Extreme Gradient Boosting algorithm classification for predicting lifespan-extending chemical compounds

Abstract

Human ageing has a great impact on global economy and society’s health with the risk factors for many chronic diseases. Discovery of the pharmaceutical interventions with the potential of promoting longevity and delaying the onset of age-associated diseases is one of the most challenging tasks in anti-ageing research today. The aim of this study was to build a new machine learning model based on the data of the DrugAge database to predict whether a chemical compound will extend the lifespan of the worm species Caenorhabditis elegans. The predictive models were built using the optimized Extreme Gradient Boosting algorithm with molecular fingerprints and molecular descriptors as features. The ranking of the models’ features was done with the built-in Extreme Gradient Boosting feature importance function and interpreted with confidence using Shapley values. The top 15 most important features included 2D molecular descriptors related to the subdivided surface areas, atom and bond counts, and electrostatic properties. The best performing model was applied to predict the class of compounds in the external database, DrugBank, consisting of approved small-molecules. The chemical compounds of the external database with a predictive probability of for increasing the lifespan of Caenorhabditis elegans were broadly separated into (i) flavonoids and isoflavonoids, (ii) fatty acids and conjugates, and (iii) other classes of compounds.

Predicting Lifespan-Extending Chemical Compounds with Machine Learning and Biologically Interpretable Features

“…Thus, all predictive features in our datasets are related to the proteins that interact with each compound, namely: the protein interactors themselves, the GO Term annotations and the Physiology Phenotypes associated with those interacting proteins, and whether the protein interactors of a compound are coded by an ageing-related gene. Notably, two of the four most related works [17] [19] [18] [20] (applying machine learning to DrugAge data) also use GO Term features, namely [17] and [18], but none of those four works used protein interactors, phenotypes or ageing-related genes as predictive features.…”

“…Hence, a promising research direction consists of analysing the data in such databases using machine learning algorithms that highlight patterns in data, particularly classification algorithms, which learn predictive models from data [16]. Therefore, recently there has been growing interest on applying classification algorithms to the data in DrugAge [17] [18] [19] [20], in order to learn models that predict which compounds are more likely to extend the lifespan of a given organism, which is also the overall goal of this work.…”

“…The novel contributions of our study compared to the related work in [17] [18] [19] [20] are as follows. First, we created datasets using four different types of predictive features, based on Gene Ontology (GO) terms, drug (compound)-protein interactions, interactions between compounds and proteins encoded specifically by ageing-related genes, and physiology terms from a Phenotype Ontology for C. elegans .…”

“…GO terms have been used as predictive features in [17] [18], but the other three types of features proposed here are new types of features for predicting a compound’s effect on the lifespan of an organism using machine learning, to the best of our knowledge. Note that we do not use chemical descriptors as features, an approach used in [17] [18] [19] [20], which generated models with good predictive accuracy. We do not use chemical features because they represent very specific chemical information which is not very meaningful for biogerontologists.…”

Predicting lifespan-extending chemical compounds for C. elegans with machine learning and biologically interpretable features

“…Hence, a promising research direction consists of analysing the data in such databases using machine learning algorithms that highlight patterns in data, particularly classification algorithms, which learn predictive models from data [16]. Therefore, recently there has been growing interest on applying classification algorithms to the data in DrugAge [17][18][19][20], in order to learn models that predict which compounds are more likely to extend the lifespan of a given organism, which is also the overall goal of this work.…”