ANNE Definitions (.net)
Two files are generated for each ANN ensemble, an [ANN model definition file (.ann)], and a network file (.net). The naming convention used is:
<data file name>_<number of inputs>_<number of Neurons>.<extension>
The network file records the entire architecture of each network in an ANN ensemble. The file consists of the following blocks, one for each network in the ensemble.
Network header
<program identifier> <networks per ensemble> <number of neurons> <number of descriptors>
type <network type> (all ADMET Modeler networks are multi-layer perceptron [MLP or LBN] type)
error_fn <error function type> (all ADMET Modeler networks use sum of squared errors)
no_layers <number of layers> (<neurons in layer 1> ... <neurons in layer n>)
Layer 1 to 3
layer <layer number>
no_units <number of neurons>
width <width> (all ADMET Modeler networks use width=1)
act_fn <activation function: Linear|Hyperbolic>
psp_fn <post-synapse potential>
[<threshold 1>
...
<threshold n>]
[fixed weights <layer number - 1>
<weight 0,0> <weight 0,1> ... <weight 0,i>
...
<weight n,0> <weight n,1> ... <weight n,i>]
LBN type, layer 3 only
[linear_boost_weights
<linear weight 0> <linear weight 1> ... <linear weight i>]
Settings
normalisation
classify_by error
accept <accept>
reject <reject>
use_single_threshold <if ANNC averaging was used>
bbfxns <confidence analysis parameters>
use_enmb_thresh <ensemble threshold>
pred_value <predictive values>
steps <steps>
lookahead <lookahead>
input_normalisation <input normalization>
output_normalisation <output normalization>
no_inputs <number of network inputs>
no_outputs <number of network outputs>
no_normed_inputs <number of normalized network inputs>
no_normed_outputs <number of normalized network outputs>
attributes
<selected input name 1> (names of the inputs that were kept for this network)
...
<selected input name i>
factors
Minimax Mean <normalized shift 1> <normalized scale 1> <shift 1> <scale 1>
...
Minimax Mean <normalized shift i> <normalized scale i> <shift 1> <scale i>
missing values
<missing value 1>
...
<missing value i>
A portion of a network file is shown below; ellipses (...) show where lines have been omitted.
# ADMET Modeler(TM) 4.1.0055 Network 32 9 60
type MLP
error_fn "Sum-squared"
no_layers 3 (5,3,1)
layer 1
no_units 5
width 1
act_fn "Linear"
psp_fn "Linear"
1
…
5
layer 2
no_units 3
width 1
act_fn "Hyperbolic"
psp_fn "Linear"
-0.542247
…
fixed_weights 1
-1.2080445 0.513286 -0.206622 0.707976 -0.708153
…
layer 3
no_units 1
width 1
act_fn "Linear"
psp_fn "Linear"
-0.957435
fixed_weights 2
1.421535 -1.058365 1.252600
normalisation
classify_by error
accept 0.5
reject 0.5
use_single_threshold 0
steps 1
lookahead 0
input_normalisation "None"
output_normalisation "None"
no_inputs 5
no_outputs 1
no_normed_inputs 5
no_normed_outputs 1
attributes
A_NeutAQ
A_IoniCC
pKa_1
factors
Minimax Mean 0 1 0.596706 6.346309
…
Minimax Mean 0 1 0.431893 0.110741