Functional Predictions & Conversation Scores in VSClinical

In our previous webcast, we discussed the splice site algorithms for clinical genomics within VSClinical. We took it a step further in yesterday’s webcast and looked at the functional predictions and conservation scores. We had a great turnout for this event with lots of great questions from the attendees. I’d like to recap our Q&A for anyone else who might be interested in learning more about the functional prediction and conservation scores in VSClinical.

Do synonymous or intronic variants get ranked differently? Does RNA folding get taken into consideration for a deleterious classification?

None of the algorithms take RNA folding into account. Conservation scores are computed for every variant, whereas SIFT & PolyPhen only run for missense variants. For intronic variants, only impact on spicing and conservation is taken into account. If all of the algorithms predict the variant to disrupt splicing, then PP3 is recommended. For synonymous variants, only splicing disruption is considered evidence for PP3.

There are differences in predicting variant effect. How do you decide if the even prediction is completely conflicting with each other?

So basically, you’re asking how is the evidence conflicting? If 3 of 4 of the splicing algorithms predict a variant to be disruptive, then we recommend PP3 regardless of what’s predicted by the functional prediction and conservation scores. If the conservation scores and functional prediction algorithms are both consistent with deleterious effect, then PP3 is recommended regardless of the splice site algorithms. In this way, we treat different variant effects as isolated types of computational evidence. Only agreement amongst algorithms of the same type is required to recommend PP3.

How about non-functional variants in the regulatory region. Most of the tools you mentioned are extronic.

The only pieces of evidence that we apply to intronic regions are conservation and splice prediction.

How do you compare these methods as they use different criteria and approach?

Excellent question – they definitely use a different approach and different criteria. Our goal here is just to find what’s the best classifier for determining if a variants going to be pathogenic or not. Ultimately, all we care about is getting the best predictions we can. And, as we’ve seen from our results, even if you use many different criteria (as with at PolyPhen2, which incorporates numerous different pieces of evidence), it only seems to give very modest improvements over simple methods that are just looking at amino acids probability scores such as SIFT.

Does VSClinical support automatic variant classification using the ACMG guidelines?

Yes, it definitely does. Without even using our VSClinical workflow at all you can go ahead and import your variants into VarSeq and you can run our automatic ACMG guidelines classifier, assuming that you have the ACMG guidelines VSClinical license. Then, you can filter on these classifications so, for example, you can filter out any variants that were automatically classified as benign to get down to a useful set of variants based on all of the automatic criteria that we showed in this webcast.

Can functional predictions and conservation scores be used to sort and filter variants in VarSeq?

Yes they certainly can which is kind of related to the previous question – anytime you run an annotation against a gene track, we’ll compute splice site predictions for you and likewise we’ll actually ship with tracks for all of our functional prediction scores and conservation scores as well for all of the clinically relevant transcripts for every single gene. So, you can go ahead and annotate against these and for every single variant y,ou can obtain a functional prediction score/conservation score that you can then filter on or sort by within VarSeq.

Are the predictions transcript specific? If I want to use a different transcript in my lab, how do I do that?

You can definitely use different transcripts, the predictions are done for the clinically relevant transcript by default. But, with VSClinical, you can always change your analysis to be switched over to any transcript and you can of course save those preferences for next time so when you see a variant in that gene, you can use your preferred transcript instead of the default clinically relevant one. All of the functional prediction and conservation scores can be rerun from scratch on that selected transcript.

Can you change the prediction tool cutoffs to customize it? I typically used a CADD score much higher than 5.

Within VarSeq, for your filter chain, you can actually customize these cutoffs to be whatever value you would like them to be. So, when you’re actually filtering your variants within VarSeq to get down to a set that you’re interested in, you can adjust those thresholds. In terms of actually looking at VSClinical, the computational evidence that we incorporate are SIFT and PolyPhen scores for automatic classification. But, we provide all of that score information for you so you can actually look at the score yourself by hand and if you, let’s say, decide that our thresholds too high, you can actually override our recommendation based on the score you see there.

Is this VSclinical part of VarSeq or separately sold?

It is sold as a separate license, but it is fully integrated into VarSeq. So once you have that license, you will have complete support for using the VSClinical workflow right within your VarSeq product. If you are interested in learning more about the licensing structure for this, please email our team at!

5 thoughts on “Functional Predictions & Conversation Scores in VSClinical

  1. Wang Ling

    Hi Nathan, thank you for your sharing. I think it’s great. So what do you think needs attention when using VSclinical? In which operation may be the user’s own operational errors may occur?

    1. Golden HelixGolden Helix

      In general, there are a number of caveats that must be considered when following the ACMG guidelines. If not properly considered, these caveats could result in user error, contributing to incorrect classifications. For instance, PVS1 is applied to null variants in genes where LOF is a known mechanism of disease, but should not be applied to variants in or near the last exon of the gene. For the most part, we identify these caveats for you and provide informative warnings to the user to limit operational errors.

      However, some criteria require special attention from the user. For example, PS1 is applied when a variant has the same amino acid change as a previously established pathogenic variant. We will recommend these criteria when a matching pathogenic variant is found in ClinVar, but we recommend that the user review the literature and evaluate the evidence for pathogenicity. While VSClinical will provide the user with the relevant ClinVar links to facilitate such review, ultimately such evidence must be evaluated manually, and if the evidence cannot be confirmed, the criteria PP5 should be applied instead.

  2. Wang Ling

    Thank you very much.

    In addition, it was mentioned that Clingen suggested that PP5 and BP6 be removed from the ACMG guidelines, and that the VSClinical would also make appropriate adjustments?

    1. Golden HelixGolden Helix

      Great question – we’ve decided against removing these two criteria and have instead opted to recommend that the user upgrades PP5 to PS1 when the relevant evidence is available. However, we allow users to ignore any criteria, so there’s nothing preventing a lab from excluding these two criteria.

      1. Wang Ling

        Thanks !

        I have a question. Since we have already run the ACMG Classifier and clicked on VSClinial in the tab, we already have the evidence. the classificatioin, the ACMG Classification Criteria. Why do we have to click “Start New Evaluation” again? What kind of variant is usually need to “start a new evaluation”? when or at what conditions should “some criteria require special attention from the user”?


Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.