Forget the biology for a second, the mechanism is pattern recognition on a noisy signal. A sequencer produces a messy waveform; somewhere in that mess is the true sequence of bases. Classic base callers used carefully engineered signal-processing rules. The 2020 Illumina publication US20200302297A1 instead trains a neural network to map signal to letters directly.
The CPC tags are the giveaway: G06N 3/08 (neural-network learning) and G06N 3/04 (network architecture). This is a general deep-learning method pointed at a very specific, very valuable problem. The assignee, Illumina, dominates DNA sequencing, so improving base-call accuracy by even a fraction translates into real downstream value across genomics.
“The technology disclosed processes input data through a neural network and produces an alternative representation of the input data. The input data includes per-cycle image data for each of one or more sequencing cycles of a sequencing run.”— U.S. Patent Application 2020/0302297 A1 source
Why this is a neat AI-sector story: it's the same neural-network toolkit that powers image and speech recognition, redirected at genomic signal. A model that learns what the signal for a G looks like, in context, is doing exactly what a speech model does when it learns what a phoneme sounds like, in context. The substrate differs; the mechanism rhymes.
Connect this to the broader pattern and you get the AI-for-science thesis in miniature: the value isn't a flashy chatbot, it's a few extra points of accuracy on a measurement that gets run billions of times. That's where deep learning quietly paid for itself in 2020.
Standard caveat: this is a published application, and accuracy claims in a patent are not a peer-reviewed benchmark. But the filing tells you that by 2020 a sequencing leader considered learned base calling core enough to protect, a useful data point about where AI was actually deployed versus merely demoed.