When running inference with a protein sequence of length 635 aa, ESMFeaturizer throws an AssertionError during featurization

[DPmooncake]

Logs:
2025-09-13 14:46:34,425 [protenix\data\infer_data_pipeline.py:246] INFO protenix.data.infer_data_pipeline: Featurizing model_005_temp_0.05_id1...
2025-09-13 14:46:44,979 [protenix\data\esm_featurizer.py:119] WARNING protenix.data.esm_featurizer: [model_005_temp_0.05_id1] ESM error: :
Traceback (most recent call last):
File "protenix\data\esm_featurizer.py", line 103, in call
x_esm = self.load_esm_embedding(sequence)
File "protenix\data\esm_featurizer.py", line 54, in load_esm_embedding
assert x.size(0) == len(sequence)
AssertionError

Json:
[
{
"name": "model_005_temp_0.05_id1",
"sequences": [
{
"proteinChain": {
"sequence": "MAPVTIEKFVNQEERHLVSNRSATIPFGEYIFKRLLSIDTKSVFGVPGDFNLSLLEYLYSPSVESAGLRWVGTCNELNAAYAADGYSRYSNKIGCLITTYGVGELSALNGIAGSFAENVKVLHIVGVAKSIDSRSSNFSDRNLHHLVPQLHDSNFKGPNHKVYHDMVKDRVACSVAYLEDIETACDQVDNVIRDIYKYSKPGYIFVPADFADMSVTCDNLVNVPRISQQDCIVYPSENQLSDIINKITSWIYSSKTPAILGDVLTDRYGVSNFLNKLICKTGIWNFSTISGKSVIDESNPTYMGQYNGKEGLKQVYEHFELCDLVLHFGVDINEINNGHYTFTYKPNAKIIQFHPNYIRLVDTRQGNEQMFKGINFAPILKELYKRIDVSKLSLQYDSNVTQYTNETMRLEDPTNGQSSIITQVHLQKTMPKFLNPGDVVVCEVGSFQFSVRDFAFPSQLKYISQGFFQSIGMALPAALGVGIAMQDHSNAHINGGNVKEDYKPRLILFEGDGAAQMTIQELSTILKCNIPLEVIIWNNNGYTIERARMGPTRSYNDVMSWKWTKLFEAFGDFDGKYTNSTLIQCPSKLALKLEELKNSNKRSGIELLEVKLGELDFPEQLKCMVEAEALKRNKK",
"count": 1
}
},
{
"ligand": {
"ligand": "O=C(Cc1ccccc1)C(=O)O",
"count": 1
}
}
],
"covalent_bonds": []
}
]

[DPmooncake]

Environment

OS: Windows 10

Python: 3.10

PyTorch: 2.3.1

Protenix: v0.5.3

[zhangyuxuann]

@DPmooncake Since it was detected that the ESM embeddings had already been computed, the calculation was skipped to avoid redundant computation—especially useful when running inference with multiple seeds. See: https://github.com/bytedance/Protenix/blob/a1d39d30fa957c807d46528f6b5f951d878d2bc6/protenix/data/compute_esm.py#L123 . Currently, ESM saves precomputed embeddings based on the "name" field in the JSON file. The error occurs because embeddings were previously computed for a certain "name", and then a different JSON file with the same "name" was used, causing ESM to fail loading the existing embeddings. You can resolve this by deleting the local esm_embeddings directory and retrying.

[DPmooncake]

@zhangyuxuann thanks!