AI・テクノロジー
AlphaFold 3.5 Released — Protein×Nucleic Acid Complex Prediction Exceeds 90%, Reshaping Drug Discovery Design
機械翻訳 / Machine-translated
機械翻訳 / Machine-translated
Google DeepMind released "AlphaFold 3.5" on September 10, 2026. Raising prediction accuracy approximately 12 points above the previous-generation AlphaFold 3, it achieved 90.2% accuracy in predicting the complex structures of proteins and nucleic acids (RNA and DNA). Three major pharmaceutical companies have already announced pipeline integration, marking what is seen as an inflection point at which the "discovery phase" of drug development becomes effectively AI-driven.
DeepMind published a paper in Nature Biotechnology dated September 10, 2026, and simultaneously released the API to the public. Key improvements are as follows:
"Structure confirmation that used to take three days is now done before lunch. The scheduling of wet labs has changed from the ground up." (Pharmaceutical researcher, quoted anonymously from a post on X)
The AlphaFold series largely solved the accuracy problem for single-protein structure prediction with version 2 in 2020. AlphaFold 3, released in 2024, added support for intermolecular interactions (protein × small molecule, protein × nucleic acid), significantly improving the accuracy of drug target identification.
In practical drug discovery applications, however, "dynamic behavior of complexes" and "prediction speed" remained persistent bottlenecks. Pharmaceutical laboratory workflows still required verification via X-ray crystallography or cryo-electron microscopy (cryo-EM) after AI predictions, and the view that "development timelines aren't shortened even with AI" remained widespread.
What AlphaFold 3.5 emphasizes is that surpassing 90% accuracy in complex prediction makes it realistic to "proceed to the next synthesis step without cryo-EM verification." Novo Nordisk, AstraZeneca, and Daiichi Sankyo have already announced trial integration into their pipelines.
In DeepMind's internal trials, the "target identification → hit compound selection" phase of small-molecule drug discovery was shortened by an average of 52%. The industry standard places this phase at two to four years, and if a compression on the scale of one to two years is realized, the front-loaded burden of clinical trial costs would be reduced. In new drug development, where average costs are said to reach around 260 billion yen, intervention at this stage has a direct managerial impact.
With AlphaFold 3.5 now offering full-fledged support for complex prediction with RNA, applications in target discovery for siRNA and mRNA therapeutics are expected to expand. The mRNA therapeutics market is estimated at $22 billion as of 2025, and improved prediction accuracy has the potential to fundamentally change the cost structure of development in this space.
The reduction from $0.38 to $0.12 per query means the cost has dropped below the threshold at which startups and university labs with limited capital can commit to production use. The free tier for academic institutions (500 queries per month) also continues, and we are entering a phase in which the access gap in drug discovery AI is rapidly narrowing.
AlphaFold's predictions remain static structural snapshots. In living organisms, structures fluctuate with temperature, pH, and competing molecules, so "high-accuracy prediction ≠ an effective drug" is not a direct equivalence. Version 3.5 has not improved on this point either, and integration with dynamic MD simulation is seen as an ongoing challenge.
The evolution of AlphaFold is a textbook example of "one AI clearing a bottleneck in an industry." Pharmaceutical companies have been running wet labs and AI predictions in parallel, but now that complex prediction accuracy has reached practical levels, reducing wet lab workloads is expected to emerge as a management-level agenda item.
Specifically, over the next 12 to 18 months, we are likely to see a movement in which "the role of structure-verification specialist teams is redefined." This is not a conversation about headcount reduction; rather, it is more likely that a resource reallocation will occur in which cryo-EM specialists focus specifically on "screening AI prediction outliers" and "verifying dynamic behavior."
Within Japan, Daiichi Sankyo has announced integration, and industry observers believe Takeda Pharmaceutical, Eisai, and Shionogi are also in the evaluation stage. For biotech startups, including "AlphaFold 3.5 integrated" in fundraising pitches may become standard practice from Q4 2026 onward.
At the same time, DeepMind's own commercial strategy warrants close attention. While lowering API prices accelerates adoption, how the lock-in structure is designed once pharmaceutical companies embed the tool in their own infrastructure has not yet been fully disclosed.
AlphaFold 3.5 has simultaneously updated all three metrics — accuracy, speed, and cost — marking the moment at which drug discovery AI transitions from "research support tool" to "core infrastructure for the development process." Challenges such as dynamic structure prediction and in vivo correlation remain, but with multiple major pharmaceutical companies moving to production integration, the industry's design philosophy is beginning to shift.
The next thing to watch is how cryo-EM equipment manufacturers and structural analysis CROs respond to this improvement in accuracy and how they redefine their business models.
This article was written by an AI writer (AI News) from the Mirai News editorial team.
