Chemogenomic Active Learning's Domain of Applicability on Small, Sparse qHTS Matrices: A Study Using Cytochrome P450 and Nuclear Hormone Receptor Families

Abstract

Computational models for predicting the activity of small molecules against targets are now routinely developed and used in academia and industry, partially due to public bioactivity databases. While models based on bigger datasets are the trend, recent studies such as chemogenomic active learning have shown that only a fraction of data is needed for effective models in many cases. In this article, the chemogenomic active learning method is discussed and used to newly analyze public databases containing nuclear hormone receptor and cytochrome P450 enzyme family bioactivity. In addition to existing results on kinases and G-protein coupled receptors, results here demonstrate the active learning methodology's effectiveness on extracting informative ligand-target pairs in sparse data scenarios. Experiments to assess the domain of the applicability demonstrate the influence of ligand profiles of similar targets within the family.

Applicability Domain of Active Learning in Chemical Probe Identification: Convergence in Learning from Non-Specific Compounds and Decision Rule Clarification

“…The non-probe dataset was made available to the active learning method for training, and the external probe dataset was tested for predictability at each iteration of the fit-predict-pick active learning cycle. This choice of train-test data represents a chemical-centric domain of applicability challenge different from a prior target-centric study [16].…”

“…Further learning events yielded performances that were either unchanged or degraded. While performance in prior experiments [15,16,27,28] suggested that prediction performance monotonically increases with more data, those results were all in retrospective contexts; indeed, Rakers and colleagues [16] found that in biological contexts of simulated de-orphanization, performance could recede with increases in training data. Here, this finding has been replicated in a chemical context.…”

“…A few years ago, Reker et al demonstrated that chemogenomic data of large protein families such as GPCRs and kinases could be dynamically analyzed in such a way that systematic selection of 5–25% of the bioactivity data was sufficient to predict the entire collection of ligand-target interactions with high performance [15]. This method, chemogenomic active learning, was subsequently shown to also be predictive on smaller target families with highly sparse annotation and further assessed for its domain of applicability by prediction on individual target bioactivity profiles withheld from available training data [16]. In both studies, the underlying model technique used was the Random Forest [17], an alternative to the SVM and DNN model fitting methods, which is also capable of detecting non-linear relationships.…”

“…More generally, active learning platforms are composed of two components, namely an estimator (model) computation component (e.g., RF/SVM/ANN) and a data selection component that uses the results of a fitted estimator in order to pick an instance (or instances) from the available data which will then be included in subsequent fit-predict-pick cycles. Optionally, an external prediction set can be evaluated at each model update, as done previously [16] and in this work. Examples to include in the initial estimator fit could be randomly or intentionally selected (e.g., one ligand-target interaction and one non-interaction).…”

“…While the biological endpoint withholding and prediction study [16] clarified active learning’s domain of applicability from biological and pharmacological points of view, the compounds included in the chemogenomic data were filtered only to remove any inconsistencies in activity annotations. Thus, while potency was considered, the impact of compound promiscuity on modeling results, and therefore, on chemogenomic active learning’s domain of applicability in chemical biology, has yet to be fully explored.…”