Selective PFAS Detection with Functionalized Cyclodextrin Probes Designed via Bayesian Optimization

This repository contains the computational assets associated with a study on cyclodextrin-based molecular recognition of PFAS, with an emphasis on improving PFOS selectivity over structurally similar surfactants such as SDS.

Citation

Title: Selective PFAS Detection with Functionalized Cyclodextrin Probes Designed via Bayesian Optimization

DOI: Pending. Replace the badge and this line with the final DOI link after publication, for example:

[![DOI](https://img.shields.io/badge/DOI-10.XXXX%2FXXXXX-blue)](https://doi.org/10.XXXX/XXXXX)

Paper Summary

Per- and polyfluoroalkyl substances (PFAS) are persistent environmental contaminants that demand highly selective molecular recognition strategies for field-deployable detection. beta-Cyclodextrin-based field-effect transistor (FET) sensors demonstrate high sensitivity to perfluorooctanesulfonic acid (PFOS), achieving sub-ppt detection limits, yet exhibit limited selectivity in the presence of structurally similar surfactants such as sodium dodecyl sulfate (SDS). Here, we screen a synthetically accessible library of 1,629 functionalized alpha-, beta-, and gamma-cyclodextrins to quantify competitive binding thermodynamics using docking and all-atom molecular dynamics simulations. We identify host architectures with sub-nanomolar PFOS affinity and high selectivity, and use regression analysis to connect binding behavior to structural and electronic descriptors. Together, these results establish quantitative structure-selectivity relationships for cyclodextrin-based PFOS recognition and provide design principles for next-generation PFAS sensing materials.

Repository Overview

The repository is organized into four main parts:

chem_space_data/

Chemical-space datasets and structure libraries for the screened cyclodextrin hosts.

• chem_space.pkl stores the enumerated library and associated thermodynamic and descriptor fields.
• chem_space_export.csv is a tabular export of selected screened candidates for quick inspection outside Python.
• chem_space_pdb_files/ contains the candidate 3D structures as PDB files.
• prim_cleaved_structs/ contains reference alpha-, beta-, and gamma-cyclodextrin scaffolds.
• analyze_chem_space.py provides a lightweight way to inspect the dataset contents.

This is the best starting point if you want to understand the screened design space or inspect individual candidates.

metadynamics/

Simulation setup files and helper scripts for the all-atom molecular dynamics and metadynamics workflows used to evaluate host-guest binding.

• example system directories such as bcd-pfos/, bcd-sds/, and 00464-pfos/
• shared GROMACS and PLUMED templates in common.files/
• helper scripts for generating PLUMED plane definitions and cyclodextrin backbone restraints
• ff-parameterize/ for guest force-field generation and topology preparation

This is the best starting point if you want to reproduce or inspect the simulation setup workflow.

bayesianoptimization/

Gaussian-process and candidate-ranking utilities used to connect descriptors and MD-derived data to chemical-space prioritization.

• morganKernel.py defines additive GP models
• training.py contains the model-training loop
• get_candidates_delta.ipynb is the interactive analysis notebook
• data/ stores the serialized training, MD, and candidate datasets used by the modeling workflow

This is the best starting point if you want to inspect the surrogate-modeling and candidate-selection components.

lasso/

Descriptor-analysis notebook for sparse, interpretable regression on the screened chemical space.

• lasso.ipynb loads chem_space.pkl
• computes RDKit molecular descriptors and charge-based features for substituent sets
• supports feature selection and regression-style analysis alongside the Bayesian-optimization workflow

This is the best starting point if you want a more interpretable descriptor-based model rather than the Gaussian-process workflow.

How To Read The CSV

The file chem_space_data/chem_space_export.csv is a compact export of selected probe designs. It currently contains 79 probe entries plus a header row.

The columns are:

• probe ID Repository-style identifier for the candidate, matching the numeric naming convention used elsewhere in the chemical-space data.
• CD type Cyclodextrin family for the probe, reported as alpha-CD, beta-CD, or gamma-CD.
• primary A string-encoded list of substituents on the primary face of the cyclodextrin. These entries are written as SMILES-like fragments inside a Python-style list.
• secondary A string-encoded list of substituents on the secondary face of the cyclodextrin, in the same format.
• dG_md MD-derived PFOS binding free energy stored as [mean, uncertainty].
• ddG_md Relative selectivity-style free energy term, also stored as [mean, uncertainty].
• Kd_md Dissociation constant for PFOS from the MD workflow, stored as [mean, uncertainty].
• Kd_SDS/Kd_PFOS Selectivity ratio between SDS and PFOS binding, stored as [mean, uncertainty]. Larger values indicate stronger preference for PFOS over SDS.

Two formatting details are important:

• The primary and secondary columns are not plain text labels. They are serialized lists of substituent strings.
• The thermodynamic columns are not single numbers. Each cell is a two-element array, where the first value is the central estimate and the second value is the uncertainty.

Example interpretation of one row:

• probe ID = 00001
• CD type = gamma-CD This probe is built on a gamma-cyclodextrin scaffold.
• primary = ["[Br]", ...] Every primary-site substituent in that candidate is bromine.
• secondary = ["[OH]", ...] The secondary sites remain hydroxylated.
• dG_md = [-31.52, 0.40] Mean binding free energy of about -31.5 with uncertainty 0.4 in the stored units.
• Kd_SDS/Kd_PFOS = [22.5, 8.8] SDS is predicted to bind more weakly than PFOS by roughly a factor of 22.5, with the listed uncertainty.

Suggested Entry Points

• Start with chem_space_data/README.md to understand the screened library and stored structures.
• Read metadynamics/README.md for the simulation setup and analysis workflow.
• Read bayesianoptimization/README.md for the GP modeling and prioritization workflow.
• Open lasso/lasso.ipynb for descriptor extraction and sparse regression analysis.

Notes

• The repository is focused on computational workflow components and intermediate data products rather than a polished software package.
• Several simulation scripts reflect the original HPC environment and may require path updates before reuse on another system.