The Dirt on "Forever Chemicals" (And Why We've Been Measuring Them Wrong)

Here's a question that should bother you: what if the billions we're spending to clean up contaminated soil is based on numbers that don't mean what we think they mean?

PFAS - those tenacious synthetic compounds nicknamed "forever chemicals" because their carbon-fluorine bonds laugh in the face of degradation - are everywhere. In your soil, your water, probably your blood. There are over 15,000 varieties, they've been in production since the 1940s, and they're turning up in agricultural fields, military bases, and suburban backyards across the globe. Regulators have been scrambling to set safe limits, and here's the catch: almost every soil standard on the books is based on total PFAS concentration. As in, every last molecule in that handful of dirt.

The problem? Your gut doesn't extract every last molecule. Not even close.

Your Stomach Is a Lousy Chemist (And That's Actually Good News)

Enter bioaccessibility - the fraction of a contaminant that actually dissolves during digestion and becomes available for your body to absorb. Think of it this way: if contaminated soil were a buffet, bioavailability is what ends up on your plate, and bioaccessibility is what makes it past the sneeze guard. Total concentration? That's counting every dish in the kitchen, including the ones nobody ordered.

Yang et al. (DOI: 10.1021/acs.est.6c01678) just published work in Environmental Science & Technology that tackles this gap head-on. They built standardized in vitro methods - essentially test-tube stomachs and intestines - to measure how much PFAS from soil actually becomes bioaccessible. Then they trained machine learning models to predict bioaccessibility across different soil types and PFAS compounds, and used species sensitivity distributions (SSDs) to derive soil environmental criteria based on what's biologically relevant rather than what's merely present.

It's the difference between knowing how much beer is in the keg versus how much you actually drank. Regulators have been counting the keg.

Machine Learning Does the Heavy Lifting (So Lab Rats Don't Have To)

Traditionally, figuring out bioavailability meant feeding contaminated soil to animals and measuring what showed up in their blood. It's expensive, slow, ethically fraught, and about as scalable as hand-delivering mail. The in vitro methods this team developed simulate human digestion in the lab - saliva, stomach acid, intestinal fluids, the whole gastrointestinal greatest hits album - without a single animal.

But running these assays for every PFAS compound across every soil type is still a massive undertaking. That's where the ML models come in. By training on measured bioaccessibility data, the models can predict outcomes for untested combinations of soil properties and PFAS characteristics. Previous work applied this same approach to heavy metals like cadmium and lead (DOI: 10.1021/envhealth.4c00035), and separate teams have built ML tools predicting PFAS sorption behavior in soils (DOI: 10.1021/acs.est.4c13284). Yang's team connects these threads specifically for setting regulatory criteria - the numbers that determine whether a site gets remediated or gets a shrug.

Why This Matters More Than You Think

With nearly 350 PFAS bills considered across 39 U.S. states in 2025 alone, and EPA maintaining drinking water limits at 4 parts per trillion for PFOA and PFOS, the regulatory landscape is a pressure cooker. States like New Jersey have already set interim soil remediation standards, and billions in cleanup costs hang in the balance.

If bioaccessibility-based criteria show that only 30-50% of soil PFAS is actually available for uptake - as prior work on 33 soils suggests (PMID: 37740396) - then current total-concentration standards may be dramatically overestimating risk. That doesn't mean contaminated sites are safe. It means we could direct resources more intelligently, prioritizing sites where PFAS is genuinely entering food chains over sites where it's locked in soil particles doing nothing particularly sinister.

For farmers dealing with PFAS in biosolid-amended fields, this distinction isn't academic. It's the difference between losing a livelihood and getting a science-based all-clear.

The Fine Print (Because There's Always Fine Print)

This is still early-stage framework development. The ML models need validation across diverse geographies and PFAS subtypes. In vitro methods, while promising, produce different results depending on which protocol you use - the field hasn't fully standardized yet. And species sensitivity distributions are only as good as the toxicity data feeding them, which for many PFAS compounds remains sparse. A recent predictive SSD framework for emerging contaminants (DOI: 10.1021/acs.est.4c12654) highlighted that marine and freshwater species respond differently to PFAS, adding another layer of complexity.

But the direction is right. Moving from "how much is in the dirt" to "how much actually gets into organisms" is the kind of boring-sounding refinement that quietly reshapes entire regulatory frameworks. If you're into mind-mapping complex regulatory and scientific workflows, tools like mapb2.io can help visualize how bioaccessibility, ML prediction, and SSD analysis connect in a framework like this one.

References

1. Yang, X., Ren, M., Juhasz, A., Kong, Y., Zhou, P., Song, C., & Cui, X. (2026). Development of Bioaccessibility-Based Soil Environmental Criteria for PFAS through the Establishment of In Vitro Methods and Machine Learning Models. Environmental Science & Technology. DOI: 10.1021/acs.est.6c01678

2. Juhasz, A. L., et al. (2024). In vitro modeling of PFAS bioaccessibility in soil and house dust. Toxicological Sciences, 197(1), 95. PMID: 37740396

3. Predicting bioaccessibility of soil Cd, Pb, and As with machine learning. Environment & Health (ACS), 2024. DOI: 10.1021/envhealth.4c00035

4. PFASorptionML: Modeling PFAS sorption in soils. Environmental Science & Technology, 2025. DOI: 10.1021/acs.est.4c13284

5. Predictive framework for species sensitivity distribution curves of emerging contaminants. Environmental Science & Technology, 2025. DOI: 10.1021/acs.est.4c12654

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.