XGBoost model definiton file (.xgb)
An XGBoost ensemble requires manually prepared .xgb file, even when the ensemble is composed of only one member. The file is mainly used to pass the list of descriptors and their training set ranges to ADMET Predictor. It is divided into four sections, three of which are identified by a keyword. The # character in the first columns indicates a comment ignored by the program. The first line is essential and its absence will result in error:
S+XGB Input File <version> <author initials>
The subsequent sections contain lines of tab-delimited elements. The output section for regression models contains:
output
<endpoint name> <endpoint minimum> <endpoint maximum>
The minima and maxima of the endpoint and modeling descriptors are determined over the training set. The output section for classification models is quite different:
output
<endpoint name> <number of classes> [<space-separated class labels in single quotes>] [<space-separated class numbers>]
The next section contains all N modeling descriptors used by the ensemble:
descriptors
<first descriptor name> <first descriptor minimum> <first descriptor maximum>
<second descriptor name> <second descriptor minimum> <second descriptor maximum>
<third descriptor name> <third descriptor minimum> <third descriptor maximum>
...
<N-th descriptor name> <N-th descriptor minimum> <N-th descriptor maximum>
Last section contains all K members of the ensemble, one XGBoost model per line:
components
<first model file>
<second model file>
<third model file>
...
<K-th model file>
Naturally, the .xgb file and all the component model files must be present in the ADMET Predictor data folder. Here is an example of a regression XGB model file, ellipses indicate omitted lines for brevity:
S+XGB Input File v. 1.0 (RF)
# Model for human_VDss_L_per_kg
output
human_VDss_L_per_kg -1.4559319556497243 1.9395192526186185
descriptors
S+logP -22.38042092 11.52741104
MWt 42.04044 3419.92626
MolVol 49.0 3619.0
VMcGowan 49.43 4051.21
N_Atoms 3 239
...
QAtomNeg 0.0 7.36211764
QAtomPos 0.0 5.96105399
F_NLP 0.0 1.79338527
F_HBP -0.00070782 0.79968843
components
0-S_Vd_2D.ubj
1-S_Vd_2D.ubj
...
11-S_Vd_2D.ubj
12-S_Vd_2D.ubj
Here is an example of a classification XGB model file, ellipses indicate omitted lines for brevity:
S+XGB Input File v. 1.0 (RF)
# Model for BBB_Binary
output
BBB_Binary 2 ['High' 'Low'] [0 1]
descriptors
MWt 16.04298 1449.28225
MolVol 28.0 1456.0
VMcGowan 22.22 1730.36
...
F_NLP 0.0 1.79338527
F_HBP -0.00340641 0.99999969
components
53-S_BBB_2D.ubj
...
46-S_BBB_2D.ubj
45-S_BBB_2D.ubj