Estimating conformational heterogeneity of tryptophan synthase with a template‐based Alphafold2 approach

Abstract

The three‐dimensional structure of the enzymes provides very relevant information on the arrangement of the catalytic machinery and structural elements gating the active site pocket. The recent success of the neural network Alphafold2 in predicting the folded structure of proteins from the primary sequence with high levels of accuracy has revolutionized the protein design field. However, the application of Alphafold2 for understanding and engineering function directly from the obtained single static picture is not straightforward. Indeed, understanding enzymatic function requires the exploration of the ensemble of thermally accessible conformations that enzymes adopt in solution. In the present study, we evaluate the potential of Alphafold2 in assessing the effect of the mutations on the conformational landscape of the beta subunit of tryptophan synthase (TrpB). Specifically, we develop a template‐based Alphafold2 approach for estimating the conformational heterogeneity of several TrpB enzymes, which is needed for enhanced stand‐alone activity. Our results show the potential of Alphafold2, especially if combined with molecular dynamics simulations, for elucidating the changes induced by mutation in the conformational landscapes at a rather reduced computational cost, thus revealing its plausible application in computational enzyme design.

Curse and Blessing of Non‐Proteinogenic Parts in Computational Enzyme Engineering

“…For example, the multiple sequence alignment (MSA) generated by AlphaFold can be manipulated to reveal alternative protein conformations. [47][48][49][50] Additional information can also be extracted from the per-residue uncertainty factor called the predicted local distance difference test (pLDDT), which was shown to correlate with protein dynamics. [51][52] While confident structure prediction enabled a plethora of opportunities in protein design, such as the de novo design of backbones from random sequences via Monte Carlo "hallucination" [53] or diffusion models, [54] the reliable prediction of the structure of point mutations, lamentably, is still an unmet need.…”

“…Some of the currently popular AI structure prediction tools create valuable, sequence‐related output. For example, the multiple sequence alignment (MSA) generated by AlphaFold can be manipulated to reveal alternative protein conformations [47–50] . Additional information can also be extracted from the per‐residue uncertainty factor called the predicted local distance difference test (pLDDT), which was shown to correlate with protein dynamics [51–52] .…”

AlphaFold2 and Deep Learning for Elucidating Enzyme Conformational Flexibility and Its Application for Design

“…Inspired by these previous publications showing AF2’s ability to sample additional conformational states, we developed a template-based AF2 approach to assess the conformational heterogeneity and how this is altered by mutations on the β subunit of several tryptophan synthase enzymes (TrpB, see Figure ). − As done in the work of del Alamo et al, we tested the effect of reducing the provided number of sequences in the MSA, but we additionally assessed how AF2 predictions are altered when different templates displaying multiple conformational states are provided . We tested the template-based AF2 pipeline by providing either X-ray based or conformations extracted from MD simulations as templates.…”

“…The new FEL generated from the 10 ns MD simulations starting from the ca. 1000 AF2 outputs at different MSA is shown in a blue to red colormap on top of the computationally reconstructed FEL obtained via well-tempered multiple-walker metadynamics simulations (in gray). , The x and y axis of the reconstructed FELs indicate the Open-to-Closed (O-to-C) transition of the COMM domain of TrpB that covers the active site, and the Mean Square Deviation (MSD) from the path of the generated O-to-C structures, respectively. − The input sequence is the 0B2- pf TrpB variant . Reproduced with permission from ref .…”

“…Overview of strategies developed for predicting alternate states with AlphaFold2 (AF2). As done in del Alamo et al12 the Multiple Sequence Alignment (MSA) depth can be altered, some of the MSA positions can be masked as shown by Stein and McHaourab,13 and the MSA can be clustered as in the Kern and Ovchinnikov preprint paper.37 The provided set of templates can also be changed as done in some cases in del Alamo et al12 and also in our recent publication 34.…”