Proteina-Complexa

What is Proteina-Complexa?

Proteina-Complexa is an atomistic flow-matching generative model for de novo protein complex and binder design from NVIDIA BioNeMo. It jointly generates a binder backbone, side-chain conformations, and sequence while conditioning on a target structure, then applies inference-time search and structural reward models to prioritize candidates.

The ProteinIQ workflow runs the protein-binder configuration only. It designs a new protein binder against selected residues of one uploaded protein target. Small-molecule ligand binders and antigen-motif epitope scaffolding use different Proteina-Complexa models and are not part of this tool.

Binder design is a candidate-generation problem, not a binding assay. Generated structures and model-based ranking help focus downstream work, but expression, stability, specificity, and affinity still need experimental confirmation.

Why it is used

Structure-conditioned binder design is useful when a target surface or epitope has already been identified. Rather than choosing from a fixed library, Proteina-Complexa samples new proteins in a requested length range and scores the resulting complexes during search. The full workflow then refolds and analyzes the retained candidates, making it useful for building a shortlist before sequence review, experimental design, or a separate protein binder design workflow.

How to use Proteina-Complexa online

ProteinIQ runs Proteina-Complexa from an uploaded target PDB. Select the target chain or residue ranges, optionally name hotspot residues, choose a binder length range and sampling settings, then receive generated protein-target complex structures, source reward and evaluation tables, logs, and downloadable run files. The full pipeline evaluates shortlisted designs; generate-and-filter returns candidates sooner without that evaluation step.

Inputs

Input	Requirements and use
Target protein	One PDB file, up to 50 MB, containing protein ATOM records. A structure can also be fetched from RCSB in PDB format.
Target residues	The target surface supplied to the design model. Use a chain such as A, a contiguous range such as A1-150, or comma-separated selections such as A1-100,B1-50. Every residue in an explicit range must be present in the uploaded PDB.
Hotspot residues	Optional residues intended to contact the designed binder, for example A35,A57,A91. Separate identifiers with commas or spaces. A residue identifier consists of a chain letter, residue number, and optional insertion code.

Selecting a whole chain expands to its protein residues. ProteinIQ prepares the selected target from PDB ATOM records before the run, so ligands, ions, glycans, waters, and other HETATM records are not part of the current protein-target design input. Selection should therefore describe the protein epitope, not a co-crystallized ligand or non-protein residue.

Pipeline mode

Mode	What runs	When it fits
Full design pipeline	Generate, filter, evaluate, and analyze.	Final candidate ranking where folding-based evaluation is needed.
Generate and filter only	Generate candidates, score them, and retain the filtered set.	Early exploration of a target or parameter set before spending time on evaluation.

The full design mode is the default. It can take substantially longer because it includes folding-based evaluation of retained designs.

Design and sampling settings

Setting	Description
Minimum binder length / Maximum binder length	Inclusive length range for generated binders, from 20 to 400 residues. The default range is 80 to 120 residues.
Target samples	Number of requested target samples during generation, from 1 to 64. The default is 4.
Diffusion steps	Number of generation steps, from 50 to 1000. The default 400 follows the source protein-binder configuration. More steps increase sampling work.
Search algorithm	Single pass generates without search replication, Best of N compares replicated proposals, and Beam search retains a beam while sampling. Best of N is the default.
Best-of-N replicas	Replicas compared by Best of N, from 1 to 16. The default is 2. This setting matters only with Best of N.
Beam width	Number of candidates retained by beam search, from 1 to 16. The default is 4. This setting matters only with Beam search.
Batch size	Maximum generation batch size, from 1 to 64. The default is 16.
Random seed	Reproducibility seed for the source pipeline. The default is 5; changing it produces a different stochastic sampling trajectory.

Filtering and evaluation settings

Setting	Description
Filter sample limit	Maximum number of generated candidates retained by the filtering stage, from 1 to 5000. The default is 1000.
Reward threshold	Optional numeric cutoff passed to the source filtering stage. Leave it blank to retain the source configuration's standard behavior. Thresholds are meaningful only for comparable runs using the same model and settings.
Redesigned sequences	Number of inverse-folded sequences evaluated per retained design, from 0 to 16. The default is 2. Setting it to 0 skips these additional sequence redesigns.

How Proteina-Complexa works

The model begins with a target structure and samples a binder in all-atom representation. Its flow-based generative prior was pretrained on protein interaction data, including synthetic binder-target pairs, then adapted for binder design. The model does not merely attach a sequence to a fixed scaffold. It proposes the binder geometry and sequence together.

During generation, the selected search strategy uses reward information to rank proposals. In the protein-binder configuration used here, the default reward model is AlphaFold2 Multimer-based and includes an interface predicted-alignment-error term. Best of N searches over replicated proposals, while beam search retains a set of promising partial trajectories. These are in-silico criteria, not measured affinities.

Filtering removes duplicate sequences and keeps the configured number of candidates. Full design then evaluates retained complexes with the source binder evaluation configuration: ColabDesign refolding for binder metrics, SolubleMPNN inverse folding for the requested redesign sequences, ESMFold monomer metrics, designability and co-designability metrics, and secondary-structure proportions. The exact columns written to evaluation tables can vary with the successful stages and source version.

Understanding the results

The Viewer contains the returned PDB structures. Each structure should be inspected as a target-binder complex: check that the intended epitope is contacted, that the binder has a plausible compact fold, and that there are no obvious clashes or disconnected segments. A favourable predicted complex is still a hypothesis, especially when the target structure or selected epitope is uncertain.

Result	Meaning
Generated complex PDB files	Candidate protein-target complex structures from generation, and refolded structures when produced during full evaluation.
rewards_*.csv	Source scoring values recorded during generation and search.
all_rewards_*.csv	Filter-stage records for generated candidates.
top_samples_*.csv	Candidates retained after filtering. This is the most useful starting table for reviewing the generated shortlist.
Evaluation and analysis CSV files	Folding, sequence-design, structural, or aggregate metrics emitted by the full pipeline. Their availability depends on which stages completed.
Config, target dictionary, logs, and additional files	The generated run configuration, selected target description, command log, and any JSON, NPZ, FASTA, text, or other source artifacts written by the job.

total_reward is a composite model score, not a physical binding free energy or an experimentally calibrated affinity. Its absolute scale and direction depend on the active reward terms, so it is best used to compare candidates within the same run and configuration. ProteinIQ retains the source scoring semantics and result order.

Evaluation metrics add independent checks, but none should be treated as a universal pass or fail threshold. Designs that remain promising across the structure view, source ranking, refolding results, and sequence-level assessment are stronger candidates for experimental follow-up than candidates selected on a single column alone.

When to use Proteina-Complexa vs alternatives

Proteina-Complexa is a good fit when all-atom, target-conditioned de novo binder generation and inference-time optimization are the main goals. It is less appropriate when the task starts from a known scaffold with a defined redesign mask, or when the primary need is rapid large-scale backbone generation. In those cases, a ProteinMPNN-compatible design workflow or RFdiffusion can be a better starting point, depending on the available structure and constraints.

For candidate validation, refolding the selected sequences with a separate structure-prediction method can provide a useful cross-check. AlphaFold2 is suitable when an MSA-assisted complex prediction is appropriate; agreement between tools increases confidence, while disagreement is a reason to inspect the design rather than average the scores.