Uncovering Hidden PFAS with Machine Learning

With over 700 PFAS features identified and 81 previously unknown compounds, machine learning reveals the scale of global contamination.

PFAS persists in the environment and poses serious health risks, including liver and kidney damage, developmental problems, and an increased risk of cancer. While regulatory bans have targeted PFAS like PFOA and PFOS, industries have substituted them with new PFAS, many of which remain unidentified due to proprietary trade secrets and environmental transformations.

Traditional methods for identifying PFAS in environmental samples focus on targeted screening, which misses unknown compounds. To fix this, Wang et al. created APP-ID, a machine learning-enhanced molecular network. It uses high-resolution mass spectrometry (HRMS) data and automated structure prediction to improve PFAS identification. Their study, published in Science Advances, shows that APP-ID detects hundreds of previously unknown PFAS, indicating widespread global exposure to these harmful substances.

You can also read: PFAS Characterization and the Push for Information.

Challenges in Identifying Unknown PFAS

PFAS identification in environmental samples is challenging due to their diversity of chemical structures. While nontarget screening methods using HRMS have been employed to detect unknown PFAS, these approaches assume that new compounds resemble known homologs. However, PFAS structures are becoming increasingly complex and do not always follow expected patterns. Additionally, current identification platforms, such as CFM-ID and MetFrag, rely on databases that contain only a limited number of PFAS structures, making it difficult to detect novel compounds.

To overcome these limitations, researchers designed APP-ID. This automated identification system combines a PFAS-specific molecular network, PFAS_link, and a machine learning–based PFAS identification module, PFAS_ID. This platform improves traditional methods by reducing false positives and enabling the discovery of previously unreported PFAS.

How APP-ID Works

PFAS_link: A Molecular Network for PFAS Identification

APP-ID begins by analyzing HRMS data to construct a PFAS-specific molecular network called PFAS_link. Unlike general molecular networking tools, PFAS_link recognizes PFAS features while filtering out non-PFAS compounds. It employs a novel Flink algorithm, which improves the accuracy of spectral matching by:

1. Removing common fragment matches that are not specific to PFAS.
2. Excluding neutral loss matches, which can cause false identifications.
3. Enhancing the detection of diverse PFAS by incorporating both fragment and neutral loss matching.

PFAS_ID: Machine Learning for Structural Identification

Once PFAS_link isolates potential PFAS candidates, the PFAS_ID module uses machine learning to predict their structures. The model uses 194 known PFAS spectra. These are chemical fingerprints recorded from previously identified PFAS compounds. The system uses a machine learning approach called a support vector machine (SVM) algorithm to learn how specific molecular features correspond to known PFAS structures.

When analyzing new samples, PFAS_ID follows these steps:

1. Retrieves structure candidates from databases like PubChem and an in-house PFAS dataset.
2. Looks for exact chemical matches and considers possible transformations, such as breakdown products or variations of known PFAS.
3. Ranks candidate structures based on their similarity to predicted fingerprints.

When tested on an external dataset, APP-ID correctly identified the top match 66.7% of the time when the molecular formula was known and 58.3% of the time when the formula was unknown.

Discovering Unknown PFAS

To demonstrate the effectiveness of APP-ID, researchers applied it to wastewater samples from a fluorochemical industrial park in China. Their findings were striking:

• A total of 733 PFAS features in wastewater, representing a 24- to 28-fold increase compared to traditional spectral matching methods.
• A total of 31 distinct PFAS classes, including 17 classes not previously reported.
• The molecular network identified 130 PFAS with high confidence, outperforming traditional database search techniques.

Global Implications of PFAS Contamination

To determine if the newly discovered PFAS were already present worldwide, the researchers used MASST, a tool that scans public mass spectrometry databases. Their search revealed 126 PFAS signals in samples from 20 countries, including the U.S., Canada, China, India, and several European nations.

Shockingly, 81 of these PFAS were previously unreported, indicating widespread, undocumented exposure. These findings emphasize the urgent need for further investigation and regulatory action to assess the risks posed by these unknown compounds.

The Future

Wang et al.’s study marks a significant advancement in PFAS research, integrating machine learning with molecular networking. APP-ID identifies both known and previously unrecognized PFAS, highlighting more widespread global contamination. The researchers call for expanded PFAS databases, regulatory action, and further use of APP-ID in human biomonitoring studies.