MISsense deleTeriousness predICtor


The diffusion of next-generation sequencing technologies has revolutionized research and diagnosis in the field of rare Mendelian disorders, notably via whole-exome sequencing (WES). However, one of the main issues hampering achievement of a diagnosis via WES analyses is the extended list of variants of unknown significance (VUS), mostly composed of missense variants. Hence, improved solutions are needed to address the challenges of identifying potentially deleterious variants and ranking them in a prioritized short list. We present MISTIC (MISsense deleTeriousness predICtor), a new prediction tool based on an original combination of two complementary machine learning algorithms using a soft voting system that integrates 113 missense features, ranging from multi-ethnic minor allele frequencies and evolutionary conservation, to physiochemical and biochemical properties of amino acids. Our approach also uses training sets with a wide spectrum of variant profiles, including both high-confidence positive (deleterious) and negative (benign) variants. Compared to recent state-of-the-art prediction tools in various benchmark tests and independent evaluation scenarios, MISTIC exhibits the best and most consistent performance, notably with the highest AUC value (> 0.95). Importantly, MISTIC maintains its high performance in the specific case of discriminating deleterious variants from benign variants that are rare or population-specific. In a clinical context, MISTIC drastically reduces the list of VUS (<30%) and significantly improves the ranking of “causative” deleterious variants. Pre-computed MISTIC scores for all possible human missense variants are available at http://lbgi.fr/mistic.

Benchmark results

Fig 1. Performance of missense prediction tools on VarTest set.
Figure 1
MISTIC was compared to individual component features (MetaSVM, MetaLR, VEST4, Condel, CADD, PolyPhen2, SIFT) used in its model (in grey) and the best-performing tools recently published (in color). The Area Under the receiver operating characteristics Curve (AUC) is shown in brackets.

Fig 2. Evaluation of prediction tools on different variant analysis scenarios.
Figure 2
The performance of MISTIC was compared to other missense prediction tools for the discrimination of deleterious variants from rare benign variants and population-specific missense variants. All prediction tools were evaluated using novel deleterious variants (Fig 2A - ClinVarNew and Benign_EvalSet set), known deleterious variants from diverse sources (Fig 2B - DoCM and Benign_EvalSet set), rare benign variants with MAF data (<0.01, <0.005, <0.001, <0.0001, singleton) or benign variants without MAF (ClinVarNew/DoCM and PopSpe_EvalSet : UK10K, SweGen, WesternAsia; Fig 2C).

Fig 3. Evaluation of the different missense prediction tools using simulated and real disease exomes.
Figure 3
A – Distribution of the percentage of predicted deleterious variants in the simulated disease exomes.
B – Ranking of the “causative” deleterious variants introduced in simulated disease exomes.
C – Distribution of the percentage of predicted deleterious variants on the exomes of the MyoCapture project.
D – Ranking of the causative deleterious variants identified in real congenital myopathy exomes from the MyoCapture project.

Fig 4. Distribution of scores for deleterious and benign variants.
Figure 4
The variants of the deleterious and benign evaluation sets were pooled and the distribution of the scores for deleterious and benign variants were represented using violin plots. Red area – distribution of scores for deleterious variants. Green area – distribution of scores for benign variants. Black line – recommended threshold.

How to cite

Chennen K, Weber T, Lornage X, Kress A, Böhm J, Thompson J, Laporte J, Poch O. MISTIC: A prediction tool to reveal disease-relevant deleterious missense variants. 2020 Jul 31;15(7):e0236962. doi: https://doi.org/10.1371/journal.pone.0236962. eCollection 2020.