Corey
by Corey
1 min read

Categories

  • articles

Tags

  • Bioinformatics

Several days ago, we found that one specific gene of Copy Number Variation(CNV) in Life sequencing platform is not so accurate as we expect.

We used to construct a baseline from several negative samples (CN=2; N>10), which is employed to call CNV of every new samples in a long time. But the problem is that the CNVs of new samples become increasingly biased from the FISH result (a gold-standard of CNV) along with the time going. We think that the problem is everything is changing except for this baseline, especially baseline is extremely biologically sensitive to the experimental condition, pipeline, enzyme.

One of my genius colleagues has reached out an idea, why not convert static baseline to dynamical baseline. So when we want to call a CNV of a specific sample, an automatic benchmark is yielded from the same batch of this sample(ignoring positive or negative). After first calling, remove the samples which possess the highest and lowest CN value and reconstruct the baseline. At last, utilize the latest benchmark to call CNV of the sample.

I am pleased to be able to work with this guy. Now he is decided to leave, and I believe he has his own choice and plan. Wish him a good future.