Rapid protein stability prediction using deep learning representations

Abstract

Predicting the thermodynamic stability of proteins is a common and widely used step in protein engineering, and when elucidating the molecular mechanisms behind evolution and disease. Here, we present RaSP, a method for making rapid and accurate predictions of changes in protein stability by leveraging deep learning representations. RaSP performs on-par with biophysics-based methods and enables saturation mutagenesis stability predictions in less than a second per residue. We use RaSP to calculate ∼ 300 million stability changes for nearly all single amino acid changes in the human proteome, and examine variants observed in the human population. We find that variants that are common in the population are substantially depleted for severe destabilization, and that there are substantial differences between benign and pathogenic variants, highlighting the role of protein stability in genetic diseases. RaSP is freely available-including via a Web interface-and enables large-scale analyses of stability in experimental and predicted protein structures.

Correspondence between functional scores from deep mutational scans and predicted effects on protein stability

“…AlphaFold2 models were used for proteins or select residues in infrequent cases where they were not covered by experimental structures. It has been recently demonstrated that accurate stability predictions can also be delivered using modeled protein structures, with AlphaFold2 providing suitable inputs for FoldX even for proteins without homologs in the training set (Akdel et al, 2021; Blaabjerg et al, 2023; Pak & Ivankov, 2022).…”

“…We tested nine stability predictors, seven of which we have previously also explored for their ability to distinguish between pathogenic and putatively benign human variants (Gerasimavicius et al, 2020). On top of FoldX, Rosetta, INPS3D, PoPMuSiC, mCSM, ENCoM, and DynaMut2, we have included DDGun3D, an “untrained” stability prediction method, as well as the recently released RaSP which offers rapid evaluation of variants based upon sequence alone through a neural network model (Alford et al, 2017; Blaabjerg et al, 2023; Dehouck et al, 2011; Delgado et al, 2019; Frappier et al, 2015; Montanucci et al, 2022; Pires et al, 2014; Rodrigues et al, 2021; Savojardo et al, 2016). While most methods only offer functionality of evaluating stability perturbing effects of mutations on monomeric structures, FoldX, ENCoM and Rosetta were also evaluated in terms of full protein complex structures, if they were available, as this functionality is easily accessible in these predictors.…”

“…Instead, the high agreement is likely due to an underlying association of loss‐of‐function mechanisms with the protein and phenotype in question. RaSP (Blaabjerg et al, 2023), a deep learning‐based method that simplifies structural protein representations, benefits indirectly through being parametrized against Rosetta predictions, while leveraging considerably improved computational speed. More novel approaches combining machine learning with sequence‐ or structure‐based features, such as mCSM, DDGun3D, PoPMuSiC and INPS3D, do not seem to overall be as effective at ranking functional impacts.…”

“…The highest correlations of 0.76 were observed between DDGun3D and INPS3D, as well as 0.75 between RaSP and Rosetta. Indeed, one would expect RaSP and Rosetta to show a high degree of correlation, as RaSP has been trained not on experimental ΔΔG values, but on Rosetta predictions (Blaabjerg et al, 2023). The high correlation between DDGun3D and INPS3D is likely due to the fact that both methods are essentially extensions of their underlying sequence and alignment-based models, and therefore both directly utilize information about evolutionary sequence conservation, in contrast to all other predictors.…”

“…For complex evaluation, PDB structures 4JZW and 4JZZ were modified to introduce cysteines instead of some nonstandard residues for Rosetta to recognize the disulfide bonds. RaSP (Blaabjerg et al, 2023) was run on a modified Google Colab notebook, based on a mix of code from GitHub commits 26b0b1a and 518624e. INPS3D (Savojardo et al, 2016), mCSM (Pires et al, 2014), DynaMut2 (Rodrigues et al, 2021), and PoPMuSiC (Dehouck et al, 2011) webservers were queried in Python 3.8.8 using RoboBrowser (https://github.com/jmcarp/robobrowser) or Selenium (https://github.com/baijum/selenium-python) packages.…”