Final Notes -
Additional Information -
Abstract - Existing approaches for 3D analysis of spatial transcriptomics often overlook either spatial information or experiment-induced distortions which causes discrepancies between the ground-truth in-vivo cell locations and the reconstructed results. The authors propose ST-GEARS, generating anchors (also see 
Spatial Landmark Detection and Tissue Registration With Deep Learning) and connect them across slices. It is a two-step process where first the sections are aligned in a rigid fashion and second, elastic fields are solved to counteract experimental distortions.
Introduction
- Distortion effects are caused by “picking, holding and relocation of the sections”.
- Mentions that
- PASTE(2) can produce rotational misalignments.
- GPSA “considers shape distortions, resulting in shape inconsistency in its loss function, which can cause the model to overfit to local gene expression similarities, leading to mistaken distortions of spatial information”.
- SLAT “doesn’t provide a methodology to reconstruct 3D transcriptomic profiles“
- “Other tools such focus on analysis and visualization of 3D data, such as Spateo, VT3D and StereoPy”
- ST-GEARS is a “3D geospatial profile recovery approach” for ST data through FGW-OT. This is different than other OT-approaches due to the introduction of distributive constraints, increasing the accuracy of anchors obtained from other methods (e.g. Spatial Landmark Detection and Tissue Registration With Deep Learning)
Results
ST-GEARS algorithm
- Personally, the paper language is… sloppy to say the least. At least in this paragraph.
- While not formatted to proper English, the idea seems to be to first perform rigid alignment + anchor computation using FGW-OT with the distributive constraint. Next, using elastic fields and gaussian denoising (i.e. gaussian blur, based on the mention of convolution?). Finally, “Bi-sectional Fields Application” corrects the deformation of each section according to its field, which was calculated using the upper and lower sections.
- This last method is worth a look… They keep mentioning it is mathematically proven (for whatever that means in terms of alignment).
Enhancement of anchor retrieval accuracy through distributive constraints
- The method outperforms in some metrics, but to be completely honest the paper does a bad job explaining what this metric describes exactly, making it hard to properly judge it.