ComplexDesign

ComplexDesign is a hallucination-based framework for multichain protein design that simultaneously folds each protein chain and forms the inter-chain interactions that hold a complex together. It was developed by Jing Xu, Milong Ren, Ning Qi, Xinru Zhang, Zaikai He, Chungong Yu, and Dongbo Bu (corresponding author) at the Institute of Computing Technology, Chinese Academy of Sciences (ICT/CAS), and posted to bioRxiv in June 2026.

Most protein-design methods operate on a single chain or on one binder against one fixed target. ComplexDesign instead casts complex design as a structure-prediction-guided sequence optimization problem: starting from random sequence, it iteratively updates the amino-acid sequences of all chains so that a structure predictor confidently folds them into a coherent multichain assembly. Because the optimization explicitly rewards both intra-chain folding and inter-chain contacts, it can generate dimers, trimers, and tetramers in a single unified procedure rather than stitching together separately designed parts.

The defining novelty is a specialized masking mechanism that lets the method explore the relative arrangement of two target proteins during optimization. This makes ComplexDesign particularly suited to designing a "bridging binder" — a designed protein that recruits two distinct target proteins simultaneously, pulling them into a ternary complex. That capability connects the work to therapeutic strategies such as induced proximity, where forcing two proteins together (for example, an effector and a target) is the desired mechanism of action.

#Key Features

• Multichain sequence hallucination: ComplexDesign performs structure-prediction-guided sequence optimization that folds each chain and forms inter-chain interfaces at once, producing dimers, trimers, and tetramers within a single framework.
• Masking mechanism for relative arrangement: A specialized masking scheme lets the method explore how two target proteins are positioned relative to each other, which is the key to designing binders that bridge two targets.
• Bridging (multi-target) binder design: The framework can design a single binder that simultaneously engages two distinct target proteins and pulls them into a ternary complex, a setting most binder-design tools do not address.
• Frozen-predictor, fine-tune-free design: Like other hallucination methods (HalluDesign, PXDesign), ComplexDesign trains no model of its own; it drives design entirely through a pretrained, frozen structure predictor used as the scoring engine.

#Technical Details

ComplexDesign is an inference-time optimization framework rather than a trained generative model: it has no ComplexDesign-specific pretrained weights and instead drives design through the forward pass of a frozen, pretrained structure predictor, in the same hallucination paradigm as HalluDesign and PXDesign. The method optimizes the sequences of all chains in a target complex so that the predictor folds them into the intended multichain assembly, using its confidence and predicted geometry as the optimization signal; a specialized masking mechanism governs how the relative arrangement of two target proteins is explored so that bridging interfaces can form. On in-silico benchmarks the authors report greater than 50% design success across dimers, trimers, and tetramers, and for the harder multi-target setting they report successful bridging-binder designs for 8 of 10 target pairs. As reported in the preprint, code "will be released upon publication" and is not yet publicly available.

#Applications

ComplexDesign targets protein engineers and structural biologists who need to design multichain assemblies or binders that engage more than one target at once. The most distinctive use case is the bridging binder: a designed protein that simultaneously recruits two target proteins into a ternary complex, which is directly relevant to induced-proximity therapeutics, synthetic signaling, and the assembly of designed protein machines. Because the approach is fine-tune-free and driven by a frozen structure predictor, groups can in principle apply it to new target pairs without assembling training data, once code is released.

#Impact

ComplexDesign extends the hallucination paradigm — previously applied mostly to monomers and single-target binders — into explicit multichain and multi-target design, addressing the underexplored problem of designing one protein that bridges two others. Reported benchmark results (over 50% success on dimers through tetramers, and 8 of 10 validated target pairs for multi-target binders) suggest the masking-based exploration of relative arrangements is an effective handle on ternary-complex geometry. Important caveats apply: the work is a bioRxiv preprint (distributed CC BY) whose results await peer review and independent reproduction; success depends on the quality of the underlying structure predictor's confidence signal; the reported successes are in-silico rather than experimentally validated binders; and the code is not yet available, so reproduction is currently not possible.