- Add
jamie
corr method warning
- Add GPU implementation warning
- Complete reruns of all results (NOTE: Results likely differ when running with and without pre-trained models)
- Correct unintuitive sequential evaluation behavior in
evaluate_impact
which caused erroneous chromatin region classification and negatively affected gene prioritization results np.int
deprecation update forMMD-MA.ipynb
- Fix
sklearn
issue withnp.matrix
inscGLUE
notebook run.sh
revised for more complex development solutionsSHAP
version updated to fix deprecatednp.bool
reference- Small notebook revisions, as well as changes from package updates
- Actually upgrade
torch
andtorchvision
- Adjust
README.md
- Fix
pip-tools
generation ofrequirements*.txt
files
- Fix certain compatibility issues, specifically for
requirements.dev
- Fix
sklearn
dependency - GPU compatibility
np.int
deprecation updatesREADME.md
no clone install support- Slightly changed
edModelVar.impute
behavior to expect preprocessed input - Suppress many unimpactful warnings
- Upgrade
PyTorch
version
- Small
sequential
argument fix - SVG visualizations
- Added
return_statistic
toplot_auroc...
function family - Additional statistic reporting surrounding dropout
- Revised visualizations for loss clarity
- Additional runs for imputation
- Clarifications
- More options
- Small terminology fixes
- Add memory logging
- Added
hybrid
method toplot_integrated
- Added
plot_sample
for showing individual cell correlations - Added
sort_type
option tosort_by_interest
utility function - Fixed ordering for
plot_distribution_alone
- More runs, more data
- New visualization for
ATAC -> RNA
imputation gene importance - Visualization fixes
- Figure changes
- Simulation data update
- Added loss logging
- Additional runs
- Alternative phenotypes
- Tuning
- Bugfixes
- More runs with partial correspondence
- Reorganization of notebooks
- Reruns
- Revised hashing function
- Reproducibility for UMAP
- Added
GNU GPL V3.0
License - Better sampling during training
- Delete
generate_plot
class - File structure revisions
- Raw ephys testing
README
updates- Reproducibility for SHAP
- Import hierarchy fix
umap-learn
import fixWR2MD
import versioning
README.md
updates- Reruns
- Reruns
- Small
README.md
changes
- Change name from
ComManDo
toJAMIE
- Include data in repo
README.md
update
- Reruns
- Run leave-one-out on COLO320DM
- Visualization changes and reporting
- Reruns
- Small change to
evaluate_impact
background preview
- Reruns
- Visualization updates
- Weight standardization updates to reduce dimension dependence
- Added
feature_dict
argument toplot_distribution_alone
for customxticks
- Additional outlier protection for
plot_distribution_alone
visualizations - Small fix with label formatting
- Add gradient clipping
- Added
plot_impact
toevaluation
module - Change early stop behavior and add
min_epochs
, mainly for KL annealing - Fix scaling on
F
andP
subsets during training - Increase model robustness
- Lots of small fixes, errors mainly appeared with large datasets
- Model saving for all algorithms and outputs
- Reruns
- Visualization additions and changes
- Add
adjustText
for cleaner text notations - Add
batch_step
option for typical AE iteration - Add more SHAP visualizations
- Add outlier detection utility
- Add scDART
- Added auto-amending kwarg
pca_dim
- Added BABEL datasets
- Adjusted evaluation figure text size
- Applied outlier detection to
plot_integrated
- Change visualizations, especially for distributions
- Changed losses to VAE by default,
cosine
,F
- Changed sampling logic on distribution similarity calculation
- Fix bug for non-aligned datasets in
commando
module - Fix SVD solver option in automated PCA
- Implement full VAE
- Implemented saving and loading models
- New interesting feature finding algorithm
- Stylistic changes in
evaluation
- Various visualizations in
evaluation
module
- Fixed a bug concerning min-max normalization in the
compute_distances
function - Implement SHAP and add visualizations
- Reruns
- Distribution similarity measure
- Evaluation style changes
- Reruns
- Add MMD-MA
- Add VAE functionality
- Add feature distribution previews
- Add inversion to model preprocessing
- Add trainable weighting to model aggregation function
- Additional error handling
- Additional visualization clarity
- Include
BrainChromatin
dataset - Reruns
- Separate plots from
generate_figure
module - Update evaluation module with imputed AUROC distribution
- Visualization reformatting
- Add configurable legend to
generate_figure
- Added correlation and p-value to calibration plots
- Added
feature_names
argument togenerate_figure
- Many Formatting changes for
generate_figure
- Rename and format calibration plot
- Reruns
- Bugfix with
None
pca argument - Reruns, small notebook formatting changes
- Added changeable
k
fortest_LabelTA
and integrated intogenerate_figure
- Automatically chooses appropriate
k
- Added optional
integrated_alg_shortnames
togenerate_figure
- Fixed a bug where PCA was used on singular
None
modalities - Fixed (subverted) a python bug which culls vars exclusively in lambda functions
- Previously prevented proper usage of multiple
pca_dim
generate_figure
formatting changes- Reruns, including PCA fix for
scMNC-Visual
generate_figure
add multi-column colorsgenerate_figure
formatting, title, layout changes- Reruns
- Small code formatting change
- Add more comparison methods to notebooks
- Change
generate_figure
default layout - Change optimizer for main
JAMIE
model - Change plotting style for
generate_figure
stats - Reruns
- Add custom colors to
generate_figure
- Add metric heatmap to
generate_figure
- Fixes to dimensions and generalization in
generate_figure
- Formatting changes to default
generate_figure
- Reruns on all
general_analysis
notebooks
- Rerun
scGEM
data
- Add experimental
corr_method
- Added
dist_method
choice forsim_dist_func
which takescosine
,euclidean
- Generalize
F
representation to allow for negatives - Fixes problem with negative
sim
,diff
,F
, orP
- Reruns
- Use euclidean distance for
sim_dist_func
- Added plug-in pre-calculated
F
matrix throughmatch_result
- Allow for per-dataset
pca_dim
- Fixed
preprocessing
variable default inmodel
- Test patch-seq dataset
- Add pca before use on full model
- Run on MMD-MA simulation data
- Add Davies-Bouldin Index
- Experimental losses
- Run on motor data
- Use
PF_Ratio
in runs
- Fix
scMNC
data loading generate_figure
- Add
_get_..._shape()
functions - Add
_group_plot()
function for automated group partitioning - Add auto-selection of latent features for reconstruction visualization
- Extend shape implementation
- General reorganization
- Streamlined main function
height_ratio
fixes
- Add
- Remove mapping timer printing
- Added group partitioning functionality to
generate_figure
- Refactored
generate_figure
to a class object
- Added custom model functionality with
model_class
argument - Added simple model weight visualization to
generate_figure
- Reruns on various levels of alignment
- Revised
generate_figure
formatting and standardized module format
- Added
integrated_use_pca
option forgenerate_figure
- Experimented with new losses
- New scaling on each loss variable
- Notebook preprocessing changes
- Reran notebooks
- Revised sim-dist measure to only include positives
- Revised model structure
- Stepping with epochs now rather than batches
- Mute output on various functions
- Add tuning function
utilities.tune_cm()
- Add early stopping
- Add experimental loss parameters
- Add alternative similarity measures
- Add correlation visualization in
generate_figure
- Revise
loss_weights
parameter - Revise loss bookkeeping
- Fix F normalization
- Fix model normalization issue
- Remove KNN usage in main model
- Reruns with higher construction losses
generate_figure
- Added 3D plotting capability
- Various formatting changes and style options
- Silhouette coefficient visualizations
- More modality prediction comparisons
- Add MMD-MA comparison
- Add scMNC data
- Reruns
- Added
generate_figure
to more concisely show results - Small changes in several algorithms to mute output
- Renamed
joint_embedding
folder togeneral_analysis
- Re-ran notebooks
- Added
test_label_dist
to show inter-cell distance - Re-run and revise notebooks, especially for
modality prediction
aligned_idx
is nowP
, a matrix filled with priors- Combined
P
andF
matrices into aggregatecorr
- Changes improved unaligned performance significantly
- Slightly reduced aligned performance
- Re-ran joint embedding notebooks
- Fixed bug in
knn
calculation - Added
perfect_alignment
toggle for separate knn graph calculation method - Cleaned up no-longer-used files
- Add evaluation graph for alignment assumptions
- Merge notebooks
- Example directory reorganization
- Removed certain errant checks
- Add modality prediction samples
- Added compatibility for partially aligned datasets using overlapping average vectors
- Added compatibility for differently-sized datasets
- Added "mix-in" metrics to control how much training is done on aligned sets
- Added more visualization for differently aligned datasets
- Make loss function more modular
- Add switchable distance function (Euclidean, Manhattan, Cosine, etc.)
- Simplify loss function
- Fix similarity function
- Re-add connected KNN to F
- Notebook reruns
- Added inverse cross loss
- Notebook reruns
- Implement encoder-decoder model
- Add custom
model
module - Re-run notebooks (doing well!)
- Removed temporary
neighborhood
module - Added neighbor graph utility
knn
tonn_funcs
module - Added guaranteed connectivity to
knn
- Tuning and reruns on all notebooks
- Modified
visualize
UnionCom function - Cleaned loss handling in
project_nlma
- More robust loss output
- Notebook rerun
- Implemented
test_closer
, measuring fraction of samples closer to true match - Moved auxiliary loss calculations to new
nn_funcs
module - Revised loss function
- Refactored code
- Added matrix versions of UC term and NLMA
- Added naïve implementation of Gromov-Wasserstein distance
- Renamed and added notebooks
scGEM
andMMD-MA
- Removed
comparison.ipynb
and added comparisons in each notebook - Reran notebooks
- Added
ALL FIRST
andBATCH FIRST
calculation modes to hybridproject_nlma
- Renamed default NN step timing to
BATCH FIRST
- Finished vanilla loss run with improved label transfer accuracy (
comparison_no-hybrid.ipynb
) - Renamed and reran unfinished hybrid loss run (
sample.ipynb
)
- Deprecated existing
project_nlma
- New
project_nlma
ontsne
backend - Hybrid loss function
- Unfinished test runs
- Temporarily added partial
neighborhood
module from ManiNetCluster
- Remove two-step, gradient optimizations temporarily
- Comparison notebook
- Fix NLMA scaling (coefficient fixes)
- Re-added
tsne
projection method
- Small
construct_sparse
fix
- Cleanup, notes, file removal
- CPU fixes
- Large matrix exclusion fixes
- Expand/shrink matrix normalization
- Reruns
- Notes on coefficient/F-combination difficulties
- Small fixes
- Reruns
- AnnData input support
- Cell cycle partitioning support
- Notebook reruns
- Two-step optimization
two_step_include_large
- Redundant calculation
- NLMA projection fixes
- GPU compatibility
- Sparse compatibility
- Utility functions
- Initial release