This page describes all methods used throughout MANTA in relative detail. The methods are ordered based on where in the MANTA pipeline they are necessary.
Voxelization
Multiscale Voxelization
- We assume the user provides the increasingly smaller voxel extents in order to voxelize the data.
- Since ideally, the lattices of each voxel are elastically transformed, voxelization will need to be redone after each alignment scale → it will need to be fast!
Voxel Features
We collect a voxel feature for each voxel.
Average Gene Expression
In the most simple model, we collect the average gene expression profile across all samples in the volume of a voxel:
While this approach is extremely scalable and fast, there are several serious drawbacks to this method:
- There is no spatial information contained within the voxels. All gene expression is aggregated into one number, which is representative of the whole voxel. This incorrectly assumes uniform gene expression of particular genes across a voxel. As voxels get smaller, this assumption is correct, however we do not operate on such small voxels.
- In theory, observations are simply a sample of the whole tissue in question. As such, we should treat is as a sample and not a ground truth.
Novae
Voxel Matching
In order to align voxel, they first need to be matched against each other. In this step, we assume to have a (set of) voxel feature(s) available. To match voxels we can initially look at the correlation between them.
Rigid Voxel Alignment
Alignment | Rigid | Correlation-based
- For each voxel in source dataset:
- Find -best correlators (user specified, hyperparameter?) in target dataset based on feature vector.
- Find centroid location of best correlators (⇒ average location). In the case , this is simply the centroid of the most correlated voxel!
- Save translation vector in translation distribution.
- Next, find KDE over translation distribution and choose the mode (i.e. point of highest density) = mode vector
- Move all points by the mode vector.