The way in which Canvas fingerprints are computed and stored allows a given “on” bit to be mapped back to the fragment which is responsible for setting that bit. Thus, unlike hashed fingerprints in other software, where many different fragments may collide to set the same bit, Canvas fingerprint bits are perfectly suited to QSAR applications and may be used just like any other calculated descriptor.
KPLS is a public domain methodology that possesses many of the advantages of other kernel-based methods, including the patent-protected technique known as support-vector machines.7 Furthermore, as a non-linear extension of PLS, it is entirely appropriate for underdetermined systems where the number of independent variables far exceeds the number of compounds, which is typically the case when the independent variables are Canvas fingerprint bits.
Figure 4 illustrates the ability of fingerprint-based KPLS to accurately predict potency. Here, a set of 71 compounds with measured inhibition against CDK2 was randomly divided into a training set of 54 and a test set of 17 compounds. Canvas dendritic fingerprints were then generated and used to build KPLS models containing between 1 and 5 factors. As seen in this plot, the 5-factor model performs well on the test set compounds over about four orders of magnitude in IC50.
Figure 4: Scatter plot for a 5-factor KPLS model of CDK2 inhibition using Canvas dendritic fingerprints.
Because a Canvas fingerprint bit can be mapped back to the atoms in the associated fragment, the contribution a given bit makes to the predicted activity can be divided equally over the atoms in the fragment. The non-linearity of KPLS makes it impossible to associate a single coefficient with a single bit, but sensitivity analysis may be employed to assess how the predicted activity ŷ responds to a small change in an “on” bit value: 1→1+dx. The change in predicted activity dŷ—based on that small change to the “on” bit value (1→1+dx)—is then used to compute the sensitivity dŷ/dx which is divided equally among the atoms in the fragment. This exercise is repeated for all bits, with accumulation of sensitivities over the atoms in the structure.