AlphaFold 3: Overhyped with Limited Improvements
A recent headline announced DeepMind's AlphaFold 3 as revolutionary, but a closer examination by leading bio labs suggests otherwise. Despite the excitement, AlphaFold 3 appears overhyped, with only marginal improvements over AlphaFold 2, and remains far from replacing traditional methods like Cryo-Electron Microscopy.
Initial Evaluation
In a preliminary assessment at CTOL.digital, we predicted the MST1 protein structure using AlphaFold3 and compared it with 3COM data from the RCSB PDB Protein Data Bank using PyMOL. The results revealed that AlphaFold 3's predicted MST1 structure and the published structure (PDB: 3COM) were poorly aligned, with an RMSD value of 14.892 Å over 934 aligned atoms. This significant discrepancy points to substantial differences between the two structures. Our continued analysis of several non-trivial proteins also yielded disappointing results, which were corroborated by other researchers in China. However, we remain open to the possibility of incorrect usage and welcome corrections. Regardless of the test result, AlphaFold3's structure generation speed amazed our whole lab.
Results from Other Laboratories
Other labs have reported similarly underwhelming outcomes. According to sources, AlphaFold 3 struggled with the classic STRING and cGAMP ligand tasks, delivering a low "binding score." Nieng's lab also noted minimal improvements in "Protein-Protein Interaction" over the previous generation. Another source tested RNA structure prediction and reported unsatisfactory results. While the protein-ligand tasks showed acceptable performance, they were far from the "magical" marketing claims.
Limited Functionality and Strict Terms of Service
Users have expressed frustration over the limited ligand choices in AlphaFold 3, further compounded by explicit bans on future molecular docking research. Molecular docking is crucial in translating structural data into functional insights, particularly for drug discovery. Widely used tools like Glide and AutoDock are explicitly prohibited in AlphaFold 3's terms of service.
Analysis: Overhyped Marketing by Tech Giants
Tech giants are increasingly over-marketing their AI solutions. OpenAI, for instance, has teased multiple new tools for months without public release. Facing pressure to reclaim its AI dominance, Google's launch of Gemini and the overhyped promotion of AlphaFold 3 have not brought the company closer to its goal. We acknowledge the fact that, to win in the heated financial market, it is crucial to tell a good story and constantly catch the public attentions. However, this can also create a future hard landing on the stock price.
Opinion: AlphaFold's Real Value is Questionable in Pharma
AlphaFold predicts the protein structure in a probablistic approach. However, in Pharma industry, a atom-level accurary is required. Only time will tell if AlphaFold3 is truly valuable in industries.