Classifying disease-associated variants using measures of protein activity and stability

Abstract

Decreased cost of human exome and genome sequencing provides new opportunities for diagnosing genetic disorders, but we need better and more robust methods for interpreting sequencing results including determining whether and by which mechanism a specific missense variants may be pathogenic. Using the protein PTEN (phosphatase and tensin homolog) as an example, we show how recent developments in both experiments and computational modelling can be used to determine whether a missense variant is likely to be pathogenic. One approach relies on multiplexed experiments that enable determination of the effect of all possible individual missense variants in a cellular assay. Another approach is to use computational methods to predict variant effects. We compare two different multiplexed experiments and two computational methods to classify variant effects in PTEN. We distinguish between methods that focus on effects on protein stability and proteinspecific methods that are more directly related to enzyme activity. Our results on PTEN suggest that ~60% of pathogenic variants cause loss of function because they destabilise the folded protein which is subsequently degraded. Methods that quantify a broader range of effects on PTEN activity perform better at predicting variant effects. Either experimental method performs better than the corresponding computational predictions.

Computational and cellular studies reveal structural destabilization and degradation of MLH1 variants in Lynch syndrome

“…Our observation that single amino acid changes in MLH1 are sufficient to cause degradation is similar to results from multiple other proteins including recent deep mutational scans on PTEN and TPMT (Matreyek et al, 2018), our previous results on MSH2 in human cells (Nielsen et al, 2017), and earlier observations on Lynch syndrome MSH2 variants in yeast (Gammie et al, 2007; Arlow et al, 2013). Our structural stability calculations predict that a relatively mild destabilization of just a few (~3) kcal/mol is sufficient to trigger MLH1 degradation, an observation in line with previous studies on other proteins (Bullock et al, 2000; Nielsen et al, 2017; DDD Study et al, 2017; Scheller et al, 2019; Jepsen et al, 2019; Caswell et al, 2019). Although the absolute thermodynamic stability of MLH1 is unknown, both in vitro and in a cellular context, it is possible that the 3 kcal/mol destabilization necessary to trigger degradation is lower than that required to reach the fully unfolded state.…”

“…While those have slightly higher overall accuracy, they do not directly indicate the underlying molecular reason for pathogenicity. Incidentally, we note that by analyzing both calculations and multiplexed assays of variant effects, we recently found that ~60% of disease-causing variants in the protein PTEN were caused by destabilization and a resulting drop in cellular abundance (Jepsen et al, 2019). On the other hand, while stability prediction is very useful for accurate identification of many pathogenic variants, it may have lower overall sensitivity.…”

Predicting and interpreting large scale mutagenesis data using analyses of protein stability and conservation

“…Pinpointing which variants are in the detrimental group, and the biochemical and biophysical mechanisms underlying loss of fitness, is important for example for assessing pathogenicity of so-called variants of uncertain significance (Richards et al, 2015) and understanding the mechanistic origins of disease. methods are important to predict and understand variant effects, and in some cases they may be even be more accurate than MAVEs for this purpose (Jepsen et al, 2020;Frazer et al, 2020).…”

“…In such tests, it has been shown that various sequence-based approaches, including deep-learning methods, can achieve very high accuracy (Riesselman et al, 2018;Livesey and Marsh, 2020). Indeed, we and others have successfully applied sequence analysis and biophysical stability calculations for identification and analysis of pathogenic variants, although these methods were not trained on clinical variants (Pey et al, 2007;Yin et al, 2017;Nielsen et al, 2017;Gray et al, 2018;Scheller et al, 2019;Cline et al, 2019;Abildgaard et al, 2019;Jepsen et al, 2020;Frazer et al, 2020).…”

“…As evolution disfavours unstable proteins (Bloom et al, 2006;Echave and Wilke, 2017), there is some correlation between the two approaches, however, there are also differences, e.g. for active sites, where variants are typically only identified as detrimental by evolutionary sequence analysis (Cheng et al, 2005;Echave, 2019;Jepsen et al, 2020;Cagiada et al, 2021). We train a machine learning model that uses ΔΔG and ΔΔE as input to predict variant effects as probed by MAVE experiments.…”