What the published comparisons actually show | AI Learning Guide

The distinction that matters most here is between recovering known answers and producing new ones. In a retrospective study, the model is scored on antibodies that were already characterized, so a good result may only mean the test set looked like the training set. A prospective study is harder: the model proposes sequences nobody has made, and those sequences are then expressed and measured. That is the only design that supports a claim about real design ability, and it is also the design that is expensive and therefore rare. So when you read that an AI designed a working antibody, the first question is which of these two it was.

0:00 / 0:00

Auto-play

Two kinds of evidence, two different claims

Retrospective evaluation

Model is tested on antibodies already characterized in databases
Measures whether the model recovers known binders
Cheap, fast, and common
Cannot show that a genuinely new antibody would work

Prospective evaluation

Model proposes sequences that have never been made or tested
Those sequences are expressed and assayed in the lab
Expensive, slow, and less common
The only design that supports a real performance claim

AI-designed antibodies have been shown to bind their targets in real experiments, but the strongest claims rest on a small number of prospective studies. Most published comparisons are retrospective, and true head-to-head trials against human-designed antibodies on the same target are rare. The defensible statement is that AI design works for some targets, not that it outperforms human design in general.

A high success rate in a retrospective study is partly a measure of how well the test set resembles the training set. When the same databases supply both, strong scores can reflect memorization of familiar antibody families rather than general design ability.