Professor Eaton’s description [see below] of the extremely complicated issues identified ~30 years ago — further underscores the need for new technology such as machine-learning to try to advance research knowledge in the PFAS field (as described in the original article from the Belcher lab).
Essentially, machine-learning is best suited for problems where: [a] there is a large amount of data available to train the models; [b] the problem involves identifying patterns and making predictions; [c] the solution can be improved over time as more data become available; and [d] the solution doesn’t need to be fully explainable, or even an approximate solution can be acceptable.
DwN
From: David Eaton
Sent: Wednesday, May 28, 2025 6:40 PM
Dan-
Many, many years ago I did some consulting for 3M on the potential health risks of PFOA and PFOS (long before the controversies over PFAS chemicals). In those days there was a reasonable amount of tox data on PFOA and PFOS (not on the many other PFAS), most of it from animal studies, but there was also quite a bit of occupational data from people who worked with them on a daily basis. Workplace monitoring ‘in the day’ found essentially no evidence of any kind of adverse effects, even though workers had huge blood concentrations, when compared to blood concentrations in animal studies.
The half-life of PFOA in rats was measured in hours, with about a 10-fold sex difference (males had much faster renal elimination compared to females). In a relatively recent PBPK (Physiologically Based Pharmacokinetic modeling — a type of mathematical modeling used to predict how a drug or chemical will behave within a human or animal body) study of PFOA in Sprague-Dawley rats, the half-life after intravenous administration in males was 250 hrs, whereas in females it was ~ 5 hrs (See 2nd attached paper from NIEHS investigators). But the half-life in humans (albeit determined based on long-term repeated exposure and not single-dose data) is measured in YEARS (2-4 years), with no noticeable sex differences, but some age-related effects, with longer half-lifes, as you get older (See 3rd attached article).
Although neither of these papers, nor the paper you sent around, spent a lot of time talking about these important differences between rats and humans (and between male vs. female rats), most of the differences between rats and humans likely are related to differences in protein-binding — making the Starnes et al., paper you sent around highly relevant. The sex differences in rats seems to be largely related to differences in renal reuptake, based on sex differences in the expression of OAT transporters (SLC22 family) in renal tissues.
This species difference created a huge problem in human-based risk assessment 30 years ago — when the only data available were from rat studies, plus the occupational studies, which of course were hard to extrapolate to the general human population (e.g., the ‘healthy worker effect’ may have missed many possible adverse outcomes from chronic exposures).
This is part of the reason why I have avoided getting into the PFAS controversy. Thirty years ago, based on animal data and human occupational health data, there didn’t seem to be any significant concerns with exposures — less than what routinely occurs in the occupational environment, with no evident effects in workers with years of exposure and very high blood levels. This is a great example of ‘G x E interactions’, given that the species and gender differences in pharmacokinetics (PK) of many (but not all) PFAS — enormously complicates any attempts to extrapolate rodent data to humans.
Dave
Dean of Science, Univ Washington Medical Center, Seattle [retired]
From: Nebert, Daniel (nebertdw)
Sent: Wednesday, May 28, 2025 5:39 PM
Per- and polyfluoroalkyl substances (PFAS) are a diverse class of synthetic organofluoride chemicals that are used widely in commercial and industrial applications. [Recently, these GEITP pages had a long email chain of discussion about these “forever chemicals”.] Owing to the strength of their carbon-fluorine bonds, PFAS and their terminal breakdown products are resistant to enzymatic and thermal degradation, and thus accumulate in the environment. The persistence, mobility, and worldwide use of PFAS — since the 1950s — have resulted in the ubiquitous detection of these chemicals in many environmental matrices, and in humans, plants, and wildlife.
As discussed in these GEITP pages, the toxicity of some PFAS — targets several organ systems (hepatic, renal, immune, endocrine, etc.) and has caused escalating public and regulatory concern regarding the adverse health impacts resulting from exposure. Although PFAS comprise a large heterogenous class of thousands of chemicals, research has predominantly focused on a small subset of perfluoroalkyl acid congeners, whereas the fundamental physiochemical properties of several thousand PFAS remain undefined.
Due to major physiochemical differences between the highly fluorinated PFAS and other persistent organic pollutants, which tend to be lipophilic, data and computational models describing the transport and fate of other organic pollutants do not adequately provide us with accurate data about PFAS toxicokinetics. As amphiphilic compounds [chemical compounds possessing both hydrophilic (water-loving, polar) and lipophilic (fat-loving, nonpolar) groups], many PFAS are bound by proteins and accumulate in highly perfused protein-rich tissue compartments (e.g., blood, liver, and kidneys). Protein–PFAS interactions are thus a critical variable for computationally determining transport, bioaccumulation, and half-life of PFAS.
Human serum albumin (HSA) is the primary transport protein in human blood for many endogenous and exogenous ligands, including many PFAS. Increased HSA–PFAS binding affinities are associated with increased circulatory half-life, delayed elimination, and decreased plasma-free fraction of PFAS in the body. Whereas an increased HSA–PFAS binding affinity leads to increased bioaccumulative potential of the PFAS, however, the unbound or free fraction of PFAS in the plasma is the more bioavailable fraction — which is often associated with greater likelihood of tissue transport, reactivity, and toxicity. Authors [see attached pdf file] believe that accurate measures of albumin–PFAS binding might help improve the accuracy of pharmacologically-based pharmacokinetic (PBPK) models, and other computational models aimed at elucidating PFAS toxicokinetics, which remain poorly understood.
Yet, experimentally-determined HAS-binding affinity estimates exist for only a small number of PFAS congeners. Although several advanced computational approaches hold great potential for predicting PFAS–protein interactions, as well as PFAS bioactivity and toxicity, the efficacy of these models is limited by this lack of basic physiochemical data. One contributing factor for the paucity of PFAS-binding data is that several of the most common approaches for measuring albumin–PFAS binding — including equilibrium dialysis and isothermal titration calorimetry — are expensive, time-consuming, and/or require large volumes of protein, making them insufficient for large-scale determinations of binding affinities for thousands of PFAS congeners.
To address the need for directly comparable data describing albumin binding of PFAS, these authors [see attached; the senior author Scott Belcher is a GEITP-er, so if anyone has questions, fire away — directly at him ] — developed an in vitro differential scanning fluorimetry (DSF) assay to determine relative binding affinities of HSA to structurally informative sets of different PFAS. [The DSF assay measures concentration-dependent changes in the melting temperature (TM) of HSA with increasing concentrations of a PFAS, and those data are then used to estimate binding affinity (Ka) values.] Using this DSF assay, authors have determined relative PFAS-binding affinities in serum albumin from humans, as well as from laboratory and livestock animals. However, due to the vast number of PFAS in production and use (with a least 14,000 PFAS chemical structures currently identified), there is a critical need for high-throughput methodologies to facilitate the rapid evaluation of protein–PFAS interactions. This is where the AI subset, machine-learning, steps in. Machine-learning approaches are uniquely positioned to address these data gaps, because machine-learning is a powerful, artificial intelligence (AI)-based tool used to automate analytical model building, and to allow for robust pattern identification in complex datasets.
To address the lack of experimental binding data for thousands of PFAS congeners, authors developed a stepwise, high-throughput approach in which [i] DSF is used to calculate HSA-binding affinities for diverse PFAS, [ii] DSF binding data is then used to train and optimize machine-learning algorithms to build: [iii] optimized classification models to predict binary binding (yes/no) of PFAS, and [iv] optimized regression models to predict HSA–PFAS binding affinities.
In this study [see attached], authors first used DSF to determine HSA-binding affinities for a curated binding affinity dataset of 65 PFAS from 20 chemical classes. DSF binding-affinity data for 3 fatty acids — representing non-fluorinated structural analogs of some perfluoroalkyl carboxylic acids (PFCAs), were also included in the training data for machine-learning models (these fatty acids were included to further assess the impact of fluorination on HSA-binding of PFAS). Authors trained 7 machine-learning models to make binary binding predictions (yes/no binding), and 9 regression-based models were trained to make quantitative binding-affinity (Ka) predictions. The optimally performing binary classifier and quantitative regression models were ultimately used to generate binding predictions for hundreds of PFAS in the EPAPFASINV chemical inventory list, allowing the direct comparison of HSA binding across hundreds of diverse PFAS congeners.
The top-performing classifier model, logistic regression, had an AUROC (area under the receiver operating characteristic curve) statistic of 0.936. The top-performing regression model, support vector regression, had an R2 of 0.854. These top-performing models were then used to predict HSA–PFAS binding for chemicals in EPA’s Computational Toxicology Chemicals Dashboard (EPAPFASINV) list of 430 PFAS. These data — which developed new in vitro and in silico methodologies — represent a high-throughput framework for predicting protein–PFAS binding based on empirical data, and generate directly comparable binding data of potential use in predictive modeling of PFAS bioaccumulation and other toxicokinetic endpoints. ☹
DwN
Toxicol Sci Jan 2025; 203: 67–78