Prediction of PARP Inhibition with Proteochemometric Modelling and Conformal Prediction

Abstract

Poly(ADP-ribose) polymerases (PARPs) play a key role in DNA damage repair. PARP inhibitors act as chemo- and radio- sensitizers and thus potentiate the cytotoxicity of DNA damaging agents. Although PARP inhibitors are currently investigated as chemotherapeutic agents, their cross-reactivity with other members of the PARP family remains unclear. Here, we apply Proteochemometric Modelling (PCM) to model the activity of 181 compounds on 12 human PARPs. We demonstrate that PCM (R0 (2) test =0.65-0.69; RMSEtest =0.95-1.01 °C) displays higher performance on the test set (interpolation) than Family QSAR and Family QSAM (Tukey's HSD, α 0.05), and outperforms Inductive Transfer knowledge among targets (Tukey's HSD, α 0.05). We benchmark the predictive signal of 8 amino acid and 11 full-protein sequence descriptors, obtaining that all of them (except for SOCN) perform at the same level of statistical significance (Tukey's HSD, α 0.05). The extrapolation power of PCM to new compounds (RMSE=1.02±0.80 °C) and targets (RMSE=1.03±0.50 °C) is comparable to interpolation, although the extrapolation ability is not uniform across the chemical and the target space. For this reason, we also provide confidence intervals calculated with conformal prediction. In addition, we present the R package conformal, which permits the calculation of confidence intervals for regression and classification caret models.

Elucidating Compound Mechanism of Action and Predicting Cytotoxicity Using Machine Learning Approaches, Taking Prediction Confidence into Account

“…The R programming language, which has proved a versatile framework for medicinal chemistry applications (Kuhn, 2008;Mente & Kuhn, 2012), will be used to generate the models. Given the increasing awareness of the importance of the uncertainty of the data for bioactivity modeling, we show how to generate individual errors of prediction calculated using the conformal prediction framework (Cortés-Ciriano, Bender, & Malliavin, 2015aNorinder, Carlsson, Boyer, & Drakakis et al Given the increasing awareness of the importance of the uncertainty of the data for bioactivity modeling, we show how to generate individual errors of prediction calculated using the conformal prediction framework (Cortés-Ciriano, Bender, & Malliavin, 2015aNorinder, Carlsson, Boyer, & Drakakis et al …”

“…We illustrate how individual models can then be integrated into more predictive meta-models using ensemble modeling, thus leading to more accurate predictions (Caruana et al, 2004;Cortés-Ciriano et al, 2014). Given the increasing awareness of the importance of the uncertainty of the data for bioactivity modeling, we show how to generate individual errors of prediction calculated using the conformal prediction framework (Cortés-Ciriano, Bender, & Malliavin, 2015aNorinder, Carlsson, Boyer, & Drakakis et al…”

Chemoinformatic Classification Methods and their Applicability Domain

“…Comprehensive benchmarks concerning the efficiency of nonconformity measures for chemoinformatics applications are missing thus far. Finally, conformal prediction is not restricted to classification but can also be applied to regression problems 117,119,123…”

Kinome‐Wide Profiling Prediction of Small Molecules

“…This is possible, because PCM modelsa re able to interpolate to new compounds for known targets and vice versa, as well as to extrapolate to new kinases for new compounds. [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.…”

“…This is possible, because PCM modelsa re able to interpolate to new compounds for known targets and vice versa, as well as to extrapolate to new kinases for new compounds. [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.…”