Model

Lu et al. (2022) developed a metal ion-binding site prediction and modeling server called MIB2 [1]. This server allowed us to input amino acid sequences and identify potential residues involved in metal ion binding as well as simulate docking of the metal ion within the protein's three-dimensional structure.

Based on the results of MIB2, we then went ahead and utilized another software tool ProteinMPNN from Dauparas et. Al (2023) to design new protein sequences through deep learning [2]. Based on the residues identified in the previous step, we instructed ProteinMPNN to keep the binding residues unchanged but allow for mutation in any other parts of the sequence. Ideally, one of these mutations would result in a gain-of-function mutation that would improve metal binding through structural stability and provide increased molecular contacts with the ligand. Through ProteinMPNN, 10 mutated sequences were generated for our target proteins as well as their predicted structures through AlphaFold.

The plan was then to compare our quantitative results from MIB2 on the wild-type proteins to the newly generated mutated constructs. At this point, the MIB2 server had crashed and did not allow us to verify our results. In order to continue with the experimental workflow while MIB2 was not available, other software was utilized like mebipred from Aptekmann et. al (2022) and LMetalSite from Yuan et. al (2022) to establish any improvement from wild type [3,4]. Based on these results, we sent specific mutated sequences to wet lab to experimentally determine whether the predicted binding capacity can be confirmed. We were able to find a more quantitative software from Zheng et. al (2017) called CheckMyMetal as a replacement for MIB2 [5]. After comparing the wild-type and the specific mutated sequences, we were able to show that there were small improvements in the predictions to binding their respective metals which helped inform the wet lab’s experimental workflow.

In our project, we aimed to develop a machine-learning model that could identify the key features of metal-binding proteins by learning from a library of mutated variants. Our goal was to construct a neural network that would analyze this library, recognize important mutations, and predict new, potentially more effective sequences for metal-binding proteins. To build this library, we based our approach on the hypothesis that mutations improving the metal-binding capabilities of proteins would enhance the survival rate of E. coli in high metal concentrations. Thus, our plan involved error-prone PCR to generate a diverse set of mutations in the target metal-binding protein. We then planned to transform these mutated genes into E. coli and screen the resulting library by exposing it to a slightly higher metal concentration than the original E. coli strain with the unmutated protein could survive. By doing this, we expected that E. coli strains which survived the increased metal concentration would possess more effective metal-binding proteins. This survival would imply that the metal-binding protein mutations in those strains were beneficial, allowing us to select these variants for inclusion in our library. Our neural network model would then learn from this library to identify which specific mutations contribute to improved metal-binding efficiency. However, the results from the toxicity assays were not as conclusive as we had hoped. A clear, consistent trend in protein effectiveness was crucial for building the library, as it would form the training data for our neural network.