LA-UR-25-28183 Approved for public release; distribution is unlimited. Title: Using Machine Learning to Improve Thermostability of MHETase Author(s): Mao, Steven Granja Travez, Rommel Santiago Jha, Ramesh Kumar Intended for: Report Issued: 2025-08-08 Los Alamos National Laboratory, an affirmative action/equal opportunity employer, is operated by Triad National Security, LLC for the National Nuclear Security Administration of U.S. Department of Energy under contract 89233218CNA000001. By approving this article, the publisher recognizes that the U.S. Government retains nonexclusive, royalty-free license to publish or reproduce the published form of this contribution, or to allow others to do so, for U.S. Government purposes. Los Alamos National Laboratory requests that the publisher identify this article as work performed under the auspices of the U.S. Department of Energy. Los Alamos National Laboratory strongly supports academic freedom and a researcher's right to publish; as an institution, however, the Laboratory does not endorse the viewpoint of a publication or guarantee its technical correctness. Using Machine Learning to Improve Thermostability of MHETase Steven Mao, Rommel Granja Travez, Ramesh Jha Abstract: Protein engineering is a field which utilizes proteins as tools, which has many useful applications in medicine, industry, biofuels and more.1 One such protein is MHETase, which is a protein that plays an important function in the degradation of polyethylene terephthalate (PET) plastics, which are commonly used in water and soda bottles.2 However, these proteins are adapted to work in specific conditions, and may not satisfy the desired properties that a new application would desire, or could be improved. For instance, a more thermostable MHETase would be more effective in the plastic degradation conditions.3 To make these desired changes, the primary structure of the protein is mutated, but there are many possible mutations and positions to mutate to make with the 20 canonical amino acids. Therefore, to narrow down the possibilities and to make the process of finding a thermostable MHETase variant, we used sequence design tools that are grounded in machine learning to find mutations that would improve thermostability of MHETase.4 In particular, we used the tools Protein MPNN and FireProt to design a more thermostable MHETase enzyme. We then compiled these mutations into a library and grew these proteins using bacteria colonies, and measured their effectiveness using a fluorescent protein marker. Thermostable proteins and their marker would fold correctly and fluorescence would be seen, but if neither folded correctly then there would be no marker detected. We grew these proteins in bacteria and then intend to use these methods to evaluate their thermostability. Introduction: 1 Aside from improving the thermostability of MHETase, which is the main focus of this internship, another focus was to improve the thermostability of another protein family of endoglucanases, namely Cel9G and Cel9R which can break down cellulose. The optimization of these proteins through protein engineering could provide a method to improve the breakdown and decomposition of biomass. Again, protein engineering alters the primary structure, or the string of amino acids that make up a protein. Once this string is made, it will fold into a specific shape of the protein which matches its corresponding function. By altering this string, we also alter how it works and how strong the bonds between different parts of the protein are, potentially improving the protein function or thermostability. Therefore, this internship will focus on improving the thermostability of both these two endoglucanases and MHETase. In particular, I used computational methods to find stronger mutations to improve the thermostability of the two endoglucanases, while most of the wet work in the internship was spent on the experiments needed to verify and test the mutations that the machine learning libraries suggested for MHETase. To try and enhance protein activity and stability, it is also important to understand what the fitness landscape means. Consider that the current effectiveness of a protein is its “height,” and all of its neighbors are a distance away, and that distance is measured in mutations between the two variants. By using machine learning, we can find some neighbors of the protein that are likely to have increased thermostability based on the training data, and then we can test those neighbors. Then, we take the most effective of those neighbors, find their neighbors, and begin a greedy uphill walk in order to maximize the efficiency and thermostability of the protein. Additionally, we can hopefully train an algorithm to identify the commonalities between the thermostable and extremophile protein, so that way we have a solid understanding of the traits of 2 the extremely stable proteins, and can use this information to design new proteins or alter existing ones to have the same levels of thermostability. Computational Work: Most computational work was about finding the most favorable mutations to enhance the thermostability in the endoglucanases. To first identify mutations, we used the machine learning tools which are ProteinMPNN and FireProt to generate a library of favorable mutations. Then, my mentor asked me to develop a ranking of which mutations should be prioritized to test first, and compare it with his ranking, and the commonly prioritized mutations would be tested later. To develop this ranking, I first looked at which mutations were most commonly suggested by the algorithms, the Shannon entropy of a given position, the variation of that position, the number of flags that the FireProt algorithm gave it, and more. Combining all of these factors, a mutation that is commonly seen throughout the results that the algorithm returned, and promises to have a high thermostability based on the computer evaluation, will have a higher score. By isolating the top 20-30 or so mutations in this way and recording the position and mutation residue, we can get a picture of what mutations are likely to improve the thermostability of an enzyme and then test it. This computational work took place most of June while I was completing training courses required by the lab, in order to begin wet work. As said before, ranking the mutations suggested by FireProt required the development of a score system, and the scores of the mutations would then be compared to determine the best possible mutations to attempt to implement. This score is the sum of mutability, variance, and the number of flags that FireProt assigned to the mutation, where mutability ranks how much/often the wild type residue was changed, variance is self-explanatory, and the number of flags is a positive indication that the 3 algorithm assigned to the mutation. Over the following weeks, I developed Python programs which used the Pandas and NumPy libraries in order to create excel sheets to visualize the ranking of the proteins, and shared these sheets with my mentor so that he could compare it with his own rankings. This way, in the future he could determine which ones are most favorable to test. One crucial difference between the way that my mentor and I computed our scores is with a measure called the FoldX norm which is a value that FireProt gives to a mutation, where a higher score is less favorable (meaning more unstable) and a lower score is more favorable, meaning that it is more stable. However, the FoldX score is very variable, as the lowest score usually goes to around -3 while the highest score can be in the high 30s or 40s. In order to compensate for these outliers, I used a logarithm formula on the positive scores, in a sense “compressing” the range of the positive scores, so that the score would become less dependent on outliers. I believe that it is this difference that made in general my scores to be higher than my mentor’s. I concluded by writing a program that compared our scores together, as well as which parts of the scores (mutability, variability) that made our scores different. After troubleshooting and debugging many of my programs we have a much higher rate of matching mutations at 70% that are worthy of testing, up from 40% which was the percentage of hits at the start of these programs. Writing this program returned another set of mutations with different scores, and now the common mutations that have high scores on both sides can be tested to see their effectiveness. Another focus of these programs is trying to identify the evolutionary pathways that make an enzyme more thermostable. When making a table with the mutations with the highest scores, we look at the residue present in Cel9R, which is a thermostable variant of the protein. 4 Therefore, when evaluating the mutations present in other endoglucanases, we look at how those mutations compare to the current residue in Cel9R. This way, we can see how the mutations suggested by FireProt line up with the mutations that were made through evolutionary pressure to make the proteins in this family more thermostable. Wild Position Type Mutation R. Score S. Score 276 N G 1.382403 1.812 100 K R 0.306168 1.732 443 F Y 1.23164 1.699 483 R K 0.295272 1.652 148 G S 0.167718 1.638 495 K R 1.239464 1.536 232 E D 0.373497 1.429 566 K R -0.25588 1.408 115 E V 1.075649 1.405 607 D L 0.116098 1.405 115 E V 1.075649 1.405 607 D L 0.116098 1.405 354 E N 0.158413 1.388 94 A Y 1.182009 1.376 446 F Y -0.00047 1.352 341 K R 1.374484 1.351 484 A V 0.798012 1.346 5 459 K P 1.346289 1.345 527 Y T 1.38557 1.34 564 A T 0.346709 1.321 562 Q E 1.380337 1.32 446 F E 0.347427 1.316 237 D E 0.245405 1.309 148 G N 0.337163 1.301 453 L T 1.259079 1.3 522 K T 1.139937 1.28 244 S K 1.358832 1.275 115 E K 1.346719 1.274 601 R D 1.354787 1.27 499 Q K 0.301356 1.269 86 Y F 0.329952 1.266 464 D T 0.293514 1.262 Figure 1. This is a chart made from a program that I wrote to compare the scores found by my mentor (R. Score) and my scores (S. Score). Methods: To measure the thermostability of MHETase we use the following process. We first grow a plasmid containing the MHETase wild type in NEB E. Coli, and then once we have enough plasmid, we then extract the DNA from the plasmid. We then send this to be sequenced in order to verify that the DNA we extracted is the correct plasmid backbone. Afterwards, we cut out the wild MHETase part of the plasmid and replace it with the library which we ordered from TWIST 6 bioscience. When we successfully transform this into another set of bacteria and grow it, we will then scrape the colonies and then grow them on a rich medium at a high temperature to isolate for the selection of thermostable bacteria, grow those, and then evaluate thermostability using the fluorescent reporter. Again, the glowing reporter will only glow when the MHETase enzyme itself folds correctly, and this is because proteins in general have their function dependent on their shape, and even one small misfolding or error in making the amino acids in a protein can completely alter the shape, thus changing the function and in our case removing the fluorescence from the fluorescent reporter. After this process, we will then purify all our product and then put it through a machine called a flow cytometer. This is a machine that has a tube that is very slender, so it only allows one cell, protein, and more to pass through at a time. By shining a light through the current cell passing through, we can effectively “count” how many times the fluorescent protein successfully expressed, and therefore how thermostable this variant is. So, using this method, we have a way to measure the thermostability of the protein. 7 Figure 2. This figure from Shin. Et al. demonstrates the experimental process in a clear flowchart. Wet Lab Work: Most of the wet lab work goes through this process, except for giving my mentor a hand when needed. At the beginning of the wet lab work, I aliquoted the antibiotics that my mentor and some other colleagues were using, as we were running low, and I made the antibiotic agarose LB media plates that we would use later too. The reasoning behind these antibacterial media plates is that only the bacteria that takes in our desired plasmid with the resistance gene will be able to grow. In this way, we only grow the bacteria that we want to grow, because all others are killed off by the antibiotics. 8 We use this method to grow a plasmid with the wild type MHETase, which we will later cut out and insert in the library version of MHETase. This library version of MHETase was generated by ProteinMPNN and bought from TWIST bioscience. We successfully grew the bacteria that had this plasmid, with a blank control that had no growth (which is a good indication), and had it sequenced to verify that the plasmid was correct, and the plasmid was the one that we intended to insert. We grew the bacteria (NEB Turbo) on a LB agar plate with carbenicillin antibiotic so only our desired plasmid would be there. After this, we used a restriction digest to successfully cut out the wild type MHETase that was in the plasmid using the Nhe1 and BstB1 enzymes and added Antarctic Phosphatase to stop the plasmid fragments from reattaching to each other. This was important as we added in the library through ligation with T4 DNA ligase, as if the plasmid reattached to become a circle again it would be impossible to insert the library fragment. 9 Figure 3: This is a picture of a gel that I ran after the restriction digest to make sure it worked. The larger band that is highlighted (larger because it is higher up on the ladder) is the backbone we want, while the smaller band is the wild type MHETase fragment that we cut out. However, the attempts so far have been mixed, as although there is growth in the antibiotic plate that is not the control, indicating that likely some library has ligated into the backbone, there is also growth on the control, which means that the ligation process likely went wrong somewhere, or the plasmid reattached to itself. Although there is decent growth on the library plate, combined with the fact that not all of those colonies have the library inserted, as seen with the control, this means that there are not enough colonies for the next part. Currently, we are at an impasse because we have run out of digested library, which is important because the 10 library insert that has been digested with the same restriction enzymes will have a much easier time binding to the backbone. The plan now is to make new digested library, which would be easier said than done because we have no more of one restriction enzyme, only an older version which is buffer-incompatible. However, another mentor has found a workaround to this, where we will use one restriction enzyme to digest, then clean it up using a kit, and then digest with another enzyme, bypassing the incompatible buffer problem. This workaround does have a big risk of loss, so I chose an alternative, to borrow Nhe1 enzyme from another lab to use in the time before our new order of Nhe1 arrives. Discussion: The primary problem that we are facing at this moment is that there is growth on the control, and that our ligation reaction may not have been done properly. The first time we attempted to grow the plasmid with the library to test the results, there was very little growth in the library plates, and we think this could be because during the restriction digest process we used incompatible buffers, although according to the gel electrophoresis it turned out fine. This poses a serious question of what went wrong, because then it is very difficult to fix this error in future experiments. As for this run, because there were so many colonies on the control plate where there were not supposed to be any colonies at all, we suspect that there may also have been a problem with the restriction digest that allowed the plasmid to close in on itself. In order to resolve this issue, we changed the enzymes to the more standard ones that the lab uses. Additionally, after new library digested insert is made we will start the process again. 11