The day in conversation for 2026-10-31

discussion

Transcript

The discussion: Ines: The deep learning work on non-coding variants really highlights that the bottleneck isn't just the model architecture, but obtaining sufficiently granular, validated molecular phenotypes to ground those predictions against. Marcus: True. When we look at scTREAT’s approach with the shared latent expression space, it seems like a necessary abstraction when direct phenotypic assays are too sparse for complex drug responses; how does that latent space map onto the specific causal drivers identified in those non-coding variant analyses? Yuki: I wonder if we are overcomplicating the connection. That paper on persistent homology in organoids showed that structured topology emerges even at low node counts, suggesting perhaps those deep learning models aren't just pattern recognition but are implicitly capturing some underlying structural constraint that topological methods look for explicitly. Ines: But the ground truth needed for the variant models is incredibly difficult to generate reliably across diverse genetic backgrounds, unlike the controlled environments used in organoid activity where we can map spatiotemporal propagation patterns with pyAvalanches. If we could translate those low-node count topological insights into a feature engineering input for our variant predictors, wouldn't that stabilize the training data? Marcus: That would be interesting—it shifts the dependency from external wet-lab validation to internal structural consistency within the biological system itself. However, we still need to define what

More episodes

← Home