Protenix-v2

Overview

Managed version: SubSeq pins Protenix v2.0.0 source, model weights, runtime data, and dependencies. Numerical behavior may change after a future managed release upgrade.

Protein, DNA, RNA, ligand, ion, and mixed-complex prediction.
Typed assembly building for common systems.
Native Protenix JSON for prepared or advanced inputs.
Batch prediction and standardized structure-artifact handoff.
Managed FP32 inference through the upstream compatibility path.

Workflows

Mode	Input shape	When to use it
`build_assembly` Build Assembly Default	No upload required.	Build native input from typed polymer chains and optional ligand or ion entities.
`custom_json` Custom JSON	One selected JSON file.	Run a prepared Protenix input containing advanced entities, modifications, bonds, or local features.
`batch_input` Batch Input	A folder or previous-step source.	Recursively run all valid Protenix JSON inputs from the selected source.

Canonical Job Configuration

The typed fields are returned by GET /api/v1/program/params?program=protenix-v2 and submitted as the params JSON object to POST /api/v1/job/submit.

Setting	What it does
Assembly entities	Add protein, DNA, or RNA chains and optional ligand or ion entities.
Samples	Choose the number of candidate structures generated per job and seed.
Diffusion steps	Control diffusion sampling depth; the managed default is 200.
Pairformer cycles	Control recycling depth; the managed default is 10.
Random seed	Set the model seed used for sampling.
Training-free guidance	Apply optional geometry guidance for suitable protein-ligand systems.
Atom confidence	Write detailed per-atom confidence JSON in addition to summary metrics.

SubSeq manages precision, checkpoint paths, device selection, triangle kernels, and output placement.

Local Feature Policy

Jobs run without network access or hosted preprocessing.
Build Assembly uses single-sequence inputs.
Prepared Custom JSON and Batch Input sources may reference local protein MSA, RNA MSA, and template files under the selected input source.
Enable a local-feature setting only when every applicable entity has its required prepared path.
Paths outside the mounted input source and service-managed reference tree are rejected.

Outputs And Results

/outputs/<job_name>/seed_<seed>/predictions/
  <job_name>_sample_<rank>.cif
  <job_name>_summary_confidence_sample_<rank>.json
  <job_name>_full_data_sample_<rank>.json  # optional

Native CIF and confidence files remain available in the raw output manifest.
Rank 0 is the highest-ranked sample.
Structures are also published as standardized SubSeq artifacts for preview and pipeline handoff.
plddt is local confidence on a 0-100 scale; higher is better.
gpde is predicted global distance error in angstroms; lower is better.
ptm and iptm estimate global and interface confidence; higher is better.
The result reader displays sample ranking, clash, and optional protein-binding fields when present.

Use the raw manifest for full confidence matrices, per-chain values, and detailed atom-confidence sidecars.

Common Examples

Protein monomer: Build Assembly with one protein chain and one sample for a quick check.
Protein-ligand complex: add a SMILES or CCD ligand and optionally enable training-free guidance.
Nucleic-acid complex: add DNA or RNA chains, using the reverse-complement action when useful.
Prepared feature run: use Custom JSON with local feature paths and enable only the corresponding feature settings.

Minimal native Protenix JSON

[
  {
    "name": "tiny",
    "modelSeeds": [101],
    "sequences": [
      {
        "proteinChain": {
          "sequence": "ACDEFGHIK",
          "count": 1
        }
      }
    ]
  }
]

Caveats

Each input job is limited to 2560 polymer residues or bases, five model seeds, and 64 entity instances; batches are limited to 64 jobs.
Single-sequence prediction may be less accurate than prediction with an informative prepared MSA.
Increasing samples, steps, cycles, sequence length, or assembly size increases runtime and memory use.
Confidence and ranking values are computational prioritization signals, not measured affinity, activity, or experimental validation.

Advanced Submit

Advanced submit accepts direct managed arguments through POST /api/v1/job/submit-advanced. Prefer canonical configuration for normal jobs and pipelines.

curl -X POST https://subseq.bio/api/v1/job/submit-advanced \
  -H "Authorization: Bearer <api_key>" \
  -F program=protenix-v2 \
  -F 'args=--input=/inputs/input.json' \
  -F 'args=--samples=5' \
  -F 'args=--steps=200' \
  -F 'args=--cycles=10' \
  -F 'args=--seeds=101' \
  -F 'args=--use-msa=false' \
  -F 'args=--use-templates=false' \
  -F 'args=--use-rna-msa=false' \
  -F 'args=--tfg=false' \
  -F 'args=--atom-confidence=false' \
  -F 'folder=@input.json;filename=input.json'