389
12 Consensus Drug Design Using IT Microcosm
(12.18)
The weighting coefficients are
(12.19)
where
p kli = (n kli + 1)/(n ali + n nli + 2) is an a priori probability of classifying compound C
for the i-type descriptor into class k = a, n by method l;
n kli
is the number of compounds in class k of the training
set that are classified as active for the i-type descriptor
by method l; and
δ Сli
is the result of classifying compound C into classes
a (δ Cli = 1) or n (δ Cli = 0) for the i-type descriptor by
method l.
Within the framework of method b (with superior accuracy), compound C is considered to be active in the whole of the QL representation if the separating function
(calculated as in (18)) L b ( C) ≥ 0; otherwise, the compound is classified as inactive.
In the normal strategy, the membership function of compound C belonging to
the activity class k is
(12.20)
where
L b (C)
is the meaning of the separating function (18) for compound C in method
b of superior accuracy;
L b , low
is the minimal value of the separating function (18) in method b of superior accuracy (a sum of w 0 and all negative w i ); and
L b, high
is the maximum value of the separating function (18) in method b of
superior accuracy (a sum of w 0 and all positive w i ).
When using the normal strategy of prediction, we consider both standard and nonstandard regularities for this type of property. Such an approach is advantageous
when modifying the structures of atypical compounds; it expands the prediction
area to the nearest boundary of the training set, but it also increases the risk of
artifacts.
11
( )
,
1, ..., 4
0
1
L C w
w
l
l
l
li Cli
i
δ
=
+
⋅
=
∑
=
11
0
1
1,
1
(1
)
log
log
,
1
(1
)
..., 4
ali
ali
nli
l
li
i
nli
nli
ali
p
P
p
w
w
p
P
p
l
=
-
⋅ -
=
=
-
⋅
=
-
∑
,
,
,
,
,
,
( )
( )
(
) max
, 1
,
2
2
b
b low
b
b low
b high
b low
b high
b low
L C L
L C L
Fb C k
L
L
L
L
β
β
β
β
-
+
-
+
∈ =
-
-
+ ⋅
-
+ ⋅
12 Consensus Drug Design Using IT Microcosm
(12.18)
The weighting coefficients are
(12.19)
where
p kli = (n kli + 1)/(n ali + n nli + 2) is an a priori probability of classifying compound C
for the i-type descriptor into class k = a, n by method l;
n kli
is the number of compounds in class k of the training
set that are classified as active for the i-type descriptor
by method l; and
δ Сli
is the result of classifying compound C into classes
a (δ Cli = 1) or n (δ Cli = 0) for the i-type descriptor by
method l.
Within the framework of method b (with superior accuracy), compound C is considered to be active in the whole of the QL representation if the separating function
(calculated as in (18)) L b ( C) ≥ 0; otherwise, the compound is classified as inactive.
In the normal strategy, the membership function of compound C belonging to
the activity class k is
(12.20)
where
L b (C)
is the meaning of the separating function (18) for compound C in method
b of superior accuracy;
L b , low
is the minimal value of the separating function (18) in method b of superior accuracy (a sum of w 0 and all negative w i ); and
L b, high
is the maximum value of the separating function (18) in method b of
superior accuracy (a sum of w 0 and all positive w i ).
When using the normal strategy of prediction, we consider both standard and nonstandard regularities for this type of property. Such an approach is advantageous
when modifying the structures of atypical compounds; it expands the prediction
area to the nearest boundary of the training set, but it also increases the risk of
artifacts.
11
( )
,
1, ..., 4
0
1
L C w
w
l
l
l
li Cli
i
δ
=
+
⋅
=
∑
=
11
0
1
1,
1
(1
)
log
log
,
1
(1
)
..., 4
ali
ali
nli
l
li
i
nli
nli
ali
p
P
p
w
w
p
P
p
l
=
-
⋅ -
=
=
-
⋅
=
-
∑
,
,
,
,
,
,
( )
( )
(
) max
, 1
,
2
2
b
b low
b
b low
b high
b low
b high
b low
L C L
L C L
Fb C k
L
L
L
L
β
β
β
β
-
+
-
+
∈ =
-
-
+ ⋅
-
+ ⋅