DISCUSSION
The interpretation of genomic variation data in clinical settings has been one of the biggest challenges in achieving a successful use of next generation sequencing in medical practice. Given the relevance of this process for current diagnosis of genetic diseases, variant interpretation is one of the most important topics in current bioinformatics. To develop this model, we combined different machine learning techniques and scores from widely used tools, as an effort to try to improve the accuracy of classification of Variants with Uncertain Significance (VUS). Our models showed improved accuracy compared to current solutions analyzing data from a large set of variants previously classified as VUS and distributed over different consequence types..
VUS raises concerns for both patients and for clinicians working on genetic diagnosis. From the patient perspective, genetic testing can yield or confirm a diagnosis, inform the probability of developing a disease for the patient or relatives, and, if the variant is actionable, offer a possibility of treatment. Thus, the uncertainty created by VUS on genetic testing can lead to a variety of emotional responses in patients. Some of the most reported answers to a VUS result are stress and distress, both on patients and on their relatives. Additionally, VUS results are more difficult to understand as patients tend to have a more deterministic view of genetics, and in many scenarios tend to misinterpret VUS results as more similar to a negative result (Clift, et al., 2019). Moreover, due to the variability of patient medical and psychosocial contexts, there is no consensus on clinical best practices to handle VUS results. Some propose to withhold them from patients, so clinicians and labs have to deal with them on a case-by-case basis focusing on pre and post diagnosis counseling to minimize potential harm (Hoffman-Andrews, 2017).
From the clinician point of view, a VUS result raises concerns on how to counsel the affected patient and their family, and how it might change the clinical management. The ACMG guidelines state that a VUS should not be used as part of clinical decision making. Therefore, it is advised that, whenever feasible, the clinician should pursue additional efforts to classify the variant. Additional monitoring and tracing of the patient might be needed if the variant is reclassified (Hoffman-Andrews, 2017). In any case, these efforts involve important time and monetary investments: they can be directed at the patient and family level (i.e. testing for the variant on parents and other relatives) or, if the laboratory has research facilities, functional studies to validate the variant consequence. Other approaches include the work by Sun, et al. (2020), which aims to tackle this problem by proactively creating comprehensive maps of cell-based assays for the missense variants of specific genes, or the work by Walsh, et al. (2019), which compares variant frequency between patient cohorts and reference population cohorts. However, so far these approaches are available for a number of selected genes and diseases. Thus, in resource limited settings VUS prioritization is a paramount need and our tools can help us select VUS with the highest probability of being pathogenic with a high accuracy.
As demonstrated by Liu, Wu, Li, & Boerwinle, et al. (2016), combining information of several predictive scores increases the predictive accuracy of missense variant classification. Here, we show that combining the information of high accuracy conservation-based variant deleteriousness tools like CADD, SIFT, and Eigen (Ionita-Laza, McCallum, Xu, & Buxbaum, 2016) yields improved accuracy across a variety of variant types including missense , splice , intron ,intergenic , and synonymous . However, synonymous variants obtained the less accurate results with both our models and CADD (Figure 3). Recent work suggests that synonymous variants might be more deleterious than would be predicted from current clinical significance annotations (Zeng & Bromberg, 2020). We believe that additional research might be needed to ascertain the true pathogenic potential of these class of variants.