogenic biomarkers, and electromechanical
biomarkers, considered here as three
distinct channels, i.e. three distinct
sources of information for drug-induced TdP risk. Our approach for
causal disentanglement from multi-channel data is designed to search for a linear causal structure between the
generated latent variables, shifting the problem of causal discovery from a
heterogeneous multichannel space to a compact lower-dimensional one,
while encoding and decoding operations are proper to each modality to better
adapt to their own specificities. Conclusion: Our approach interestingly suggests the existence of hidden (latent) causal relationships between the three
considered sets of biomarkers, providing a rationale for including mechanical
biomarkers in TdP-risk assessment approaches. Further, MC$^2$VAE is able to
quantify the strengths of the identified causal relationships, opening up a viable avenue for actionable interventions on the established graph.