PREreview of "Large-scale exploration of protein space by automated NMR"

This Zenodo record is a permanently preserved version of a PREreview. You can view the complete PREreview at https://prereview.org/reviews/18821335. Overview The traditional paradigm of "one structure, one function" is being replaced by the more accurate view that proteins exist as an ensemble of folded structures. While many proteins are presumed to adopt a single, predominant structure, thereby minimizing free energy landscape exploration upon folding, the prevalence of this phenomenon remains underexplored due to technical challenges in capturing these ensembles. However, many recent works demonstrate how alternate conformations within the ensemble exhibit distinct functions (dubbed "functional sub-states"), which can be selected for during evolution or rationally optimized by protein engineers. To date, advanced in ensemble exploration have, thus far, primarily been constrained by limitations in scale of "ensemble detecting" techniques. Here, the authors present NMR-Automated Protein Production (NMR-APP), which aims to tackle this issue. This is a highly integrated, robot-assisted pipeline that automates steps from construct assembly and protein to purification, all the way to the collection of NMR spectra. The authors tested 384 protein designs in their pipeline, successfully purified 98.7% of the constructs, and obtained NMR spectra for 62% of them; all at a low cost of $25 per sample. They computed the spectral peak intensity coefficient of variation (CV) as a proxy for dynamics and demonstrated that these "statically" designed protein structures display pervasive dynamics. Further analysis of nine constructs to assess local conformational flexibility in backbones indicated the presence of heterogeneous and computationally unpredictable dynamical features. This work represents a significant advancement toward understanding sub-state sampling within ensembles, suggesting that such sampling may be an inherent characteristic of designed proteins with a single intended structural state. Areas for improvement / questions for the authors 1. A commonly cited limitation of NMR is protein size. That said, there are no metrics presented in this work that comment on the length of designs. What is the mean and median length? What is the distribution? Is length correlated to any other metrics that may explain success rate in the NMR-APP pipeline, like the concentration of purified protein, or the expected extent of dynamical regions? 2. The authors mention the ability for Proteina to generate less alpha helix biased designs, a great inclusion (along with the use of FoldSeek), to enable broad sampling of structural features. This also seemed to improve the experimental success rate. However, in Figure S4, mean CV versus model used was not included in the cross correlations; a potentially interesting relationship. In other words, does one model have a greater propensity to produce more dynamical structures than the other? Additionally, based on the findings, could the authors speculate on what parameters within, or between, the models could be tweaked to increase the probability of encountering dynamical features in the designs? 3. In lines 245-247, the authors mention that these NMR datasets may help train models to predict dynamics from first principles. Although the authors do not speculate on a putative timeline for such advances, it seems like an ambitious claim. To my knowledge, and as the authors mention, most structural prediction models rely heavily on sequence data. Even with the increasing availability of structural data from pipelines like NMR-APP, is this a sufficient amount of data for training of AI/ML models? 4. The observation that p2 A4 the proportion of minor states was sampled