How reliable are AI models like AlphaFold for drug discovery, and where do their predictions still fall short?

AI models like AlphaFold have significantly accelerated research in drug discovery, but they are not a replacement for experimental methods [1]. Their reliability is limited by several key factors:

• Accuracy Gaps: Even for predictions with the highest confidence scores, researchers have found that these models can have approximately twice the errors of high-quality experimental structures. About 10% of these high-confidence predictions may contain very substantial errors [3].
• Structural Limitations: These models often struggle with precise side-chain orientations, which are critical for structure-based drug design [2].
• Novelty and Complexity: Predictions tend to be less reliable when dealing with novel protein folds or disordered regions of proteins [2].

Because of these gaps, experimental structure determination remains a necessary step in the drug discovery process [1].