Phate

phateR v0.2.8 replaces the `potential.method` parameter with `gamma` to match Python PHATE v0.2.8.

Use `gamma=1` (default) for the log potential, `gamma=0` for the square root potential. Valid values of gamma are between -1 and 1.

Bugfix resolves issues with reticulate passing `NA` to Python as `True`.

PHATE now calls the much faster code written in Python to perform dimensionality reduction using the `reticulate` package. PHATE must be installed in both Python and R in order to use the R code.

Other changes:
- precomputed distance matrices are now accepted, with `phate(data=distances, knn.dist.method="precomputed")`
- precomputed affinity matrices must be provided in the same way, with `phate(data=affinities, knn.dist.method="precomputed")`, rather than via `g.kernel`.
- multiprocessing is supported - call `phate(..., n.jobs=5)` to use 5 cores or `phate(..., n.jobs = -1)` to use all available.

The Python version of PHATE now accepts both distance matrices and affinity matrices with the keyword `knn_dist='precomputed'`.

We assume distance matrices have only zeroes along the diagonal, and affinity matrices have no zeroes on the diagonal.

PHATE now accepts scanpy's native AnnData format

Version 2.0 implements fast scalable PHATE in Python (2.7, >=3.5), MATLAB and R.

PHATE now runs in seconds to minutes on tens of thousands of cells. Benchmarking shows runtime of ~3 hours on >1,000,000 cells.

Key changes:

* t is automatically chosen using Von Neumann Entropy using t='auto' by default.
* n_landmark determines the number of landmarks to use for scalable PHATE. n_landmark=None ([] on MATLAB, NA in R) specifies exact PHATE.
* default parameters have changed: k=15, a=10 but is ignored if n_cells>=n_landmark.