Discover the second part of our model, which focus on the CRISPR dCas9 part, on the Model page!

BacPROTACs, The conception of our tool

PROTACs and BacPROTACs

Some of these chemically synthesized molecules used on eukaryotic cells are in phase 3 clinical trials for treating breast and prostate cancer [3]. PROTACs are also being optimized to target other cell-specific E3 ligases, thus extending their sphere of activity.

But what about BacPROTACs?

First introduced as proof of concept in 2021 by Tim Clausen's team [4], these molecules are used in bacteria. Unlike eukaryotic cells, which have a proteasome, prokaryotic cells, like bacteria, don’t have the exact same system. While they do possess a protein complex enabling the degradation of proteins at the end of their life cycle or when they have not been properly formed, this complex differs in its composition. Indeed, bacteria have a lot of diverse systems to eliminate unwanted or malfunctioning proteins or regulate their production, since they provide essential functions. It is estimated 80% of the bacteria proteolysis is done by the Hsp100/Clp (heat shock protein-100/caseinolytic protease) and Lon families. The first one is essential in the folding, assembly, and degradation of proteins during normal growth and, mainly, under stress-inducing conditions. This family functions with different ATPase chaperones that will pair with a peptidase, ClpP (caseinolytic protease proteolytic subunit (P)) to complete the degradation. Together, they will form the chaperone–ClpP complex which is capable of degrading proteins in a specific manner: the chaperones can use ATP to promote protein folding changes and direct protein degradation by ClpP. Indeed, ClpP is an ATP-dependent peptidase that has a serine peptidase activity when linked to a chaperone. ClpC (caseinolytic protease subunit (C)) is one of these ATPase chaperones. Another one is the ClpX (caseinolytic protease subunit (X)), and most of the Gram-negative bacteria contain the ClpXP protease, and has been really well studied in E. Coli.

How did the idea come up?

As we said, our project aims at re-sensitizing E. coli against a certain type of antibiotic from the beta-lactam family: carbapenems. In fact, certain pathogenic strains of E. coli are resistant to these antibiotics, which are nevertheless antibiotics of last resort, that is, only used in hospitals after testing other antibiotics that have proved ineffective in treating a bacterial infection. These resistances in E. coli can be due to the production of carbapenemases, enzymes that destroy carbapenems by cutting them. In order to re-sensitize E. coli, these enzymes must be eliminated, so that carbapenems can once again have a bactericidal effect on E. coli.

Conception

As the bacteria doesn’t have the same degradation system as eukaryotic cells, the first challenge was to find a degradation complex, present in all our bacteria of interest, able to degrade OXA48. In the literature, we found a paper written by Tim Clausen’s team in 2022 [1], where the first BacPROTACs ever created is reported. However, their BacPROTACs were functioning with the ClpCP degradation system. And this system is only found in Gram-Positive and Cyanobacteria.

Normally, in E. coli, ClpXP recognizes specific N- and C-terminal motifs located within its substrates but we know thanks to this paper that it can work if a BacPROTACs bring together ClpX and the protein of interest. As the ClpXP was present in E. Coli and well characterized, we decided to use this one. However, we didn't have any ligand for ClpX, as the ligand Tim Clausen’s team used in his paper was ClpC specific. To overcome this problem, we decided to directly integrate the ClpX protease to our BacPROTAC by designing a linker binding it to the OXA48 ligand. In this way, it recruits nearby ClpP, which has protease properties and degrades the protein.

To link our two proteins, we needed a linker. But, on the contrary of what we thought at the beginning, a linker can’t be something we put without thinking about it further. Indeed, it’s a very important part of the BacPROTACs since it allows the two proteins to stick close enough so the degradation of the protein of interest can be done but to be far apart at the same time so there are no clashes and they don’t alter each other's function and activity.

So the first question that came to our minds was if the ClpX chaperone would still work while bind to another protein. So we looked into the literature to see if a tag with fluorescence GFP wouldn’t have been done: the GFP would have been automatically linked to the ClpX part by a linker. We found a paper by Brandon Williams written in 2014: “ClpXP and ClpAP proteolytic activity on divisome substrates is differentially regulated following the Caulobacter asymmetric cell division”, were a GFP protein had been linked to the ClpX protein on the C-terminal extremity, without perturbing the ClpX function. And since the GFP is much bigger than the nanobody (27 kDa against 10), we believe that it will be the same for our nanobody. It won’t perturbate the activity nor be degraded. After that, we decided that we could use the linker of the complex ClpX - GFP, as it has been proved to work fine.

IMP is an open source software (Integrative Modeling Platform), in particular a C++/Python library that allows us to calculate, in our case, the distribution of the different coordinates that our linker takes if it's connected to ClpX in C-terminal and nanobody. Although this library includes several modules, for our linker we would have had to create a Python script from scratch. The data we could provide as an input were the amino acid sequence structures of the ClpX protein and nanobody. As for the linker, we also had to give it the simplest possible amino acid sequence, such as 3 Glycines. In effect, the script would have simulated all possible amino acid coordinates of the linker in the environment of ClpX and the nanobody to which it is linked. It's clear that, for the script, only the linker had to be flexible, while the nanobody and ClpX had to be rigid. Otherwise, at best, the molecular dynamics system would work, but the resulting data set would take far too long to analyze and would be incoherent; at worst, the molecular dynamics system would have been aborted.

On the other part of the BacPROTAC, a protein that was specific to OXA48 was needed to recruit it. We decided to use a nanobody, since it’s a small molecule (a small molecule is smaller than 100 kDa, ours is about 10 kDa), and that could bind to any part of OXA48 in a specific way. Indeed, in order to prevent the creation of resistance, we wanted a protein that didn’t bind to the active site of the protein because it would entrave it function and so could create a resistance.

To have a starting point, we looked in the literature and found a nanobody that was CMY-2 specific, this protein being a ß-lactamase from the same group as OXA48, so really close to it in term of function, folding and structure. So we decided to try this nanobody with a group of researchers and found out that it had really few interactions with OXA48. Thus, we had two options: try an experimental modification of the nanobody that could be really long but would lead to a result, or modelize a better nanobody by modifying the previous one thanks to fold algorithms. With the help of Riccardo Pelarin, a researcher at the ICPB in Lyon, we decided to try a really recent folding model, RF Diffusion (The step by step modification method is resumed in the RF Diffusion manual, which can be found bellow). Thanks to that tool, we were able to simulate the interaction between OXA48 and the nanobody we found in the literature, and then modify it. Indeed, a nanobody possesses three hypervariable parts that will give its specificity to the nanobody. And since we already had this nanobody that was specific to CMY-2, few changes were needed to make a new nanobody from it that would be oxa48 specific. The algorithm first modified the backbone of the nanobody in order to make it interact better with OXA48. It then searched for a sequence that could fold into that backbone and verified it. We run that algorithm multiple times to have several different propositions at the end and so more chances of interaction. We ended up with around 800 results and selected 162 following criteria of pLDDT, which marks the per residue confidence score (it had to be >0.89) and with the lowest RMSD* (which rates whether the alignment between the predicted backbone and the final structure is good or not). At the end we selected the three most different ones in the first ranked nanobody (to have time to test them), to increase the chances to have an interaction.

RF diffusion manual

References