Mega-scale experimental analysis of protein folding stability in biology and protein design

Abstract

Advances in DNA sequencing and machine learning are illuminating protein sequences and structures on an enormous scale. However, the energetics driving folding are invisible in these structures and remain largely unknown. The hidden thermodynamics of folding can drive disease, shape protein evolution, and guide protein engineering, and new approaches are needed to reveal these thermodynamics for every sequence and structure. We present cDNA display proteolysis, a new method for measuring thermodynamic folding stability for up to 900,000 protein domains in a one-week experiment. From 1.8 million measurements in total, we curated a set of ~850,000 high-quality folding stabilities covering all single amino acid variants and selected double mutants of 354 natural and 188 de novo designed protein domains 40-72 amino acids in length. Using this immense dataset, we quantified (1) environmental factors influencing amino acid fitness, (2) thermodynamic couplings (including unexpected interactions) between protein sites, and (3) the global divergence between evolutionary amino acid usage and protein folding stability. We also examined how our approach could identify stability determinants in designed proteins and evaluate design methods. The cDNA display proteolysis method is fast, accurate, and uniquely scalable, and promises to reveal the quantitative rules for how amino acid sequences encode folding stability.

New mega dataset combined with deep neural network makes a progress in predicting impact of mutation on protein stability

“…The performance of the model on MegaTest set is shown in Table 1. The PCC was 0.85 which approaches the agreement to PCC = 0.91 between trypsin-and chymotrypsin-based data from the Mega dataset (Tsuboyama et al, 2022). For the MegaValidation dataset ABYSSAL showed the same PCC = 0.85, see upper panel of Fig.…”

“…We took the experimental data on protein stability changes (∆∆G) upon mutations from (Tsuboyama et al, 2022) where ∆∆G values were estimated from cleavages by proteases. We downloaded the file "K50_dG_Dataset1_Dataset2.csv" from the Zenodo repository https://zenodo.org/record/7401275#.Y6st59JBxD_ associated with the paper (Tsuboyama et al, 2022). Out of 851,552 stability change data for 542 reference proteins, we removed records (i) having tag 'unreliable', (ii) when no mutation was introduced, (iii) associated with insertions and/or deletions, (iv) associated with multiple mutations.…”

“…One of the known limitations of neural networks is that they require a lot of data for their training. Thank to the Mega dataset containing 376,918 mutations we could train the ABYSSAL on 367,858 single mutations to near-to-experimental accuracy of the ∆∆G-estimates-producing method (Tsuboyama et al, 2022). However, we wonder about the minimal dataset size for training the deep learning model we chose here.…”

“…Recently, Tsuboyama et al published the experimentally measured ΔΔG values for 851,552 mutations, with 376,918 of them being high-quality single mutations (Tsuboyama et al, 2022). The dataset is much larger than any dataset used before and has no bias towards ‘truncating’ mutations to smaller amino acids, especially to alanine.…”

“…Despite the recent tremendous success in 3D protein structure prediction, the apparently simpler problem of predicting the effect of mutations on protein stability has been hampered by the low amount of experimental data. With the recent high-throughput measurements of mutational effects in ‘mega’ experiment for ~850,000 mutations [Tsuboyama et al, bioRxiv, 2022] it becomes possible to apply the state-of-the-art deep learning methods. Here we explore the ability of ESM2 deep neural network architecture with added Light Attention mechanism to predict the change of protein stability due to single mutations.…”