Machine learning assisted single-molecule sensing of per- and polyfluoroalkyl carboxylic acids: Quantification without standards

doi:10.21203/rs.3.rs-4603074/v1

Download PDF

Article

Machine learning assisted single-molecule sensing of per- and polyfluoroalkyl carboxylic acids: Quantification without standards

https://doi.org/10.21203/rs.3.rs-4603074/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Per- and polyfluoroalkyl carboxylic acids (PFCA) are of great concern due to their ubiquitous presence in the environment. Despite a severe shortage of authentic standards, compared to the rapid increase of possible structures identified, it remains difficult to quantify a mixture of PFCA without references. Herein, a standard-free single-molecule electrochemical sensing method was developed for the first time by establishing a linear correlation between current blockades and the volumes of PFCA simulated by molecular dynamics. A nearly 100% accuracy was realized for the simultaneous determination of 13 pristine or H- / Cl-substituted PFCA, using frequency-modulated multi-feature classification. Shortlisting the 21 high-priority features reduced the required number of training data by 7.6 folds, and almost 80% quantification reliability was maintained even with interference of 100 times concentration. Moreover, the detection limit of trifluoroacetic acid (an ultrashort-chain PFCA) went down to 57 ng·L^-1, comparable to the state-of-the-art performance.

Earth and environmental sciences/Environmental sciences/Environmental chemistry/Environmental monitoring

Physical sciences/Chemistry/Analytical chemistry/Sensors

Physical sciences/Nanoscience and technology/Nanoscale devices/Nanopores

Per- and polyfluoroalkyl substances (PFAS) contamination has become a global concern due to their widespread environmental occurrenceand elevated human exposure¹. It is a complex chemical class, covering almost all the compounds with at least a perfluorinated methyl (-CF₃) or methylene (-CF₂-) functional group, according to the revised definition by the Organisation for Economic Cooperation and Development (OECD)². Till now, over 14,000 chemicals have been registered in the U.S. EPA PFAS structure list³. Growing evidence indicates that the abundant tiny structural variation in PFAS

can affect the respective partitioning⁴, transferring⁵, elimination⁶, transformation⁷, bioaccumulation⁸ behaviours, and may pose distinct health risk⁹. Hence, it is imperative to determine the concentration of multiple structurally similar PFAS species in various environmental matrices simultaneously.

The current analytical methods for PFAS, however, are hard to achieve precise quantification, high resolution and wide coverage concurrently. While great advances have been made over the past decades to enhance the resolution and detection limit of high-performance liquid chromatography-tandem mass spectrometry (HPLC-MS/MS)^10,11, as well as gas chromatography-mass spectrometry (GC-MS)^12,13, such quantitative analysis relied on the use of reference standard, of which the number of commercially available samples¹⁴ were just over 120, less than 1% of the total PFAS. In contrast, progress in high-resolution mass spectrometry (HRMS)^15,16 opens the possibility to comprehensive, nontargeted screening of unknown PFAS, especially when coupled with ion mobility spectrometry (IMS)^17,18. Nevertheless, the structure identified by HRMS is tentative¹⁹, and the concentration can, at best, be semi-quantified. So far, quantification of PFAS without standards remains challenging.

Herein, a nanopore based single-molecule electrochemical sensor is proposed as an emerging technology to bridge the gap between the urgent demand for quantitative PFAS monitoring and the severe shortage of authentic standards. By measuring the change of ion movement through nanopore, single-molecule sensor (SMS) can identify and quantify a specific analyte with the characteristics and frequency of the resulting current blockade²⁰. A linear correlation was established in this work between the magnitude of ionic current and the volume of PFAS simulated by molecular dynamics (MD), and thus the current response of unknown PFAS could be accurately predicted, avoiding the need for standards. Previous studies have demonstrated positive correlations between the magnitude of current blockade and the volume²¹or mass^22-26 of peptides^21-24 and proteins^25,26 using a variety of protein pores, e.g., α-hemolysin²², aerolysin²¹, FraC²³, ClyA²⁵, CytK²⁴ or YaxAB²⁶, but a strict linear relationship (R² = 0.9998) was realized for the first time, as far as the authors were aware, and the predicted blockade values were almost identical to experimental measurements. More importantly, a custom machine learning algorithm, based on the frequency-modulated multi-dimensional feature extraction, was developed to enhance the structural resolution of SMS, reaching an overall accuracy of 99.9% for a total of 13 per- and polyfluoroalkyl carboxylic acids (PFCA). Further optimisation of feature combination, reducing it from 43 to 21 dimensions, required only 13% of the training set for 99% accuracy. As a result, even under an interference of 100 times concentration, nanopore SMS was able to maintain 78% quantification reliability, over an order of magnitude better than ensemble analysis. Besides, a wide, interference-free linear-response range of 0.5 nM to 100 μM was achieved for trifluoroacetic acid, corresponding to the detection limit of 57 ng·L^-1, which was comparable to the state-of-the-art performance of UPLC-MS/MS²⁷ or GC²⁸ for this ultra-short PFCA.

Establishment of a linear volume-current relationship using perfluoroalkyl carboxylic acids

Building a structure-activity relationship was the prerequisite towards the development of standard-free quantification methods, and thus a linear correlation was established in the first place between the volume of PFCA molecules and the magnitude of current blockade measured by nanopore SMS. Despite many factors were proposed to influence the translocation induced current response, it was possible to compile them into two dominant factors, i.e., steric exclusion, counterion enhancement, or a combination of those two²⁹. As for steric exclusion, the effective volume of nanopore for signal transduction could change dynamically, indicating that the residence position of analyte in nanopore was crucial when sensing small molecules³⁰. Therefore, polycationic peptide probes were employed in this study to control the location of tethered PFCA in nanopore, while concentrated electrolytes were adopted to reduce the contribution of surface charge, so that the magnitude of current blockade was mainly determined by steric exclusion. More specifically, eight linear perfluoroalkyl carboxylic acids (C2 to C9) were chosen as typical PFCA to form the structure-current relation, Fig. 1a. They were connected to the N-terminal of an oligo-arginine leader (PFCA-R₆) and measured by the wild-type aerolysin (WT AeL) in 4 M KCl solution, Fig. 1b. It was found previously that WT AeL formed positive electrostatic barriers on both of its trans exit and cis entry under negative applied voltages³⁶, and so the polycationic -R₆ probe might be able to drive the non-ionic PFCA targets to the identical position within the AeL nanopore. The typical current traces of C2-C9 PFCA-R₆ at -50 mV were shown in Fig. 1c, as well as the R₆ probe (C0). It was clear that both the magnitude of current blockade and dwell time increased with the length of PFCA. Histograms of the current blockade for C0 and C2 to C9 were given in Supplementary Fig. 1.

The resulting magnitude of current blockade for C2-C9 PFCA-R₆ complexes exhibited a strong linear correlation (R² = 0.9998) with their molecular volume (shown as shallow squares in Fig. 1d). The measured differences between the blockade of C2- and C9-R₆ was 11.8%, corresponding to an increase of 1.68% per -CF₂- (ca. 73.5 Å³) or a slope of 0.023%·Å^-3 for this straight line. The effective transduction volume of 4.82 nm³, estimated at the blockade of 100%, was identical to the inner pore volume comprised between the A224 and S236 residues of WT AeL (4.82 nm³), supporting the use of R₆ probe to direct the movement of PFCA targets. A total of 61 individual measurements (at least three for each sample) were conducted to reduce experimental errors, using perfluorohexanoic acid (C6) as an internal standard for calibration³¹. The average of error between different measurements was as low as 0.022%, one order of magnitude smaller than the overall standard deviation (0.198%) for the histogram of current blockade, confirming the reliability of our approach. The hydrodynamic volume of PFCA-R₆ was calculated via all-atom molecular dynamics simulations using GROMACS. More details of the simulation process were summarized in the “Methods” section, and the obtained raw data of volumes for C0 and C2 to C9 were given in Supplementary Figs. 2 and 3.

Accurate prediction of current response of H- / Cl-substituted polyfluoroalkyl carboxylic acids

Following the establishment of a linear correlation between the volume of perfluoroalkyl carboxylic acids and their magnitude of current blockade, the next step was to examine its prediction accuracy for other polyfluoroalkyl carboxylic acids. Among all the 622 carboxylic PFAS structures that were identified previously (CAS numbers available)⁵, 49% of them contained at least one C-H functional group while 5.7% had one or more C-Cl group. Therefore, five typical H- or Cl-substituted analytes, either terminal or internal, were examined in this study, including 3H-tetrafluoropropionic acid (3H), 5H-octafluoropentanoic acid (5H), 7H-dodecafluoroheptanoic acid (7H), 3Cl-tetrafluoropropionic acid (3Cl), and 3:3 fluorotelomer carboxylic acid (FTA), which were increasingly discovered in the wastewater from fluorochemical industry as well as the surrounding surface water^32-34. Based on the MD simulated molecular volume, the predicted current blockades of H- or Cl-PFCA were in perfect accordance with experimental measurements, Fig. 1d and insert, with negligible deviation (0.022%) close to the observational errors, which clearly demonstrated the capability of using nanopore SMS for standard-free identification of PFAS. The current trace, histogram of blockade, and the simulated molecular volume of H- or Cl-PFCA were provide in Supplementary Figs. 4 and 5.

Factors that affected the standard-free prediction and single-molecule identification of PFCA

To further explore the origin of the observed linear correlation, different peptide structures (R₆K-, R₅K-, -R₇ and -R₆) were compared to analyse the role of probe length and orientation of connection. Perfluoropentanoic (C5), perfluorohexanoic (C6) and perfluoroheptanoic acid (C7) were linked to the lysine side chain of R₆K- or R₅K- probes, and to the N-terminals of -R₇ or -R₆. Errors of the actual blockade from prediction was much smaller for C6-R₆ or -R₇ than R₆K- or R₅K-C6 (Fig. 1e), probably due to the narrow lumen of WT AeL (diameter between 1-1.4 nm)³⁵. Meanwhile, the slope of the linear fit for -R₆ probe, i.e., signal sensitivity to the change of analyte volume, was 70% higher than -R₇ (Fig. 1e), suggesting the enlarged transduction volume of the latter. Although showing little improvement in linearity (R² = 0.9995-0.9999 for the linear fit of C5-, C6- and C7-R₆ in 2-4 M KCl, (Supplementary Fig. 6), the elevated salt concentration reduced the standard deviation of blockade from 0.47% for 2 M to 0.165% for 4 M, Fig. 1f, which could triple the resolution of identification. One possible reason was the prolonged dwell time of PFCA in WT AeL, caused by the higher cis-to-trans driving force in 4 M KCl (or the lower trans-to-cis electroosmotic force)²⁰. Nevertheless, it was noted that the sum of the three sigma of poly- and the adjacent per-fluoroalkyl carboxylic acids (highlighted as the error bar in Fig. 1d) were still greater than the difference between their blockade, indicating that the use of current blockade alone was unlikely to fully resolve the total 13 PFCA.

Frequency-modulated multi-dimensional feature extraction for 100% classification accuracy

An inherent advantage of SMS over ensemble methods was the ability to record multi-dimensional features of the signal generated by an individual molecule³⁶. The use of five or more signal features was demonstrated recently as an effective approach to distinguish structurally similar compounds, e.g., achieving 92.4%-99.9% accuracy for the determination of saccharides³⁷, riboses³⁸, alditols³⁹ or benzenediols⁴⁰. Herein, frequency modulation (using five low-pass filters of 2000, 800, 500, 200 or 100 Hz, as well as the wavelet transform) was applied to extend the eight common features of single-molecule signals, i.e., the magnitude (ΔI/I₀), duration (τ_on), standard deviation (I_σ), peak-to-peak (I_pp) of current blockade, and the peak (H_peak), full width at half maximum (H_FWHM), skewness (H_skew), kurtosis (H_kurt) of the all-points histograms of current blockade, to a total number of 43 (τ_on remained constant at all frequencies), Fig. 2a. All the feature inputs were normalized by an internal standard (C6, C5 or C3) and averaged by at least three parallel measurements to minimize experimental errors. The resulting 43 features of 14 analytes were all fitted with a Gaussian distribution (Supplementary Figs. 7 to 20). Interestingly, despite the closely related physical meaning of ΔI/I₀ and H_peak, or I_σ, I_pp and H_FWHM, the resolving power of these features and their frequency dependent behaviours differed remarkably (Supplementary Figs. 21 to 29), implying the possibility to integrate multi-dimensional features for enhanced resolution.

In fact, the combination of eight-dimensional feature (extracted from the raw data at 2000 Hz) increased the identification accuracy from 89.1% (using ΔI/I₀ only) to 96.7% for the 11 short-chain PFCA (excluding C8 or C9), and the further incorporation of frequency modulation (using the total 43 features) pushed it to 99.9%, Fig. 2b. A total of 2400 sets of data were collected for each analyte (2000 for training/validation and 400 for test). In total, 31 classifiers were evaluated, among which the Bagged trees model showed the highest identification accuracy, Supplementary Table 1. All the accuracies were calculated based on five or more repetitions of random holdout test sets, and a 10-fold cross-validation was applied for training. It was worth noting that the order of feature addition, from two to eight dimensions, for achieving the highest identification accuracy (left part of Fig. 2b), was distinct from the rank of their own one-dimensional accuracy (Supplementary Fig. 30), which suggested the significance of correlation and complementarity between features. The enhancement by feature addition reached a plateau when the number of dimensions exceeded four, meaning that certain features were less relevant or easier to replace in the classification of PFCA. Once frequency modulation was included, the identification accuracy quickly jumped over 99% (right part of Fig. 2b). The confusion matrix of all 13 PFCA plus the R₆ probe (also with an overall accuracy of 99.9%) showed that the biggest errors came from the misclassification of FTA with C5 and C6, Fig. 2c, and strikingly, small but non-negligible false identifications (0.01-0.04%) were constantly observed for the pairs of PFCA that could be resolved completely using blockade only (Supplementary Fig. 22). Such phenomena stressed the necessity to reduce model complexity or the number of features.

Shortlisting high-priority features to minimise the size of training set for precise determination

To optimise the combination of model input, the change of maximal validation accuracy against the number of features was examined, under various sizes of training set for C5, C6 and FTA, i.e., 2000, 200 or 20 signals for each PFCA. A similar trend was discovered in all three curves (Fig. 2d): when more features were included, the accuracy increased initially, then reached a plateau, and dropped slightly in the end, indicating that the optimal number of features should be neither too small (e.g., to avoid the deviation in training set) nor too large (e.g., to reduce the model complexity). The height, length and position of the plateau became lower, shorter, and left-shifted for smaller set, which decreased from 99.97% for 2000 signals with 28±10 features, to 98.40% for 20 signals with 21±5 features. In the meantime, the priority of features was also affected by training size: when only 20 signals were used, a general order of importance 2000 Hz < wavelet < 100 Hz < 200 Hz < 500 Hz < 800 Hz was followed; but in the case of 200 or 2000 signals, other features such as kurtosis at 200 and 500 Hz stood out, while all the filtered blockades (either wavelet or 100-800 Hz) were no longer critical, Supplementary Table 2-4. Taking into account of their rank in all three scenarios, a total of 21 features were shortlisted, Fig 2e, which was able to decrease the amount of data for training by 7.6 folds, compared to the full 43-feature model, and to increase the maximal identification accuracy from 99.58% to 99.92%.

Maintaining nearly 80% quantification reliability with interference of 100 times concentration

The aforementioned few-shot learning mode of nanopore SMS facilitated the accurate quantification of trace targets in the presence of structurally similar interferents with much higher concentrations, which was often difficult for ensemble measurements. For instance, with regard to the determination of FTA from more concentrated C5 and C6, the pre-developed 21-feature model was able to maintain over 78% accuracy with interference of 100 times concentration, Fig. 3a, and a substantial 44% was still detectable even under 1000 times. Such performance was at least one order of magnitude better than the multi-peak fit using the 2000 Hz blockade only, which was the best one-dimensional feature in this study to mimic ensemble analysis, showing an accuracy of 69% with 10 times interference and a rapid drop to zero at 20 times. In the meantime, the quantification reliability of the shortlisted 21-feature model was consistently higher than the full 43-feature counterpart as well, in accordance with the above analysis on the influence of feature number.

Interference-free quantification of trifluoroacetic acid towards a detection limit of 57 ng·L^-1

Trifluoroacetic acid (C2) was the simplest form of PFCA, with a much higher environmental level than other PFCA due to its more diverse sources²⁸. Recent evidence suggested that the rising rate of C2 was considerable, and its potential adverse health effects could be higher than expected⁴¹. The quantitative analysis of C2, however, was arduous, and the limit of detection (LOD) was among 10-500 ng·L^-1, Supplementary Table 5^27,28,42-46. Herein, linking the logarithm of C2 concentration and the logarithm of interval time measured at -50 mV and 20℃, a linear response range was established between 50 nM and 100 μM C2, insert of Fig. 3b. Interference such as C4, C5, C6 or tap water caused a tiny deviation less than 1%. Further adoption of an elevated voltage of -80 mV and a temperature of 30℃ pushed the LOD of C2 down to 0.5 nM or 57 ng·L^-1, Fig. 3b, comparable to the state-of-the-art^27,28.

This work realized for the first time standard-free quantification of PFCA through the combination of single-molecule sensing with multi-feature classification. The structure of polycationic probe was fine-tuned so that various PFCA analytes could be directed to the identical residence position within the WT AeL pore, enabling the establishment of a linear volume-blockade relationship. It was noted that difference between the current blockade of C2 and C8 was only 9.662%, but at least 452 PFCA³ (with CAS available) had been identified in this range. Hence, to fully resolve all the PFCA above, the standard deviation of blockade has to be smaller than 0.00356% (R = 1.5)²⁰, which however was around 0.2% at best for WT AeL. It was thus essential to make the most of other features of single-molecule signals.

Though the frequency modulated 43-feature model reached an overall accuracy of 99.58% for FTA/C5/C6, the hardest-to-distinguish combination of PFCAs in this study, when the size of training set was 2000 for each analyte, this value quickly dropped to 93.5% if only 20 signals were sampled. The shortlisted 21-feature model, according to the rank of importance, reduced the required amount of training data by 7.6 folds, enabling accurate quantification of trace targets at high-concentration interference. This was particularly crucial, indicating that a total of 20 signals recorded in 10-minute response time (or a capture rate of 1/30 Hz) was sufficient for precise PFCA determination. Future studies towards the measurements of all the 338 PFCA isomers by SMS would require the rationale engineering of nanopore interface in the context of multiple feature classification.

Reagents and chemicals. Potassium chloride (KCl, ≥99.9%), tris(hydroxymethyl)aminomethane (Tris, 99%), ethylenediaminetetraacetic acid (EDTA, ≥99.0%), and octane (≥99%) were purchased from Titan Scientific (Shanghai, China). The electrolyte adopted in this detection system was the aqueous solution of 4 M KCl, 10 mM Tris, and 1 mM EDTA. 1,2-diphytanoyl-sn-glycero-3-phosphocholine (DPhPC, powder, ≥99%) was purchased from Avanti Polar Lipids (Shanghai, China). DPhPC was dissolved in octane and sealed dryly at -20℃. Proaerolysin was purchased from GenScript Biotech (Nanjing, China), and trypsin agarose (25UN) was purchased from Sigma-Aldrich (Shanghai, China). Conjugates of perfluorocarboxylic acids and peptide leaders were produced by ChinaPeptides (Suzhou, China).

Single-molecule sensing of perfluorocarboxylic acids. Proaerolysin was activated with immobilized trypsin at 4℃ for 10 hours to form monomeric aerolysin, which were involved in the oligomerization for aerolysin nanopores. After activation, proaerolysin was separated from trypsin at 10000 RPM and stored at -20℃ for further use. The above electrolyte was added into the detection chamber and the temperature was set as 20℃ constantly, except other stated. The DPhPC lipid bilayer membrane supporting the nanopore was formed across the 100 μm microcavities of MECA 4 Recording Chips (Nanion Technologies, Germany). Activated aerolysin was added into the cis chamber and formed a single nanopore on the DPhPC lipid bilayer membrane under the applied bias of +200 mV. The cis side of the detection chamber was seen as virtual ground whose potential was 0 mV and the voltage was applied on the trans side. If a single aerolysin nanopore was formed, open pore current would go to -100 pA as the applied voltage was set as -50 mV. PFCA-R₆ conjugates were added into the cis chamber and translocated to the trans side under negative voltages. Sampling rate and bandwidth were set as 20 and 10 kHz, respectively.

Data acquisition and analysis. Electrical signals were recorded with a patch clamp amplifier (Orbit mini). Frequency-modulated signals were processed in MATLAB 2021b, via wavelet denoising or lowpass filter at 100, 200, 500 and 800 Hz. Machine learning models were trained, validated and predicted in MATLAB as well. Features of raw and frequency-modulated signals were extracted in Python 3.11, based on the pre-identified events from Clampfit 10.4. Feature normalization was done in Python. Graphs were drawn in OriginLab, MATLAB or Python.

Molecular modelling and simulation. Three-dimensional structures of the R₆ probe and PFCA-R conjugates were generated in ChemBio3D. Energy minimization of these 3D structures were done with MM2 method and saved as pdb files. Structure topology files were generated with Ligand Reader & Modeler on CHARMM-GUI.

Molecules were simulated and their volumes were calculated in an aqueous solution system. The aqueous solution system, including electrolyte, electric field, temperature and pressure, was modelled and stabilized in advance. A 7*7*7 nm³ sized water box was built through Solution Builder on CHARMM-GUI, with a KCl concentration of 4 M and temperature at 293 K. To eliminate the redundant steric hindrance and minimize the energy of this water box, 5000 iterations were done with the steepest descent algorithm. A 125 ps long equilibrium simulation was carried out to maintain constant temperature and reasonable distribution of total compounds. Based on this finished product, a 5 ns long simulation with a time step of 2 fs was carried out using Nose-Hoover method for temperature control and Parrinello-Rahman method for pressure control. Electrostatic interaction was calculated with Particle mesh Ewald (PME) method. Cut-off radius referring to calculations of Coulomb interaction, electrostatic interaction and Van der Waals interaction was set as 12 Å. Calculations of hydrogen bonds were constrained with LINCS algorithm.

The last frame of the simulation was adopted as the water box for later molecular dynamic simulations. Afterwards, each structure under simulated was inserted into this water box with gmx insert-molecule, replacing water molecules, potassium ions and chloride ions randomly. The whole system was controlled at 293 K with V-rescale method of temperature control. After 10000 iterations with the steepest descent algorithm, energy minimization of this system was achieved. As a final step, there was a 1 ns long NPT equilibrium. Molecular volumes of all kinds of conjugates were calculated in a 50 ns long simulation, with time step of 2 fs. In purpose of reducing the impact of R₆ leader and highlighting real volume differences between different types of PFCA, X-R conjugates (where X represented different types of targets) were designed and simulated at first, putting the volume of X at a dominant position while taking the junction between the target and the leader into consideration. The volume difference between C4-R and C4-R₆ was set as an internal standard to speculate all of the molecular volumes of X-R₆.

Data Availability

Source data are provided with the paper. Other data are available upon request.

Code Availability

The custom MATLAB and Python scripts are available at https://github.com/DEWMEME.

Acknowledgements

This study was supported by the National Key R&D Program of China (2023YFC3008803), the National Natural Science Foundation of China (21972041 and 22006037), the Natural Science Foundation of Shanghai Municipality (23ZR1416300), and the Fundamental Research Funds for the Central Universities.

Competing Interests

The authors declare no competing interests.

Additional Information

Supplementary information. The online version contains supplementary material available at.

Evich, M. G. et al. Per- and polyfluoroalkyl substances in the environment. Science375, eabg9065 (2022).
OECD. Reconciling terminology of the universe of per- and polyfluoroalkyl substances: Recommendations and practical guidance. Series on risk managementNo. 61 (2021). (accessed 14^th May 2024). https://one.oecd.org/document/ENV/CBC/MONO(2021)25/En/pdf.
US EPA. PFAS structures in DSSTox (update August 2022). https://comptox.epa.gov/dashboard/chemical-lists/PFASSTRUCTV5.
Liu, D. et al. Distribution of per- and poly-fluoroalkyl substances and their precursors in human blood. J. Hazard. Mater.441, 129908 (2023).
Wang, Q. et al. Legacy and emerging per- and polyfluoroalkyl substances in a subtropical marine food web: suspect screening, isomer profile, and identification of analytical interference. Environ. Sci. Technol.57, 8355-8364 (2023).
Gao Y. et al. Isomer-specific perfluoroalkyl acids accumulation, excretion and maternal transfer to eggs in chickens around a fluorochemical manufactory in China. Sci. Total Environ.865, 161125 (2023).
Liu, J. et al. Isomer-speciﬁc biotransformation of perfluoroalkyl sulfonamide compounds in aerobic soil. Sci. Total Environ.651, 766-774 (2019).
Liu, S. et al. Transport and transformation of perfluoroalkyl acids, isomer profiles, novel alternatives and unknown precursors from factories to dinner plates in China: new insights into crop bioaccumulation prediction and risk assessment. Environ. Int.172, 107795 (2023).
Ye, W. et al. Associations between serum isomers of perfluoroalkyl acids and metabolic syndrome in adults: isomers of C8 Health Project in China. Environ. Res.196, 110430 (2021).
Miller, K. E. & Strynar, M. J. Improved tandem mass spectrometry detection and resolution of low molecular weight perfluoroalkyl ether carboxylic acid isomers. Environ. Sci. Technol. Lett.9, 747-751 (2022).
Shan, G. et al. Identification and quantification of perfluorooctane sulfonamide isomers by liquid chromatography-tandem mass spectrometry. J. Chromatogr. A1594, 65-71 (2019).
Zhu, J. et al. Investigating isomers/enantiomers of perfluorooctanoic acid in river water by gas chromatography-mass spectrometry with chiral derivatization. Chemosphere238, 124617 (2020).
Naile, J. E. et al. Isomers/enantiomers of perfluorocarboxylic acids: Method development and detection in environmental samples. Chemosphere144, 1722-1728 (2016).
Wellington Laboratories. Standards for environmental testing and research (2023). (accessed 14^th May 2024) https://well-labs.com/wellingtoncatalogue2123.html.
Washington, J. W. et al. Nontargeted mass-spectral detection of chloroperfluoropolyether carboxylates in New Jersey soils. Science368, 1103-1107 (2020).
Yu, N. et al. Non-target and suspect screening of per- and polyfluoroalkyl substances in airborne particulate matter in China. Environ. Sci. Technol.52, 8205-8214 (2018).
Mu, H. et al. Suspect and nontarget screening of per- and polyfluoroalkyl substances based on ion mobility mass spectrometry and machine learning techniques. J. Hazard. Mater.461, 132669 (2024).
Kirkwood-Donelson, K. I. et al. Uncovering per- and polyfluoroalkyl substances (PFAS) with nontargeted ion mobility spectrometry-mass spectrometry analyses. Sci. Adv.9, eadj7048 (2023).
Charbonnet, J. A. et al. Communicating confidence of per- and polyfluoroalkyl substance identification via high-resolution mass spectrometry. Environ. Sci. Technol. Lett.9, 473-481 (2022).
Yao, X. et al. Influence of electrolyte concentration on single-molecule sensing of perfluorocarboxylic acids. Front. Chem.9, 732378 (2021).
Ouldali, H. et al. Electrical recognition of the twenty proteinogenic amino acids using an aerolysin nanopore. Nat. Biotechnol.38, 176-181 (2020).
Ghimire, M. L. et al. Selective detection and characterization of small cysteine-containing peptides with cluster-modified nanopore sensing. ACS Nano16, 17229-17241 (2022).
Lucas, F. L. R. et al. Protein identification by nanopore peptide profiling. Nat. Commun.12, 5795 (2021).
Versloot, R. C. A. et al. β‑barrel nanopores with an acidic-aromatic sensing region identify proteinogenic peptides at low pH. ACS Nano16, 7258-7268 (2022).
Zernia, S. et al. Current blockades of proteins inside nanopores for real-time metabolome analysis. ACS Nano14, 2296-2307 (2020).
Straathof, S. et al. Protein sizing with 15 nm conical biological nanopore YaxAB. ACS Nano17, 13685-13699 (2023).
Dong, B. et al. Trace analysis method based on UPLC-MS/MS for the determination of (C2-C18) per-and polyfluoroalkyl substances and its application to tap water and bottled water. Anal. Chem.95, 695-702 (2023).
Cahill. T. M. Assessment of potential accumulation of trifluoroacetate in terminal lakes. Environ. Sci. Technol.58, 2966-2972 (2024).
Liu, J., & Aksimentiev, A. Molecular determinants of current blockade produced by peptide transport through a nanopore. ACS Nanosci. Au4, 21-29 (2023).
Zuo, J. et al. Single-molecule sensors based on protein nanopores. J. Electrochem. Soc.168, 126502 (2021).
Restrepo-Peŕez, L. et al. Resolving chemical modiﬁcations to a single amino acid within a peptide using a biological nanopore. ACS Nano13, 13668-13676 (2019).
Yao, J. et al. Nontargeted identiﬁcation and temporal trends of per- and polyﬂuoroalkyl substances in a fluorochemical industrial zone and adjacent Taihu lake. Environ. Sci. Technol.56, 7986-7996 (2022).
Joerss, H. et al. Beyond the tip of the iceberg: suspect screening reveals point source-speciﬁc patterns of emerging and novel per- and polyﬂuoroalkyl substances in German and Chinese rivers. Environ. Sci. Technol.56, 5456-5465 (2022).
Wang, Y. et al. Suspect and nontarget screening of per- and polyﬂuoroalkyl substances in wastewater from a fluorochemical manufacturing park. Environ. Sci. Technol.52, 11007-11016 (2018).
Qiu, K. et al. Pore-forming conﬁned space for the innovative electrochemical methods. Curr. Opin. Electrochem.10, 46-53 (2018).
Qiu, K. et al. Toward precision measurement and manipulation of single-molecule reactions by a confined space. Small15, 1805426 (2019).
Zhang, S. et al. A nanopore-based saccharide sensor. Angew. Chem. Int. Ed.61, e202203769 (2022).
Jia, W. et al. Nanopore signatures of nucleoside drugs. Nano Lett.23, 9437-9444 (2023).
Liu, Y. et al. Nanopore identification of alditol epimers and their application in rapid analysis of alditol-containing drinks and healthcare products. J. Am. Chem. Soc.144, 13717-13728 (2022).
Hu, C. et al. Single-molecule sensing of acidic catecholamine metabolites using a programmable nanopore. Chem. Eur. J.28, e202201033 (2022).
Scheringer, M., Cousins, I. T., & Goldenman, G. Is a seismic shift in the landscape of PFAS uses occurring? Environ. Sci. Technol.58, 6843-6845 (2024).
Austin, C. et al. Hydrothermal destruction and defluorination of trifluoroacetic acid (TFA). Environ. Sci. Technol.58, 8076-8085 (2024).
Zheng, G., Eick, S. M., & Salamova, A. Elevated levels of ultrashort-and short-chain perfluoroalkyl acids in US homes and people. Environ. Sci. Technol.57, 15782-15793 (2023).
Sadia, M. et al. Occurrence, fate, and related health risks of PFAS in raw and produced drinking water. Environ. Sci. Technol.57, 3062-3074 (2023).
Zhao, M. et al. Nontarget identification of novel per-and polyfluoroalkyl substances (PFAS) in soils from an oil refinery in southwestern China: a combined approach with TOP assay. Environ. Sci. Technol.57, 20194-20205 (2023).
Cahill. T. M. Increases in triﬂuoroacetate concentrations in surface waters over two decades. Environ. Sci. Technol.56, 9428-9434 (2022).

There is NO Competing Interest.

MLreportingsummaryLHS.pdf
Machine Learning Reporting Summary
nrsoftwarepolicyLHS.pdf
Software Policy Checklist
SupplementaryInformation2024.06.28.pdf

Download PDF

Version 1

posted

You are reading this latest preprint version

Machine learning assisted single-molecule sensing of per- and polyfluoroalkyl carboxylic acids: Quantification without standards

Status:

Version 1

Abstract

Figures

Introduction

Discussion

Methods

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1