Skip to main content
Figure 3 | Journal of Cheminformatics

Figure 3

From: Proteochemometric modelling coupled to in silico target prediction: an integrated approach for the simultaneous prediction of polypharmacology and binding affinity/potency of small molecules

Figure 3

Performance of the DHFR target prediction model compared across a number of parameters. 145 data points annotated against plasmodial DHFR were used as a test set to assess the performance of the target prediction model. The top n predicted non-plasmodial targets were considered (n was varied for values between 1 and 12), after which these targets were extrapolated to plasmodial targets. When n increases, recall values rise up to 36% (with recall values of ~35% for n =3 and n = 4). On the other hand, precision values are 100% for n ≥ 2. The high precision values are likely to be explained by the fact that plasmodial DHFR inhibitors and T. gondii DHFR inhibitors occupy the same chemical space. In addition to varying the parameter n, we performed a 2-fold cross validation (averaged over 20 randomizations), which resulted in a drastic improvement as a recall value of 79% was achieved (with a standard deviation of 10.1%, which is shown as an error bar). These results show that domain-based extrapolations have added value to the prediction algorithm (correct predictions are made even when bioactivity data on plasmodial DHFR is not present in the training set) and that including plasmodial DHFR bioactivity data in the training set can drastically improve recall values.

Back to article page