Two ways to get a machine to sort DNA | AI Learning Guide

Think about how you would tell a computer to spot a suspicious stretch of DNA. The obvious way is to write the condition yourself: look for this exact string of letters near a gene. That works, but only if you already know what to look for. The other way is to hand over thousands of labeled sequences and let the machine work out its own condition. The comparison on this page is the whole idea in miniature. Notice the last row of each column: the hand-written rule can never be better than the person who wrote it, while the learned rule can pick up something nobody described — but you also cannot read it off and check it by eye.

0:00 / 0:00

Auto-play

Rule-based program

The condition is written by a person before the program runs
Applies the same condition to every sequence
Cannot improve from experience
Accuracy is capped by how well the person understood the biology

Learning system

No condition is written in advance
Learns from labeled examples and their outcomes
Adjusts itself as it sees more examples
Can capture signals nobody knew to describe

Neither approach is simply better. A rule-based program is transparent and predictable, which matters when the biology is well understood. A learning system is worth its cost when the pattern is real but unknown, and when enough labeled examples exist to infer it. Much of the DNA signal relevant to disease falls into that second category, which is why learning methods became central here.