Three Questions a Model Can Be Asked | AI Learning Guide

The comparison table is worth reading column by column, because the input line is what separates these tasks. Property prediction takes a molecule and returns something about that molecule alone — how soluble it is, whether it is likely to be toxic. Activity prediction takes a molecule and a target together, and returns a score for whether they interact. Structure prediction takes a sequence and returns a shape. Now notice the last line of each column. Property data is abundant, which is why those models tend to be reliable. Activity data is sparse and unevenly spread across targets, which is why activity predictions are the ones that most often disappoint. And structure prediction is not producing a number at all — it is producing geometry, so it is judged differently.

0:00 / 0:00

Auto-play

What each prediction type takes in and gives back

Property prediction

Input: a molecule
Output: a number or label, such as solubility or a toxicity flag
Depends mainly on the molecule alone
Training data relatively abundant

Activity prediction

Input: a molecule plus a target
Output: a score or probability of interaction
Depends on the pair, not either one alone
Training data sparse and uneven across targets

Structure prediction

Input: a sequence, or a sequence plus a ligand
Output: a three-dimensional arrangement
Depends on folding and geometry
Output is a shape, not a score

The distinction is not academic. A model can be excellent at property prediction and poor at activity prediction, because the second task requires evidence about pairs that may barely exist in the training data. When you read that an AI model predicts drug-like behavior, the first useful question is which of these three it is actually doing.