Small Random Forest Models for Effective Chemogenomic Active Learning

Abstract

The identification of new compound-protein interactions has long been the fundamental quest in the field of medicinal chemistry. With increasing amounts of biochemical data, advanced machine learning techniques such as active learning have been proven to be beneficial for building high-performance prediction models upon subsets of such complex data. In a recently published paper, chemogenomic active learning had been applied to the interaction spaces of kinases and G protein-coupled receptors featuring over 150,000 compound-protein interactions. Prediction models were actively trained based on random forest classification using 500 decision trees per experiment. In a new direction for chemogenomic active learning, we address the question of how forest size influences model evolution and performance. In addition to the original chemogenomic active learning findings that highly predictive models could be constructed from a small fraction of the available data, we find here that that model complexity as viewed by forest size can be reduced to one-fourth or one-fifth of the previously investigated forest size while still maintaining reliable prediction performance. Thus, chemogenomic active learning can yield predictive models with reduced complexity based on only a fraction of the data available for model construction.

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

“…These emerging methods are collectively known as chemogenomic active learning. Rakers, Reker, and Brown further demonstrated that model complexity built on qHTS data could be reduced by more than half of existing estimates.…”

“…The subsequent CGAL model development is based on a collection of decision trees known as a random forest that are actively trained using a specified number of iterations or data samples. Based on previous analyses of the tradeoff between the number of trees and resulting performance, the number of trees used here was fixed at 100 …”

“…Overall, the interaction spaces of CYPs and NHRs, comprising 6432 and 10 601 data points, respectively, were successfully modeled in chemogenomic active learning experiments based on only a fraction of available data. In comparison with previous studies that were conducted based on more than 100 000 protein–ligand interaction samples, results presented here indicate applicability of active learning even in sparse data scenarios.…”

“…The descriptor choice was based on previously published comparative experiments . Model training was initiated by random selection of one CPI and one non‐CPI (Figure ) and, based on previous research, models were built using 100 trees with sqrt(4496)=67 descriptors considered per tree node split. Training was terminated after 5000 iterations.…”

Classifiers and their Metrics Quantified

“…Computational experiments may consider the philosophy discussed herein to perform repeated executions of a model‐predict experiment such that less than half of the data per class is subsampled for model selection with the majority remainder used for prediction. Recent modeling methods have shown that often only a fraction of a dataset is sufficient to build a predictive model ,,. As in the prior studies, if distribution of prediction performances can be shown to be normally distributed by the Kolmogorov‐Smirnov test, we can consider using such a fact to forecast the chances of success in a true prospective experiment.…”