(191b) Active Learning of Optimal Linear Molecular Probes to Bind with per- and Polyfluoroalkyl Substances in Water | AIChE

(191b) Active Learning of Optimal Linear Molecular Probes to Bind with per- and Polyfluoroalkyl Substances in Water


Dasetty, S. - Presenter, The University of Chicago
Wang, Y., University of Chicago
Rowan, S. J., University of Chicago
Lee, S. S., Argonne National Laboratory
Darling, S. B., Argonne National Laboratory
Benmore, C. J., Argonne National Laboratory
Willet, R., University of Chicago
Jonas, E., University of Chicago
Chen, J., University of Chicago
Ferguson, A., University of Chicago
Per- and polyfluoroalkyl substances (PFAS) are highly persistent synthetic chemicals that contain at least one fully fluorinated methyl or methylene carbon group.[1,2] Potential detrimental effects caused by PFAS in humans include higher cholesterol levels, birth defects, and cancer.[1-4] To avoid the exposure of PFAS in living organisms, treatment systems capable of their selective removal from water are highly desirable.[5,6] In this work, we discovered linear molecular probes for building such water treatment systems by using an integrated approach combining molecular dynamics (MD) simulations, enhanced sampling methods, machine learning (ML), and wet-lab experiments.

We started our search by testing the effectiveness of linear probes that can bind via fluorophilic and electrostatic interactions with perfluorooctanesulfonic acid (PFOS), our initial target PFAS. We observed sensitivity of the probes moderately increased with number of fluorinated carbons but their selectivity to PFOS remained low (<-0.25 kBT) relative to sodium dodecyl sulfate (SDS), a template interferent. A similar trend was observed with increase in number of hydrogenated carbons in the probe but with slightly lower sensitivity than fluorinated molecules for certain carbon lengths. Our results show a moderate increase (~0.75 kBT) in selectivity for the shortest studied probe when one carbon group in the hydrogenated probe is replaced with a primary amine head group. To further optimize the probes, we use the collected initial data to develop a computational active learning approach involving deep representational learning of probes via variational autoencoder and multi-objective Bayesian optimization to efficiently navigate the vast molecular design space and discover linear probes with optimal sensitivity and selectivity to PFOS.

In our presentation, we will discuss our development of the computational active learning framework that includes optimization over the length of all-atom MD simulations and enhanced sampling to minimize the computational cost. We will then present our understanding of the design rules of the discovered optimal probes and their effectiveness observed in wet-lab experiments. Ultimately, the discovered optimal probes will be deployed for efficient and effective detection as well as removal of PFAS in water sources.


[1] Lindstrom AB, Strynar MJ, Libelo EL. Polyfluorinated compounds: past, present, and future. Environmental science & technology. 2011 Oct 1;45(19):7954-61.

[2] Hubal EA. PFAS: Insights from past actions to inform today’s decisions. J Expo Sci Environ Epidemol 29, 129-130 (2019)

[3] Steenland K, Woskie S. Cohort mortality study of workers exposed to perfluorooctanoic acid. American journal of epidemiology. 2012 Nov 15;176(10):909-17.

[4] Ding G, Peijnenburg WJ. Physicochemical properties and aquatic toxicity of poly-and perfluorinated compounds. Critical reviews in environmental science and technology. 2013 Jan 1;43(6):598-678.

[5] Rodriguez KL, Hwang JH, Esfahani AR, Sadmani AH, Lee WH. Recent Developments of PFAS-Detecting Sensors and Future Direction: A Review. Micromachines. 2020 Jul;11(7):667.

[6] Wang Y, Darling SB, Chen J. Selectivity of Per- and Polyfluoroalkyl substance sensors and sorbents in water. ACS Appl. Mater. Interfaces. 2021, 13, 51, 60789-60814.