Fix missing output for nonsense and essential_splice impacts

[FerriolCalvet]

Problem

Missing outputs for nonsense and essential_splice impacts:

gene    sample  impact  dnds    pvalue  lower   upper
ARID1A  P19_0044_BDO_01 missense        1.0398697       0.8973031       0.49979153      1.8980241
ARID1A  P19_0044_BDO_01 nonsense                                
ARID1A  P19_0044_BDO_01 essential_splice                                
EP300   P19_0044_BDO_01 missense        1.798834        0.028565407     1.0808547       3.3448544
EP300   P19_0044_BDO_01 nonsense                                
EP300   P19_0044_BDO_01 essential_splice   

Possible reasons

There are 0s in the lambdas vectors and the tensorflow probablity functions don't seem to like it too much.

For missense:

l equal to [0.28558427 0.32617283 0.05533479 0.13334824 0.15745339 0.09242073
 0.08529546 0.30062553 0.20218417 0.2424905  0.08098787 0.27460638
 0.13486527 0.32502422 0.00684538 0.14846061 0.01359925 0.0575599
 0.04244946 0.01481647 0.01848553 0.05727808 0.06567501 0.09264341
 0.09275755 0.06764409 0.01564826 0.02131398 0.05997038 0.06238931
 0.01891936 0.12842454 0.0938124  0.12310488 0.17774798 0.04028124
 0.14306849 0.46776822 0.30274367 0.25046015 0.08864892 0.09785648
 0.07335839 0.25764883 0.108295   0.12602973 0.11663748 0.02268062
 0.07946959 0.05692631 0.01860485 0.02654143 0.02170661 0.04168417
 0.13690487 0.03569017 0.01251115 0.05133861 0.01112057 0.07303728
 0.0157018  0.01001406 0.00796137 0.01362443 0.16251975 0.08476225
 0.3185019  0.06119071 0.03862971 0.07041591 0.07634285 0.04007868
 0.08172115 0.01557059 0.09000723 0.0718038  0.02061916 0.03482537
 0.05826616 0.03305257 0.17029199 0.01897544 0.01860485 0.00884715
 0.007079   0.01311499 0.07524361 0.01182637 0.11420705 0.01738286
 0.03394764 0.01460746 0.0127372  0.01234038 0.01358624 0.01617238]

For nonsense:

l equal to [1.2262212e-02 2.1095350e-02 4.3984400e-03 3.9292038e-03 9.4510019e-03
 3.0427321e-03 4.3038931e-03 1.4362105e-02 1.8824339e-03 5.5378429e-03
 2.5045760e-03 2.4753483e-03 7.2978713e-02 4.8018314e-02 8.4807668e-03
 3.2509621e-02 5.8391481e-04 3.7227089e-03 2.6390641e-03 4.3657818e-04
 5.9400496e-05 0.0000000e+00 0.0000000e+00 0.0000000e+00 0.0000000e+00
 0.0000000e+00 0.0000000e+00 0.0000000e+00 1.3675435e-02 0.0000000e+00
 0.0000000e+00 0.0000000e+00 3.1742804e-02 2.1667564e-03 2.5523536e-02
 0.0000000e+00 5.5269703e-02 7.7360771e-03 2.5631344e-02 0.0000000e+00
 5.3811915e-02 2.7419524e-03 0.0000000e+00 0.0000000e+00 3.2564253e-02
 0.0000000e+00 5.9980094e-03 0.0000000e+00 5.0469371e-03 3.4141592e-03
 2.1656503e-03 1.6842084e-03 0.0000000e+00 8.1689312e-04 5.1226473e-04
 2.5927636e-04 1.7851978e-04 1.2167569e-03 1.9530891e-04 0.0000000e+00
 7.4089942e-03 1.8911036e-03 4.8782704e-03 1.7760373e-03 0.0000000e+00
 0.0000000e+00 3.4833457e-03 7.0859439e-04 0.0000000e+00 0.0000000e+00
 0.0000000e+00 0.0000000e+00 0.0000000e+00 1.3559515e-04 0.0000000e+00
 0.0000000e+00 0.0000000e+00 0.0000000e+00 0.0000000e+00 0.0000000e+00
 1.0814865e-02 1.1380531e-03 2.1656503e-03 5.6140282e-04 0.0000000e+00
 0.0000000e+00 0.0000000e+00 0.0000000e+00 0.0000000e+00 1.3559515e-04
 0.0000000e+00 0.0000000e+00 1.8345994e-03 0.0000000e+00 0.0000000e+00
 0.0000000e+00]

Explanation

There are only a limited number of changes that can lead to having a nonsense impact (stop codon creation, start codon truncation ... ) and same for the essential_splice variant types.

Then the 0s of the lambdas vector are real 0s, so filling them is not correct, and we would also never expect to see a mutation with that impact in any of the channels in which there is no probability (lambda = 0).

Then what we can do is remove the positions of the vector in which there is a lambda = 0 and also remove those positions from the mutations vector changing the shape from 96 channels to whatever number of channels are left after the filtering.

This is what I implemented. (see file changes in this PR)

After solving it

gene    sample  impact  dnds    pvalue  lower   upper
ARID1A  P19_0044_BDO_01 missense        1.0398697       0.8973031       0.49979153      1.8980241
ARID1A  P19_0044_BDO_01 nonsense        0.21724111      0.34792298      0.21724111      3.645586
ARID1A  P19_0044_BDO_01 essential_splice        0.21734934      0.5623948       0.21734934      9.230412
EP300   P19_0044_BDO_01 missense        1.798834        0.028565407     1.0808547       3.3448544
EP300   P19_0044_BDO_01 nonsense        3.609274        0.09682387      0.71474063      13.938253
EP300   P19_0044_BDO_01 essential_splice        0.21835195      0.5825151       0.21835195      10.202482

[koszulordie]

It is still a question why the previous implementation failed to provide an output when the lambda (expectation) vectors had zero expectation components. But the proposed solution makes sense in that it should give in theory the same behavior as intended by the original MLE implementation and it seems to be working in practice.