that is a pretty bad example of the issue. I strongly suggest using view_select to encompass several distinct regions that are NOT the dominant view (even though there are no obvious “hotspots”), running 2D classification with ~10-30 classes for each subset, then training a topaz picker for any 2D that looks different from preferred view. Simultaneously you should consider other ways to prepare grids (LEA? grid type/coating? detergents?) as this might not be insurmountable. In future if you have prepared multiple grids and you can manage to identify strong preferential orientation early, consider screening another - sometime’s the strong bias is grid-specific as it happens very quickly post-blot and is dependent on ice thickness and time.
1 Like