Biotechnology · global
AI-Designed Proteins Enter the Lab: Changjiang Cup Announces Seven Targets, Plans to Validate About 7,000 Sequences
Protein design still faces an experimental hurdle between computational structure prediction and whether a protein can actually bind its target in a test tube. The Changjiang Cup will offer large-scale experimental validation free of charge, but a ranking of binding ability remains some distance from drug efficacy.
Whether an AI-designed protein sequence can be produced, bind a specified target, or even alter cellular responses must be answered through experiments. In early November, the 2026 Changjiang Cup Global AI Protein Design Challenge announced seven targets and its competition rules, with plans to provide free experimental validation for about 7,000 selected sequences, subjecting submitted designs to actual measurements. Registration and sequence submission will close on November 30.
The seven tracks target EGFR, TNF-α, GLP-1R, TfR1, TL1A, TNFR1, and IL-7Rα, respectively, and require either single-domain VHH antibodies or minibinders. According to an announcement from Shanghai Artificial Intelligence Laboratory, VHHs must contain complete framework and antigen-recognition regions and be 100 to 150 amino acids long; minibinders are limited to 40 to 100 amino acids. Participants face design tasks with explicit molecular specifications.
Some tracks take the challenge beyond binding. A summary of the competition rules by DrugAI, published on the BAAI Community platform, states that EGFR designs must have their affinity measured at pH 6.0, while GLP-1R designs must have agonist activity, meaning they can activate the receptor. This distinction directly affects biomedical applications: binding to a receptor is only a starting point; whether a design can elicit the intended response requires separate validation.
Under the published rules, affinity measurements are planned for 1,000 sequences in each track, with the main prizes ranked by experimentally measured dissociation constant, KD. Under the same testing conditions, a lower KD indicates stronger binding. The top 50 entries in each track can also undergo developability and cellular activity testing, but these additional tests do not affect the main prize rankings. The rankings therefore primarily reflect binding ability and cannot be treated directly as an overall ranking of drug development potential.
Competition sponsor Biointron said its RushDATA platform will provide high-throughput protein expression, binding, and developability data, with testing covering thermal stability, aggregation, and nonspecific binding, among other measures. These data can help identify molecules that perform well in predictions but may encounter problems during actual preparation or use. However, the description of the platform's capabilities comes from the supplier and does not represent completed validation results for designs entered in this year's competition.
Shanghai Artificial Intelligence Laboratory announced on November 5 that the competition had attracted more than 1,800 teams and 4,000 participants from more than 20 countries and regions, with research tools and free computing support provided through its Shusheng scientific discovery platform. Biointron announced a total pretax prize pool of RMB 1.4 million. For participating researchers, beyond the prize money, the opportunity to bring designs into the laboratory and obtain measurement data is another important resource offered by the competition.
The announcements so far disclose the competition tasks and validation arrangements but do not yet provide success rates, functional results, or safety data for this year's designs. Even if molecules perform well in affinity and cellular tests, questions about their effects in vivo, immune responses, and manufacturing quality must still be addressed before they can potentially advance toward drug development. The competition is expected to provide a set of design results supported by experimental evidence; how far it can advance biomedical applications will depend on subsequent data.