How to Use Proteina-Complexa

Proteina-Complexa is a generative binder design platform that combines flow-based protein generation with inference-time optimization and structure-prediction validation. The ICLR 2026 paper presents it as a unified alternative to separate generative and hallucination workflows, with strong performance on protein binder, ligand binder, and motif scaffolding tasks.

On Neurosnap, researchers choose a Design Mode and provide the structure inputs required for that workflow: protein targets for binder design, ligand-containing complexes for small-molecule binder design, motif-and-ligand inputs for AME scaffolding, or motif structures with contig definitions for monomer motif scaffolding. Shared advanced controls expose search depth, sampling steps, and filtering limits without requiring manual pipeline configuration.

Methodologically, Proteina-Complexa extends flow-based latent protein generation with reward-guided search and downstream refolding. Protein binder campaigns are evaluated with AlphaFold2-style confidence and structural consistency metrics, while ligand binder and AME workflows rely on RoseTTAFold3 because it can handle protein-ligand complexes. Monomer motif mode uses indexed contigs to preserve functional structural motifs while generating new scaffold segments around them.

On Neurosnap, each mode maps to a dedicated section of the form. Protein Binder mode uses Target Input to crop the uploaded structure and optional Hotspot Residues to define the interface. Ligand Binder mode adds Ligand Residue Name, optional Target Input, and either CONECT bond extraction via Use Bonds From File or a required SMILES string when bonds are regenerated. AME mode requires an atom-level Motif Atom Specification together with ligand residue names; Neurosnap auto-prepares raw uploads when needed and writes ame_prep_report.csv. Monomer Motif mode uses Contig String, motif length bounds, and Atom Selection Mode to control whether all motif atoms or only tip atoms are preserved; Filter Sample Limit does not apply to this mode.

Results should be interpreted as a ranked design campaign rather than a single structure. The service exports results.csv and metrics.csv summaries, generated design_*.pdb structures, optional refolded complex PDBs (design_*_refold_self.pdb, design_*_refold_mpnn.pdb, and design_*_refold_mpnn_fixed.pdb), and for AME jobs ame_prep_report.csv, making it easier to compare binding confidence, motif preservation, and designability before selecting candidates for expression or follow-up optimization.

• Multimodal design coverage: One service supports protein binders, ligand binders, AME motif scaffolding, and monomer motif scaffolding.
• Mode-specific checkpoints: Each workflow uses the model variant best matched to the biological task rather than forcing one generic setting.
• Search-aware generation: Shared advanced controls expose sampling depth, search strategy, and post-generation filtering from a no-code interface.
• Validation-ready outputs: Ranked CSV summaries and structure files make it easier to compare interface quality, motif preservation, and downstream designability.