-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathNEWS
112 lines (82 loc) · 5.58 KB
/
NEWS
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
CHANGES IN VERSION 0.99.12
------------------------
* Incorporating immApex Usage
* Update ```CoNGAfy()``` to support all versions of Seurat
* Update trex_example to Seurat v5
* Updated ```quietTCRgenes()``` to used VariableFeatures() call
CHANGES IN VERSION 0.99.11
------------------------
* Added geometric encoding using the BLOSUM62 matrix
* Sequence input:
- TRA: 905611
- TRB: 1276462
* Trained classical and variational autoencoders for TRA/TRB chains
- Architecture: 256-128-30-128-256
- Parameters:
Batch Size = 64
Latent Dimensions = 30
Epochs = 100
Loss = Mean Squared Error
- Optimizers: Adam
- Early stopping was set to patients of 10 for minimal validation loss and restoration of best weights
- learn rate varied by models
- classical auto encoders have batch normalization layers between the dense layers.
DEPRECATED AND DEFUNCT
* Deprecate clonalCommunity()
CHANGES IN VERSION 0.99.10
------------------------
- fixed issuer with TRB OHE encoder returning NaN or zero values
- "both" call now pulls all AF/KF without error
CHANGES IN VERSION 0.99.9
------------------------
- maTrex() now offers support for scRepertoire combineTCR() output and not just single-cell objects alone.
- CoNGAfy() function modified for additional meta variable support - will automatically check the duplicated clones for multiple values and condense with ";"
- New format of encoder models to reduce size of package
CHANGES IN VERSION 0.99.8
------------------------
- update to the database used in annotateDB(), cleaned data extensively
- Correct CoNGAfy() function to fix mean-based reduction
- Added edit distance parameter to annotateDB()
CHANGES IN VERSION 0.99.7
------------------------
- Added one hot auto encoder to model options
CHANGES IN VERSION 0.99.7
------------------------
* Dev/Maine pull - will update move towards new model architecture from here.
CHANGES IN VERSION 0.99.6
------------------------
* After comparing the outputs of various calculations Trex offers - by far the best performing is returning the auto encoded values. Removed distance calculations and gene usage statistics.
* Expanding auto encoder models to use 288,090 unique TCRs and 455,779 unique TCRBs derived from both single-cell cohorts and bulk databases - VDJdb, McPAS, IEdb, and TRAdb.
* Plan to add one-hot auto-encoder for sequence information and a database of TCRs
CHANGES IN VERSION 0.99.5
------------------------
* CoNGAfy now can reduce the size of multiple dimensions using mean or distance.
* Moving the auto encoder distance back to cosine distance - per row the calculation will be 1-abs(cos())
* Modified runTrex() to allow for generating layers without the edit distance (before it just errored out).
CHANGES IN VERSION 0.99.4
------------------------
* Added CoNGAfy() function to reduce RNA to a single cell per clonotype based on either mean expression or euclidean distance of the PCA.
* Changed normalization for length-based distance calculation, was effectively halving the true distance, over calling neighbors when using thresholding.
* AA property normalization now uses Euclidean distance instead of cosine. Simplifies the normalization.
* Removed the multicore parallelization, after implementing the above (and removing the over calling issue) the code is substantially faster when using thresholding!
CHANGES IN VERSION 0.99.3
------------------------
* Added clonalCommunity() function to allow for assignment and subsequent regression of feature space of groupings.
* Changed neighbor.manager. Reverted the nearest neighbor proximation to the original version which only randomizes clones. Threshold detection changed to use expand.grid() instead of looping through all barcodes.
* Added support for aaReduction() and distReduction to the clonalCommunity() function
* Added multicore parallelization to aaProperty() and distanceMatrix() function to speed up reduction. On an object of 3200 cells, sped the calculation up 8x.
CHANGES IN VERSION 0.99.2
------------------------
* Added seed parameter to improve reproducibility
* Added clone.proportion parameter to select for the relative number of duplicates to allow into nearest neighbor evaluation
* New internal function neighbor.manager - nearest neighbor approximation is now based on clone proportion (for clones), the remaining neighbors are randomized across unique clones in the vicinity of the clonotype/cell in question.
* Modified Mean approach for aa.property now references the correct columns in generating output frame and the cosine distance function
* Changed internal function parameters to prevent the recursive issue.
* added aaReduction() and distReduction() functions to get intermediate values for a layer.
CHANGES IN VERSION 0.99.1
------------------------
* Changed metric functions as it related to memory. Firstly, no more as.matrix() calls, this prevents upper ceiling on memory usage.
* Now distance is computed and from each row, a normalized comparison is made using the method for near call, which is then converted and stored as an edge list. This eliminates 3 matrices - pairwise distance matrix, normalized distance matrix and the adjacency matrix. Based on some benchmarking, this is a reduction in memory of around 3 (matrices) x 50.
*Edge list is now used to form multiplex graph object, with the list of edges across layers used to make supraAdjacency matrix.
* AAProperties distance calculated using cosine using the philanthropy package - this returns a distance object instead of a Pearson correlation-based pairwise matrix, saves time and memory
* Gene Usage Metrics return edge list directly, no intermediate adjacency matrix