dawson2023: Machine Learning PFAS Half-Life Predictions from Dawson et...
In httk: High-Throughput Toxicokinetics

dawson2023

R Documentation

Machine Learning PFAS Half-Life Predictions from Dawson et al. 2023

Description

Dawson et al. (2023) Supplemental Information S3 includes half-life predictions for 6603 PFAS, of which 3890 are estimated to be within the applicability domain (AD) for humans. This machine learning (ML) model predicts PFAS half-life as one of four categories. The ML model was trained to a dataset of 91 in vivo measured TK half-lives across 11 PFAS, 4 species, and two sexes. Predictions were a function of compound-specific physico-chemical descriptors, species-specific physiological descriptors, and an indicator variable for sex. The kinetics of PFAS are thought to be complicated by active transport, both through either proximal tubular resorption (into the blood) (Andersen et al. 2006) or secretion (into the urine) (Kudo et al. 2002). The ML model uses several species- and structure-derived surrogates for estimating the likelihood of active PFAS transport. Geometry of the proximal tubule was a surrogate for transporter expression: since secretion/resorption transporters line the surface of the proximal tubule, the amount of surface area provides an upper limit on the amount of transporter expression. PFAS similarity to three distinct endogenous ligands was considered as a surrogate for transporter affinity.

Usage

dawson2023

Format

data.frame

Details

The Dawson et al. (2023) half-life categories are:

Category	Range of Half-Lives
1	< 12 hours
2	< 1 week
3	< 2 months
4	> 2 months

The data.frame contains the following columns:

Column Name	Description
DTXSID	CompTox Chemicals Dashboard substance identifier
Species	Species for which the prediction was made
Sex	Sex for which the prediction was made
DosingAdj	Route of dose administration -- intravenous, oral, or other
ClassPredFull	The predicted half-life class (category)
ClassModDomain	AD estimated from chemical classes of training set
AMAD	AD including AD predicted for each model used for descriptors