protenix-v2 has no end-to-end pipeline

Summary

protenix-v2 is listed as a supported model, but there is no end-to-end pipeline for it: build_processor(model_source="protenix-v2") cannot fold from sequence input the way Boltz-2 / OpenFold3 / AlphaFold2 can. Only the optimized model forward is wired; the surrounding stages are stubs.

What I found

Using the same EngineProcessorConfig → build_processor path that folds Boltz-2 end-to-end, Protenix-v2 does not run: ProtenixFactory's tokenizer / feature_factory / postprocessor raise NotImplementedError (the model/trunk+diffusion is the only piece implemented). By contrast Boltz-2, OpenFold3 and AF2 have all stages wired and fold from a sequence request out of the box.

If you instead try to drive the optimized Protenix model directly, the forward expects an input_feature_dict on a different schema than a naive OSS ByteDance Protenix feature dump produces — e.g. it reads keys like d_lm / v_lm and drops profile / deletion_mean. So even the forward-only path needs an (undocumented) feature-schema conversion, which makes it hard to reproduce the reported Protenix speedups end-to-end.

Reproduce

from bionemo_ir.registry import register_all_factories; register_all_factories()
from bionemo_ir.pipeline.processor.engine_proc import EngineProcessorConfig, build_processor

cfg = EngineProcessorConfig(
    model_source="protenix-v2",
    runtime_args={"diffusion_samples": 1, "num_sampling_steps": 200, "recycling_steps": 10},
    writer_stage={"output_path": "/out", "format": "cif"},
)
proc = build_processor(cfg)   # protenix-v2: tokenizer/feature_factory/postprocessor NotImplementedError

(The identical pattern with model_source="boltz-2" folds fine.)

Ask

1. Is an end-to-end Protenix-v2 pipeline planned (tokenizer + featurizer + postprocessor wired into the stage framework, like Boltz-2)?
2. If not, would a PR porting the OSS ByteDance Protenix featurization into the stage framework be welcome? Happy to contribute if the direction is wanted.
3. Separately, could the expected input_feature_dict schema for the optimized Protenix forward (the d_lm/v_lm keys) be documented, so the forward can be exercised against an OSS feature dump in the meantime?

Environment

bionemo-ir 0.1.0, Python 3.12, CUDA 13.2, driver 580.95.05, NVIDIA L40S (Modal). Boltz-2 end-to-end works in the same environment.

Commencez par le chemin reproduit EngineProcessorConfig → build_processor et inspectez ProtenixFactory en parallèle du pipeline Boltz-2 fonctionnel. Confirmez le périmètre avec les maintainers avant de modifier tokenizer, feature_factory ou postprocessor, puisque l’issue demande si cette direction est souhaitée. Le travail serait considéré comme terminé lorsqu’une sequence request s’exécute de bout en bout ; séparément, le schéma d’entrée attendu de d_lm/v_lm serait documenté si ce travail est accepté.