Combining three "othogonal" datasets can be modeled in a Bayesian network:

A pair of proteins is either interacting (positive) or __not__ interacting
(negative); this is represented by node *A*. Nodes *B1*,*
B2*, and *B3* represent the three "orthogonal" experiments with
outcomes of either positive or negative. The outcome of node *C*
is a prediction based on the information from the three experiments.
The accuracy of this prediction can be expressed in terms of the conditional
probability *p(C|A)*, which is given by:

*p(C|A) = p(C|B1,B2,B3) p(B1|A) p(B2|A) p(B3|A)*

The conditional probabilities * p(B1|A)*,* p(B2|A) *and*
p(B3|A)* are functions of the accuracy of the individual experiments,
thus, the overall prediction accuracy is also a function of these.
The term *p(C|B1,B2,B3)* can be altered in order to change the outcome
of the prediction. For instance, *C* can be chosen to be positive
if any of the three experiment nodes have positive outcomes or if only
one of them has a positive outcome, while the other two are negative (plus
all other possible combinations). Obviously, this affects the overall
accuracy of the prediction. However, in different situations, different
trade-offs between sensitivity and specificity of the prediction might
be desired.

The model equations have been put together in this Excel-spreadsheet.

The three experiments are "orthogonal" in the sense that there are no
direct connections between the nodes *B1*, *B2* and *B3*.
They are conditionally independent, that is, if *A* is given, then
*B1*, *B2* and *B3* are independent.