Scalable and universal biomolecular
inverse folding with ComplexMPNN

Yijie Zhang1,6,† Pengcheng Li3,† Yanyi Wang8,† Lingdong Shen2,† Tongyue Xu11,12,† Wenjie Yan3,† Qi Liu2 Dongxu Zhang3 Jue Wang7 Tang Liu10 Shaohua Ma3 Asher Qin5 Odin Zhang13 Conghui He4 Hongmei Wang8,9,* Jinren Ni2,* Mathieu Blanchette1,6,* Cheng Tan5,*
1McGill University 2Peking University 3Tsinghua University 4Shanghai Artificial Intelligence Laboratory 5Tencent 6Mila – Quebec AI Institute 7Zhejiang University 8University of Chinese Academy of Sciences 9Beijing Institute for Stem Cell and Regenerative Medicine 10Shenzhen University 11Tongji University 12Shanghai Innovation Institute 13The Chinese University of Hong Kong

†Equal contribution  ·  *Corresponding authors

Generation

Human SRP19 bound to helix 6 of SRP RNA, with two magnesium ions included as context. Generation starts with all designable positions masked. At each step the decoder predicts amino-acid and nucleotide probabilities for every designable position, fills a selection of the masked ones by sampling from those probabilities, and masks a few filled positions again so that their identities can be revised in later steps. The protein and the RNA are updated together, so the amino acid or nucleotide chosen at one position becomes part of the context for predicting its neighbors, including those on the partner chain.

t = 0.00 0 of 0 filled
Protein
RNA
masked filled this step filled masked again Mg
Overview of ComplexMPNN
a, A hierarchical encoder represents residue contacts, chain fragments and ligand and metal-ion context. A coupled flow decoder turns that representation and the current sequence into amino-acid or nucleotide probabilities, from every designable position masked at t = 0 to a complete sequence at t = 1, revising earlier choices along the way. b, One complex, human SRP19 bound to helix 6 of SRP RNA (PDB 1JID), under four design masks: the protein, the RNA, both chains, or individual sites. Blue marks designable positions, gray fixed ones, and the orange spheres are two magnesium ions given as context.

Abstract

Inverse folding uses a target biomolecular structure to design sequences that can fold into it. In complexes, sequence design also needs to account for interactions between components. Here we present ComplexMPNN, a unified model for protein, RNA and DNA inverse folding. Its hierarchical encoder integrates structural and chemical information across the complex, and its coupled flow decoder refines sequences at selected positions. It can design individual chains or multiple components together and complete partially specified sequences. Across monomer and complex benchmarks, ComplexMPNN improves sequence recovery and refolding performance. Context analyses show that its predictions respond to partner sequences and local chemistry. Experimental tests identified rhodopsin designs with pigmentation indicating successful expression and zinc-finger designs with DNA-binding activity. By allowing interacting proteins and nucleic acids to be designed together, ComplexMPNN provides a basis for developing new recognition pairs for gene regulation.

Results

One model is evaluated on protein, RNA and DNA monomers, on protein–protein complexes and on both directions of protein–nucleic-acid design, against specialized methods retrained on the same training set. Designs are refolded with AlphaFold3, without multiple-sequence alignments or structural templates, and compared with the target structures.

On monomers, ComplexMPNN reached higher sequence recovery than the protein-specific methods across all three length groups, and the refolding success rate on long proteins was 32.0 per cent against 22.2 for PiFold. On RNA it reached higher recovery than RDesign, gRNAde and RhoDesign, with lower RMSD and higher TM-scores; on DNA, higher recovery and a higher refolding success rate than StructGNN, GraphTrans and PiFold.

Monomer sequence design
Monomer sequence design. Sequence recovery on the protein, RNA and DNA benchmarks (1,340, 54 and 23 targets), and RMSD, TM-score and refolding success rate for the designs after AlphaFold3 refolding. ComplexMPNN-Monomer is the same architecture trained on one molecular type instead of all three.

In protein–protein complexes, with all protein chains designed together, recovery was about 3 percentage points higher than ProteinMPNN retrained on the same chains, with TM-score and RMSD improving as well. Against LigandMPNN, with the nucleic acid fixed, recovery was comparable and refolding more accurate: on the subset whose native-sequence predictions reach ipTM ≥ 0.6, median RMSD was 1.05 Å against 1.44. Designing the RNA instead, with the protein fixed, beat gRNAde overall and every method at 100 nucleotides or more.

Sequence design in biomolecular complexes
Sequence design in biomolecular complexes. a–c, Protein–protein complexes. d–f, Protein design with the nucleic acid fixed, on both test sets. g, h, RNA design with the protein fixed; RhoDesign and gRNAde read the RNA alone.

What the partner contributes is local. Removing it costs the most recovery at the residues closest to it; shuffling the nucleotide identities while keeping the coordinates costs 2.9 percentage points within 3 Å and 1.2 across the 5 Å interface. Providing magnesium raises the predicted proportion of Asp and Glu nearby, and providing zinc raises His and Cys.

Molecular context and case studies of complex design
Molecular context and case studies. Distance to the partner (a), pretraining on the other nucleic acid (b), predicted composition near magnesium and zinc (c) and at the nucleic-acid interface (d, e), and two designs refolded beside their native complexes, ToxR–DNA and an ancestral Dicer–RNA (f, g).

Experiments

Five of 20 microbial rhodopsin designs produced distinctly pigmented pellets against the empty-vector control, from three templates and at 69 to 74 per cent global identity to those templates. The positive designs kept the retinal-binding lysine and the conserved motif in the third transmembrane helix.

Experimental validation of generated microbial rhodopsins
Microbial rhodopsin designs. The design and screening workflow (a), the pellets (b), the conserved motifs in the third and seventh transmembrane helices (c), and where the designs sit among reference sequences (d).

The ZIF268 protein was redesigned with its DNA duplex and the three structural zinc ions held fixed. All four selected designs were expressed and purified, and two of them, pred14 and pred4, shifted the target probe in an electrophoretic mobility shift assay.

Purification and EMSA validation of designed zinc-finger proteins
Designed zinc-finger proteins. Purification of the four designs (a–d), native ZIF268 as a positive control (e), the initial EMSA screen (f), and titrations of the two designs that shifted the probe (g).

Get started

git clone https://github.com/OpenRaiser/ComplexMPNN.git
cd ComplexMPNN
pip install -r requirements.txt

python complexmpnn/design.py \
  --input examples/1jid.pdb --out_dir out/design \
  --ckpt weights/complexmpnn.pt \
  --design_types protein \
  --temperature 0.1 --flow-steps 50 --seed 111