DeepMind's AlphaFold 2 Makes Strides but Does Not Solve Protein Folding
Despite winning the CASP14 competition, DeepMind's AI system only matches experimental structures about two-thirds of the time. Experimental biologists are still essential for verifying these computational predictions.
DeepMind announces that its artificial intelligence system, AlphaFold 2, wins the latest CASP14 competition for predicting protein structures. The media responds with breathless reports claiming that the 50-year-old protein folding problem is finally solved. However, a closer look at the actual results reveals a much more nuanced reality.
While DeepMind clearly outperforms all other computational teams in the contest, the system only produces predictions comparable to experimental structures about two-thirds of the time. This means researchers do not know which specific predictions are accurate until they compare them against actual experimental data. A 67 percent accuracy rate represents a significant step forward, but it falls short of completely solving the complex scientific challenge.
Because of this limitation, experimental scientists like protein crystallographers and cryo-electron microscopists remain essential to the field. Their careful laboratory work is still required to verify which computational models are correct. Until AI can guarantee highly accurate predictions on its own, experimental validation continues to play a crucial role in structural biology.