Prospectively Validated Proteochemometric Models for the Prediction of Small-Molecule Binding to Bromodomain Proteins

Abstract

The bromodomain-containing proteins are a ligandable family of epigenetic readers, which play important roles in oncological, cardiovascular, and inflammatory diseases. Achieving selective inhibition of specific bromodomains is challenging, due to the limited understanding of compound and target selectivity features. In this study we build and benchmark proteochemometric (PCM) classification models on bioactivity data for 15,350 data points across 31 bromodomains, using both compound fingerprints and binding site protein descriptors as input variables, achieving a maximum performance as measured by the Matthew's Correlation Coefficient (MCC) of 0.83 on the external test set. We also find that histone peptide binding data can be used as a target descriptor to build a high performing PCM model (MCC 0.80), showing the transferability of peptide interaction information to modeling small-molecule bioactivity. 1,139 compounds were selected for prospective experimental testing by performing a virtual screen using model predictions and implementing conformal prediction, which resulted in 319 correctly predicted compound-target pair actives and the correct prediction for certain selectivity profile combinations of the four bromodomains tested against. We identify that conformal prediction can be used to fine-tune the balance between hit retrieval and hit structural diversity in a virtual screening setting. PCM can be applied to future virtual screening and compound design, including off-target prediction for bromodomains.

An Analysis of Proteochemometric and Conformal Prediction Machine Learning Protein-Ligand Binding Affinity Models

“…Alternatively, PCM models featurize both protein sequence and ligand to create the final feature vector. This allows ML models to be trained on protein-ligand binding affinity data from multiple proteins at once, hence augmenting the size of the training set, and potentially allowing the model to learn from related proteins (Lapins et al, 2008 ; Gao et al, 2013 ; Subramanian et al, 2013 ; Cortés-Ciriano et al, 2015a , b ; Tresadern et al, 2017 ; Giblin et al, 2018 ).…”

“…In contrast proteochemometric (PCM) models combine both protein and ligand information to create a composite feature vector that allows the model to learn mappings between all protein-ligand pairs in a training set (Cortés-Ciriano et al, 2015a ). PCM models have been applied to a diverse number of protein families including G-protein coupled receptors (Gao et al, 2013 ), HDACs (Tresadern et al, 2017 ), kinases (Subramanian et al, 2013 ), Cytochrome P450s, HIV proteases (Lapins et al, 2008 ), Poly(ADP-ribose) polymerases (Cortés-Ciriano et al, 2015b ), and bromodomains (Giblin et al, 2018 ). Recently, PCM and multi-task neural networks have also been benchmarked using ChEMBL data where the utility of PCM modeling for binder/non-binder classification was demonstrated (Lenselink et al, 2017 ).…”

An Analysis of Proteochemometric and Conformal Prediction Machine Learning Protein-Ligand Binding Affinity Models

“…Alternatively, PCM models featurize both protein sequence and ligand to create the final feature vector. This allows ML models to be trained on protein-ligand binding affinity data from multiple proteins at once, hence augmenting the size of the training set, and potentially allowing the model to learn from related proteins [9][10][11][12][13][14][15][16] .…”

“…In contrast proteochemometric (PCM) models combine both protein and ligand information to create a composite feature vector that allows the model to learn mappings between all protein-ligand pairs in a training set 9 . PCM models have been applied to a diverse number of protein families including G-protein coupled receptors 10 , HDACs 11 , kinases 12 , Cytochrome P450s 13 , HIV proteases 14 , Poly(ADP-ribose) polymerases 15 , and bromodomains 16 .…”

Machine Learning for Acute Oral System Toxicity Regression and Classification

“…Conformal prediction defines a machine learning model's applicability domain, i.e the region of chemical space where predictions are reliably accurate [31][32][33][34][35][36][37] . Conformal prediction produces a confidence region where the true value lies with a probability determined by a user specified confidence threshold.…”