Practical application of the Average Information Content Maximization (AIC-MAX) algorithm: selection of the most important structural features for serotonin receptor ligands

Abstract

The Average Information Content Maximization algorithm (AIC-MAX) based on mutual information maximization was recently introduced to select the most discriminatory features. Here, this methodology was applied to select the most significant bits from the Klekota-Roth fingerprint for serotonin receptors ligands as well as to select the most important features for distinguishing ligands with activity for one receptor versus another. The interpretation of selected bits and machine-learning experiments performed using the reduced interpretations outperformed the raw fingerprints and indicated the most important structural features of the analyzed ligands in terms of activity and selectivity. Moreover, the AIC-MAX methodology applied here for serotonin receptor ligands can also be applied to other target classes.

Keywords: Fingerprint reduction; Fingerprints; Machine learning; Selectivity studies; Serotonin receptors; Virtual screening.

Fig. 1

One hundred of the most informative KRFP bits (shown as black squares) selected using the AIC-MAX algorithm for each serotonin receptor. The most significant common bits are marked: blue—polarizable nitrogen atoms, green—aromatic systems, red—amide moiety. Two highly specific fragments that are typical of individual receptors are shown in orange circles (phenylsulfonylamide for 5-${\text{HT}}_6\text{R}$ and o-metoxyphenyl for 5-${\text{HT}}_{1\mathrm{A}}\text{R}$). (Color figure online)

Fig. 2

One hundred (per one ‘off-target’) of the most informative bits (shown as black squares) from KRFP selected using the AIC-MAX algorithm for the 5-${\text{HT}}_{1\mathrm{A}}$ receptor to discriminate its ligands from compounds that act on different serotonin receptors. The most significant common bits are marked: blue—polarizable nitrogen atoms, green—aromatic systems. (Color figure online)

Fig. 3

Comparison between Mathews Correlation Coefficients values obtained in random forest experiments for raw (white background in panel a) and reduced fingerprints (grey background in panel a). Panel b shows when the reduced representation outperformed in conducted experiments the raw one ‘+’, vice versa ‘–’ or no changes ‘nc’. (Color figure online)