Profiling Prediction of Kinase Inhibitors: Toward the Virtual Assay

Abstract

Kinome-wide screening would have the advantage of providing structure-activity relationships against hundreds of targets simultaneously. Here, we report the generation of ligand-based activity prediction models for over 280 kinases by employing Machine Learning methods on an extensive data set of proprietary bioactivity data combined with open data. High quality (AUC > 0.7) was achieved for ∼200 kinases by (1) combining open with proprietary data, (2) choosing Random Forest over alternative tested Machine Learning methods, and (3) balancing the training data sets. Tests on left-out and external data indicate a high value for virtual screening projects. Importantly, the derived models are evenly distributed across the kinome tree, allowing reliable profiling prediction for all kinase branches. The prediction quality was further improved by employing experimental bioactivity fingerprints of a small kinase subset. Overall, the generated models can support various hit identification tasks, including virtual screening, compound repurposing, and the detection of potential off-targets.

Kinome‐Wide Profiling Prediction of Small Molecules

“…For model building the data set "Combined" from Merget et al [5] was updated with data from the most recent ChEMBL [33] version 22 and extended by recently published kinase profiling data from Christmann-Franck et al [14] Kinases with less than 20 inactive or active data points were removed, leaving 305 kinases, 48 953 compounds and 1324 732 data points. The pIC 50 values were binarized with ac utoff at 6.3.…”

“…Encouragingly,t he PCM model using the Z3 protein descriptor not only allows prediction for the entire kinome (i.e.,a lso for kinases without experimental profiling data), but also improvest he prediction powerf or the kinases for which classical QSAR models can be obtained (i.e., % 280 kinasesw ith sufficient experimental bioactivity values) [5] (Figure 2b,c, Figure 4). [29,30] This observationi si nl ine with resultsr eported on serine proteases.…”

“…[4] Accordingly,s everal attempts to predict bindingp rofiles exist, including ar ecent study by us where we described the strength of ligand-based activity prediction models for over 280 kinases by employing machine learning models on an extensived ata set of proprietaryd atac ombinedw ith open data. [5] Here, we report prediction capabilitieso fp roteochemo-metric (PCM) models, which are trained on an updated version of this data set. PCM models combine compound and target descriptors and are, thereby,a blet op redict the activity of compounds for the entire kinome (i.e.,i ncluding kinases without experimental profiling data).…”

“…[6] PCM models have already been successfully appliedt oavarietyo ft argets [7][8][9][10] including kinases, [11][12][13][14] but none of these studies aimed to profile for the entire kinomei ncluding cancer-related mutants.The PCM modelsw ere trainedu sing ar andom forest classifier as described in detail in Merget et al [5] The pIC 50 cutofft o distinguish active from inactive compounds was set at 6.3 (corresponds to an IC 50 value of 500 nm), which is an adequate thresholdt oi dentify,a part from potentialh its, also weaker offtarget interactions. [15] Compound descriptors were calculated using Morgan fingerprints, which generally exhibit very good performance in variousc heminformatics tasks; [5,[16][17][18] protein descriptors fort he binding site residues weree ncoded via zscales (Z3), which are widely used in the peptide design and PCM fields. [7,8,[19][20][21] The modelsw eree valuated via af ive-fold cross-validation (CV) in af our-stage validation scheme ( Figure 1).…”

“…Thus, compared with the individual kinase QSAR models described in ref. [5],t he null model was trained on profiling data for % 300 kinases and an on-physicochemical description of the target.F urthermore, prediction modelsw ere trained individually for all eight kinaseg roups to test whether the consideration of only closely relatedk inases would improve the prediction power.…”