Rainbow is a proprietary CNV detection platform powered by advanced AI algorithms, designed to provide highly reliable copy number variant (CNV) detection and analysis.
Rainbow by Franklin provides powerful CNV capabilities such as:
Copy Number Variants in small Panels / WES / WGS originating from Germline and Somatic samples
Detection down to two exons deletion or duplication (heterozygous) or one exon (homozygous) resolution
The workflow:
Rainbow creates a unique model by using a cohort of >30 samples from more than one batch of the same kit and laboratory - in order to reduce the batch effect
Once the model is defined, the workflow is fully automated from raw data to variant calling and interpretation
The model is frequently re-trained by using the latest samples for best optimization and accuracy
Franklin's in-house advanced AI algorithm "Rainbow"
For each exon, Rainbow builds a predictive model using over 50 unique “predictors”: other exons or regions whose coverage shows a strong statistical correlation with the exon in question.
Predictors are chosen from different chromosomes and genes to maximize robustness.
The model defines the expected coverage ratios between each predictor and the target exon.
The actual coverage is then compared against these ratios to estimate the exon’s copy number.
This approach enables Rainbow not only to predict an exon’s or region’s copy number, but also to calculate a prediction score, which indicates how reliable the model is for that specific exon or region.
CNV Confidence
Franklin’s Rainbow assigns a confidence level to each detected CNV: Failed, Low, Medium, or High. This score is based on several factors:
The predicted copy number for the exon or region.
The number of consecutive targets (exons or regions) that the CNV is composed of.
The prediction score of the model.
The level of agreement among predictors.
Each confidence level reflects the likelihood that the CNV is real, helping users assess the reliability of the call:
Failed confidence ~1-5% (probably false)
Low confidence ~5-60%
Medium confidence ~60-85%
High confidence ~85-99%
The variant confidence widget visualizes 3 AI-inferred confidence parameters:
Depth (median)
Predicted copy number
Prediction score
The aggregate values for each parameter (across all regions that the CNV is composed of) are shown at the top of each section
The sample under exploration is highlighted in blue, while additional samples using the same CNV model are displayed in the background:
Family members (if available) → light blue
Other reference samples → grey
By comparing the highlighted sample against family members and reference samples, users can better assess the reliability of the CNV and perform a more informed evaluation.
The genomic range is shown in the top menu. The user can adjust the view by directly editing the range or using the navigation buttons (zoom in/out, move left/right).
Sample filtering: additional samples can be included or hidden using the dropdown menu in the top-right corner.
Hovering over an exon point in the sample’s graph reveals the exact values for that exon.
Depth (median)
The median depth of a CNV region is calculated by averaging the median depths of all exons in that region. Hovering over the data points will display the exon-level median depth for each sample.
Predicted copy number
The variants predicted copy number, as calculated by the model:
3 - Heterozygous duplication
2 - Normal
1 - Heterozygous deletion
0 - Homozygous deletion
An exon-level predicted copy number of each sample can be observed by hovering the dots.
Prediction score
The Prediction Score reflects the potential accuracy of the CNV model in a given variant region. This score is derived directly from the model, so it is the same across all samples. At the exon level, the user can view the prediction score by hovering over the dots in the graph. A low Prediction Score indicates weaker model performance and lower confidence in the CNV call for that region.
Prediction Score | Model prediction performance |
1 | perfect |
0.95-1 | high |
0.85-0.95 | medium |
0.7-0.85 | low |
0-0.7 | None (model is not available) |
A low prediction score is usually caused by one of the following reasons:
Low coverage correlation to predictors
Complex genomic regions including homologies, repeats GC-reach (usually affect also the coverage)
Polymorphic copy number regions
Some of the exons might have lower prediction scores, meaning that a CNV in this specific region is hard to predict.
Variant confidence assessment (WES):
Still have questions? Reach out to our Support Team, they'll be happy to help!