An activity canyon characterization of the pharmacological topography.

Background: Highly chemically similar drugs usually possess similar biological activities, but sometimes, small changes in chemistry can result in a large difference in biological effects. Chemically similar drug pairs that show extreme deviations in activity represent distinctive drug interactions having important implications. These associations between chemical and biological similarity are studied as discontinuities in activity landscapes. Particularly, activity cliffs are quantified by the drop in similar activity of chemically similar drugs. In this paper, we construct a landscape using a large drug-target network and consider the rises in similarity and variation in activity along the chemical space. Detailed analysis of structure and activity gives a rigorous quantification of distinctive pairs and the probability of their occurrence.

Results: We analyze pairwise similarity (s) and variation (d) in activity of drugs on proteins. Interactions between drugs are quantified by considering pairwise s and d weights jointly with corresponding chemical similarity (c) weights. Similarity and variation in activity are measured as the number of common and uncommon targets of two drugs respectively. Distinctive interactions occur between drugs having high c and above (below) average d (s). Computation of predicted probability of distinctiveness employs joint probability of c, s and of c, d assuming independence of structure and activity. Predictions conform with the observations at different levels of distinctiveness. Results are validated on the data used and another drug ensemble. In the landscape, while s and d decrease as c increases, d maintains value more than s. c ∈ [0.3, 0.64] is the transitional region where rises in d are significantly greater than drops in s. It is fascinating that distinctive interactions filtered with high d and low s are different in nature. It is crucial that high c interactions are more probable of having above average d than s. Identification of distinctive interactions is better with high d than low s. These interactions belong to diverse classes. d is greatest between drugs and analogs prepared for treatment of same class of ailments but with different therapeutic specifications. In contrast, analogs having low s would treat ailments from distinct classes.

Conclusions: Intermittent spikes in d along the axis of c represent canyons in the activity landscape. This new representation accounts for distinctiveness through relative rises in s and d. It provides a mathematical basis for predicting the probability of occurrence of distinctiveness. It identifies the drug pairs at varying levels of distinctiveness and non-distinctiveness. The predicted probability formula is validated even if data approximately satisfy the conditions of its construction. Also, the postulated independence of structure and activity is of little significance to the overall assessment. The difference in distinctive interactions obtained by s and d highlights the importance of studying both of them, and reveals how the choice of measurement can affect the interpretation. The methods in this paper can be used to interpret whether or not drug interactions are distinctive and the probability of their occurrence. Practitioners and researchers can rely on this identification for quantitative modeling and assessment.