To begin generating novel enzymes to degrade per- and polyfluoroalkyl substances (PFAS), it was necessary to identify candidate enzymes to build effective training data to finetune our pLLMs on known PFAS-degrading enzymes or enzymes predicted to degrade PFAS and properly bias the model towards structures capable of PFAS degradation. These candidate enzymes would have to exhibit some level of catalytic activity with PFAS species, particularly the ability to cleave the strong carbon-fluorine bonds that make PFAS so difficult to degrade. Candidate enzymes also needed to be sequentially diverse from each other to ensure that when trained on this data, the pLLM would be capable of generating appropriately novel sequences and would not overfit to a particular structure. To this end, we completed an extensive literature review where we identified two enzyme families that had the potential to meet the requirements for effective training data: Radical-producing enzymes and dehalogenases.

Radical-producing enzymes produce highly reactive radical species that disrupt the structure of PFAS, leading to degradation. Dehalogenases have ligand-specific active sites that can defluorinate certain species of PFAS. The inherently destructive nature of radical-producing enzymes makes them difficult to express in microbes recombinantly. A potential approach to remedy this challenge would be to express them within a cell-free extract. However, because the radical species produced by these enzymes are generally reactive, they not only degrade PFAS, making the reactions they cause less controlled and more unpredictable. This fact made radical-producing enzymes less appealing to the team.

With that, we pivoted our focus to dehalogenases. In our search to find specific dehalogenases proven to degrade PFAS, the team discovered the corrinoid iron-sulfur reductive dehalogenase A6RdhA. A6RdhA was identified within Acidimicrobium sp. Strain A6, a soil-dwelling microbe that, when incubated with a PFOA substrate, could partially defluorinate PFOA's structure . PFOA is a legacy PFAS species that the EPA recently mandated in 2024 to be less than 4 parts per trillion (ppt), the minimal detectable limit, in drinking water.

When the gene coding for A6RdhA was knocked out, Acidimicrobium sp. Strain A6 lost its ability to degrade PFOA. This kind of concrete proof of PFAS degradation caught the team's attention, and we decided to focus our efforts towards reductive dehalogenases. Because A6RdhA has only recently been identified, its properties and complete structure have not been well characterized; in fact, the only known amino acid sequence of A6RdhA is missing a C-terminus of more than 100 amino acids. This incomplete fragment is predicted to be incapable of completing catalysis because the missing C-terminus makes up a large part of its

To find enzymes structurally similar to A6RdhA, we used the program Foldseek, which matches a query protein structure with proteins with similar structures within various protein structures and sequence databases. Through this method and a secondary literature review, the team identified 68 proteins with high structural similarity to A6RdhA (Fig. 2B).

The majority of these enzymes had a relatively low annotation score on Uniprot and were not very well characterized structurally or functionally. However, many were classified as reductive dehalogenases, which aligns with the classification of A6RdhA and fits within the family of enzymes we were researching. In addition to having high structural similarity to a known PFAS degrading enzyme, the 68 were also sequentially diverse from each other.

At this point, the computational team formulated a separate plan that aimed to reconstruct the missing C-Terminus of A6RdhA, allowing us to produce a functional novel enzyme closer to A6RdhA in structure and, therefore, function than any other candidate. To achieve this goal, the team would produce a chimera enzyme made up of the fragment of A6RdhA and one of the structurally similar candidates. The candidate chosen to reconstruct A6RdhA was enzyme T7RdhA, a reductive dehalogenase with the second-highest structural similarity to A6RdhA out of the 68 candidates. Another research group computationally verified T7RdhA to have PFAS degrading abilities similar to A6RdhA.

To construct the chimera, the structures of T7RdhA and the fragment of A6RdhA were aligned, and the point at which the fragment ends was identified. The point at which T7RdhA's sequence continues from the end of the fragment was identified. Every amino acid within T7RdhA's sequence from where the fragment sequence ends and T7RdhA's sequence begins was grafted onto the end of the A6RdhA fragment. This alteration of A6RdhA's structure reconstructed its active site (Fig. 3A).

When expressed, A6RdhA and other reductive dehalogenases form a dimer structure. To test whether the novel chimera dimerized properly, the structural file for the chimera was also input into AlphaFold3, where its dimerized structure was predicted. Its dimerized structure followed the binding formation found in other similarly structured dehalogenases and even had a higher plDDT score than that of the A6RdhA fragment's dimerized structure. The pLDDT score (predicted Local Distance Difference Test) measures the confidence of a predicted protein structure at each residue, with higher scores indicating greater reliability. This novel chimera was dubbed the A6T7 Chimera.

To attempt to characterize the candidate enzymes computationally and validate their activity with PFAS, the computation group completed a myriad of ligand-protein docking simulations with their three-dimensional structure. The results from these tests were inconclusive and seemed to show no clear pattern in binding across various PFAS ligands, specific negative control ligands, and negative control proteins. In a secondary attempt to computationally validate our candidates, we began completing molecular dynamics simulations to gauge affinity by observing whether a ligand docked within the active site of a receptor would start to drift out over time.