Language Model Assembles Proteins Like a Puzzle express
A protein is like a chain of amino-acid beads that folds into a precise shape, and that shape dictates how the protein works in a cell. Protein designers swap beads to achieve the desired form, but the number of possibilities is enormous. The neural network LigandMPNN produces one variant at a time, like someone assembling a puzzle who places a piece wrong on the first try.
RosettaSearch, a new method, uses a large language model as an idea generator. The model proposes many altered sequences, and the program RosettaFold3 quickly checks how well each folds into the target structure. The search operates under a tight computational budget: like a chef tasting a dish, adding a pinch of salt, and tasting again until the flavor is perfect. In tests on 400 proteins where LigandMPNN produced poor variants, RosettaSearch found improvements in 18–68% of cases, and overall success rose 2.5-fold. The method also works on entirely new, computer-generated structures.
Interestingly, the quality of improvements scales with the model's "reasoning" ability: more powerful versions (o4-mini and Gemini-3) perform better. You can also add vision: the model looks at an image of the predicted protein structure and factors it into generation. This is the first major demonstration that language models can serve as optimizers for protein design without laborious training.
Original abstract
as written by the authors arXiv:2604.17175
We introduce RosettaSearch, an inference-time multi-objective optimization approach for backbone conditioned protein sequence design. We use large language models (LLMs) as a generative optimizer within a search algorithm capable of controlled exploration and exploitation, using rewards computed from RosettaFold3, a structure prediction model, under a strict computational budget. In a large-scale evaluation, we apply RosettaSearch to 400 suboptimal sequences generated by LigandMPNN (a state-of-the-art model trained for protein sequence design), recovering high-fidelity designs that LigandMPNN's single-pass decoding fails to produce. RosettaSearch's designs show improvements in structural fidelity metrics ranging between 18% to 68%, translating to a 2.5x improvement in design success rate. We observe that these gains in success rate are robust when RosettaSearch-designed sequences are evaluated with an independent structure prediction oracle (Chai-1) and generalize across two distinct LLM families (o4-mini and Gemini-3), with performance scaling consistently with reasoning capability. We further demonstrate that RosettaSearch improves the sequence fidelity of ProteinMPNN designs for de novo backbones from the Dayhoff atlas, showing that the approach generalizes beyond native protein structures to computationally generated backbones. We also demonstrate a multi-modal extension of RosettaSearch with vision-language models, where images of predicted protein structures are used as feedback to incorporate structural context to guide protein sequence generation. To our knowledge, this is the first large-scale demonstration that LLMs can serve as effective generative optimizers for backbone-conditioned protein sequence design, yielding systematic gains without any model retraining.
Proteins are built from just 20 types of amino acids, but their combinations yield more possibilities than there are atoms in the universe.
In Stanisław Lem's novel "Solaris," the ocean creates structures on demand — RosettaSearch does something similar for proteins, but so far only in a test tube.