The lab of forensic geneticist Susan Walsh has pioneered some of the most accurate pigmentation-prediction systems currently available to anthropologists and law enforcement. But at the 31st annual ISHI conference, Walsh gave virtual audience members a sneak peak into her laboratory’s latest project—quantitatively predicting eye color from phenotypic and genotypic information.
“We are trying to develop and bring iris color prediction from DNA to the next level, toward continuous eye color prediction,” Walsh said during her presentation on Wednesday.
In 2018, Walsh and her lab debuted the HIrisPlex-S DNA test system, which can simultaneously predict eye, hair and skin color phenotypes from small amounts of DNA, whether left at a crime scene or swabbed from archeological remains. Her current project reaches beyond that with the creation of a trained quantification program that can break down the percentage of colors in the iris.
“We have to look at the color space and where it lives, and then we can understand the ranges that exist within the irises and if we’re capable of actually predicting these ranges,” Walsh explained.
That’s why Walsh’s program is based on HSV color space—H for hue (different ranges/classes of color), S for saturation (intensity of color), and V or B for brightness (light/dark of color). The researchers fed the program training images based on HSV ranges, telling it to group pixels into four classes of color—gray/blue, green, light brown and dark brown.
Once each pixel is assigned a color class, output comprises a visual quantification and a written log of pixel counts per color class. For example, in the photo above, the program gave a readout of: blue/gray: 78.3%; dark brown: 13.2%; green: 5.4%; and; light brown: 3.1%.
Of course, Walsh wanted to add genotypic information to the phenotypic database her lab just compiled. So, they recruited 3,300 individuals from Europe, Africa and East Asia, gathering close-up images of their irises. According to their literature search, there are about 1,500 gene variants associated with all pigmentation traits (skin, hair, eyes). With that in mind, the team finalized a list of genotyped and imputed variants for 1,300 SNPs for 33,300 individuals.
The researchers are now working on prediction model building utilizing linear regression models and neural networks. They are also fine-tuning feature selection to “make sure the variants we are introducing into our model are really responsible for each defined class of color.”
Walsh said the results generated during a recent test were encouraging. During the assessment, the researchers fed an iris image to the quantifier tool to produce a measure of predicted percentage for the four color classes.
“We then used our prediction model to assess this individual based on their genotypic data. Could we replicate the actual values?,” Walsh explained.
While there was a high proportion of errors in some of the colors, overall, the estimating percentages were good. In the photo to the right, the images inside the gray box are the real-life images with actual and predicted percentage values per color. The images outside the box are those chosen by the prediction model. As you can see, in most cases, they match pretty well.
“This quantitative way of determining color takes out the subjectiveness that individuals would have in grouping irises into certain categories. We wanted to avoid that, we wanted the computer to dictate how it sees the irises once we trained it,” Walsh concluded.
Photo credit: Susan Walsh/ISHI