APOGEE
APOGEE is a LMT-based consensus classifier.
LMT (Logistic Model Tree) is a machine learning technique
which consists of a combination of decision trees and logistic regressions at the leaves.
The model is evaluated on the basis of some predictor variables that can be used for making
decisions
in the tree construction and selected for logistic models.
The difference between decision tree and LMT is that the former classifies all the instances
belonging to a
leaf with the class having the highest frequency in the leaf. While LMT constructs a logistic model
for
classifying the instances in the same leaf by giving, to each instance, the probability of belonging
to a class.
APOGEE handles two pathogenicity classes: neutral and pathogenic. Mutations are
considered as instances of the following predictors:
- PhyloP 100V
- PhastCons 100V
- PolyPhen2 (HumDiv dataset)
- SIFT
- FatHmm (weighted version)
- PROVEAN
- Mutation Assessor
- EFIN (SwissProt dataset)
- EFIN (HumDiv dataset)
- CADD Phred
- PANTHER
- PhD-SNP
- SNAP
Once defined the classification function, we implemented and tested a bootstrap strategy,
which randomly selected 70% of the pathogenic mutations and considered the same number of neutral
mutations. In brief, for 100 iterations, we run this algorithm:
- Sampling the training set, as described above;
- Estimating the LMT;
- Predicting the pathogenicity of all the mutations stored in the database.
Each iteration gave an estimate of pathogenicity for each variant. These were summarized by
calculating the probability mean for each variant. A variant was deemed harmful if the mean of the
probabilities of being harmful calculated on 100 runs resulted > 0.5. Compared to a single run of LMT,
the bootstrap strategy implies a loss of generalization of the resulting model.
The LMT models generated during the 100 iterations can be downloaded
here.
Comparison of the performance of classification among meta-predictors
| Method |
Accuracy |
Precision |
FDR |
MCC |
MCR |
| MetaSNP |
0.54 |
0.29 |
0.71 |
0.09 |
45.83 |
| CAROL |
0.59 |
0.33 |
0.67 |
0.13 |
40.28 |
| Condel |
0.49 |
0.23 |
0.78 |
-0.08 |
51.16 |
| COVERC WMV |
0.59 |
0.33 |
0.67 |
0.12 |
41.27 |
| MToolBox DS |
0.48 |
0.28 |
0.72 |
0.06 |
51.62 |
| APOGEE bootstrap |
0.84 |
0.73 |
0.27 |
0.59 |
15.97 |
top
