DeepMind's AlphaFold Dominates Protein Structure Prediction at CASP13
DeepMind's AlphaFold takes first place at CASP13 by a wide margin in predicting novel protein folds, leaving the academic community stunned.
DeepMind makes a stunning debut at CASP13, the biennial assessment of protein structure prediction methods, by taking first place and leaving a comfortable distance between themselves and the second-place predictor. The victory occurs specifically in the free modeling category, which focuses on predicting novel protein folds that lack known structural templates. This dominant performance prompts researchers at the conference to ask each other exactly what just happened.
The AlphaFold result sparks a complex mix of excitement and melancholy among academic scientists who previously led the field. While initial reactions involve tribal reflexes and existential questions about the future of academic research in this discipline, the mood shifts toward rational appreciation of the scientific progress. The author notes that the sheer scale of DeepMind's success forces the community to reconsider the trajectory of protein structure prediction and the role of deep learning in the life sciences.
Meanwhile, the author's own end-to-end differentiable model, known as RGNs, struggles during the competition. This poor performance stems partly from a sudden increase in the value of co-evolutionary information at this specific CASP and partly from technical problems that prevent the submission of the original, unaltered predictions. The post promises a deeper dive into the specific methodological advances of AlphaFold and how it compares to existing approaches.