Why the Search Is So Hard | Visual AI Learning Guide

Look at the nesting here. The outer region is every possible sequence of the binding region — roughly a hundred positions, twenty choices each, so about twenty to the hundredth power. That is more sequences than there are atoms in the observable universe, so enumeration is out of the question. Now look at the middle region: sequences that actually fold into a stable protein. It is a small slice of the whole, because most random sequences do not fold well. And inside that, the innermost region is the sequences that fold and also bind the chosen target tightly and specifically. That innermost region is what a designer is hunting for. The lesson is that any method has to be guided, because blind search would spend essentially all its time in the outer region.

0:00 / 0:00

Auto-play

The difficulty of antibody design comes from three stacked problems. First, the number of possible antibody sequences is astronomically large — a typical binding region is roughly a hundred amino acids long, and with twenty possible amino acids at each position the count of distinct sequences runs to about \(20^{100}\), a number with more than a hundred digits. Second, almost none of those sequences fold into a stable protein at all; a random string of amino acids is far more likely to clump or fall apart than to form a well-behaved structure. Third, even among sequences that do fold, only a tiny fraction present a surface patch that binds the intended target with good specificity and affinity.

The practical consequence is that you cannot search this space by enumeration. Any design method — human or machine — must use some form of guidance to avoid wasting effort on sequences that were never going to work. That guidance is exactly where human approaches and AI approaches diverge, and it is the subject of the rest of this course.