Structure prediction for protein, RNA, and protein-RNA complexes in one AlphaFold2-derived framework that accepts MSA or language model encoders.
No providers recorded yet. Browse all providers
Deep learning reshaped protein structure prediction, but the machinery that made it work was built for polypeptides. RNA folds under different constraints: its backbone runs through a flexible five-membered ribose ring rather than a rigid peptide plane, and the sequence families supplying evolutionary signal are orders of magnitude smaller. Protein-RNA complexes fell into the gap, served well by neither.
OpenComplex is the Beijing Academy of Artificial Intelligence's answer: a single trainable framework that predicts protein monomers and multimers, RNA monomers and multimers, and protein-RNA complexes end to end. Jingcheng Yu, Zhaoming Chen, Qiwei Ye and colleagues released it as open-source PyTorch code in January 2023 and posted the accompanying technical report as a preprint in March 2025. The architecture inherits from OpenFold's reimplementation of AlphaFold 2, adds a reimplementation of AlphaFold-Multimer, and then extends both with the geometric representations a nucleotide needs.
The unifying move sits in the input pipeline: a joined sequence is split by biomolecule identifier, each chain type gets its own feature extraction — protein templates on one path, RNA secondary structure on the other — and the resulting MSAs and residue pairings are concatenated into one representation. OpenComplex is the root of a continuing line; its diffusion-based successor OpenComplex2 comes from the same lab.
The MSA encoder is a 48-block Evoformer derivative applying rotary position embeddings to node and edge representations in the attention layers. The structure module runs eight blocks pairing invariant point attention with separate intra-chain and inter-chain refinement networks, so multimer geometry is updated alongside monomer geometry. Training drew on the Protein Data Bank as of 18 November 2022 — 193,950 entries containing protein structure and 6,463 containing RNA structure, reducing after filtering to 182,652 protein monomers, 1,704 RNA monomers, and 4,679 protein-RNA complexes, split into training and validation at a 1 May 2020 release cutoff. RNA alignments came from rMSA against Rfam, RNAcentral, and nt, secondary structure from PETfold.
On the CASP15 target R1108, the chimpanzee CPEB3 ribozyme (PDB 7QR3), OpenComplex predicted a structure at 3.55 Å RMSD and TM-score 0.454, against 16.04 Å and 0.302 for the best human-track submission and 14.52 Å and 0.302 for the best server-track submission; it achieved the lowest RMSD of the 111 structures proposed in the human track. On RNA-Puzzles targets excluded from training it matched DeepFoldRNA, both deep learning methods placing ahead of FARFAR2 with ARES rescoring. For RNA dimers it reached sub-2 Å RMSD on the E. coli thiM riboswitch (7TZS), where RoseTTAFoldNA converged on a different base-pairing topology, and across 36 protein-RNA complexes the two methods performed comparably.
The framework suits structural biologists working on assemblies that mix polymer types — ribonucleoprotein particles, riboswitch-protein pairs, viral RNA bound to host factors — rather than stitching together separate protein and RNA predictors. Its second constituency is methods developers: because the repository exposes training as well as inference, it is a base for experimenting with nucleotide representations, RNA loss functions, and language-model encoders without rebuilding an AlphaFold-style stack.
OpenComplex established that AlphaFold 2's architecture generalises past proteins once the frame representation and loss functions are rebuilt for nucleotides, and it did so in permissively licensed code that supports training as well as inference — the starting point for the lab's continuing work on unified biomolecular modelling. The authors are candid about the ceiling: CASP15 showed deep learning RNA predictors trailing physics- and template-based methods on targets without homologous sequences, and their own protein-RNA failure case exceeds 35 Å RMSD on an electrostatically mediated interface despite abundant RNA homologs. The technical report describes a work-in-progress project and has not been peer reviewed. The repository states that training results and pretrained parameters will be released, and no weights have been published, so running the model requires training it.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.