NickJohr/ss-supervision-rhd-weights

Paper Model Weights

This repository contains the model weights for all model configurations reported in “Secondary-Structure Supervision Improves Protein Foundation Models for Remote Homology Detection”.

The full checkpoints and pretrained baselines are distributed here because they are too large for normal Git/Git LFS storage.

Paper configurations

The paper reports seven configurations for each of the three foundation models (ESM-C 600M, ProtT5, ProstT5). Several configurations share the same weights:

#	Paper configuration	Unique weights	Notes
1	Baseline	baseline/	Unadapted pretrained model
2	PH (Late), without LoRA	ph_no_lora/	Linear-probe checkpoint
3	FH (Late), without LoRA	fh_no_lora/	Head-only fine-tuning checkpoint
4	PH (Early), with LoRA	lora_ph/	LoRA-adapted projection head
5	FH (Early), with LoRA	lora_fh/	LoRA-adapted feature head
6	PH (Late), with LoRA	lora_ph/	Same weights as #4
7	FH (Late), with LoRA	lora_fh/	Same weights as #5

The Early/Late distinction is an extraction choice, not a different checkpoint. Early extraction uses the final backbone representation before the prediction head; Late extraction uses the representation immediately before the final classifier. Therefore, one LoRA adapter serves both the Early and Late rows for its head.

LoRA adapters use rank 4, alpha 64, and dropout 0.1092.

Directory layout

ss-supervision-rhd-weights/
├── README.md
├── full/
│   ├── esmc_600m/
│   │   ├── baseline/
│   │   ├── ph_no_lora/
│   │   ├── fh_no_lora/
│   │   ├── lora_ph/
│   │   └── lora_fh/
│   ├── prott5/
│   └── prostt5/
└── compact/
    ├── esmc_600m/
    │   ├── baseline/          # README with Hugging Face model ID only
    │   ├── ph_no_lora/        # head.safetensors
    │   ├── fh_no_lora/        # head.safetensors
    │   ├── lora_ph/           # adapter_model + adapter_config
    │   └── lora_fh/
    ├── prott5/
    └── prostt5/

Full vs. compact

full/

Contains the complete checkpoint weights needed to reproduce the paper:

• ph_no_lora/ and fh_no_lora/: full model weights (model.safetensors for ESM-C, pytorch_model.bin for the T5 models) including the frozen backbone and the trained head.
• lora_ph/ and lora_fh/: LoRA adapter weights plus adapter_config.json.
• baseline/: full pretrained baseline weights.
• custom_args.json is included where available and records the training arguments of the corresponding run.
• Tokenizer files are included for convenience.

All files in full/ are regular files; no symbolic links are used.

compact/

Contains only the parameters that are not part of the frozen backbone:

• ph_no_lora/head.safetensors and fh_no_lora/head.safetensors: the 18 non-backbone tensors (classifier, dense layers, CNN branches, gating, layer aggregation, relative-position mixer, and attention pooling).
• lora_ph/ and lora_fh/: the LoRA adapter weights and adapter_config.json.
• baseline/: no weights, only the Hugging Face model ID.

This is the version intended for the Git repository. The base model is loaded from Hugging Face and the compact head or adapter is applied on top.

Baselines

Model	Hugging Face model ID	Path in this repository
ESM-C 600M	biohub/esmc-600m-2024-12	full/esmc_600m/baseline/esmc_600m_2024_12_v0.pth
ProtT5	Rostlab/prot_t5_xl_uniref50	full/prott5/baseline/pytorch_model.bin
ProstT5	Rostlab/ProstT5	full/prostt5/baseline/pytorch_model.bin

The ProtT5 baseline is an encoder-only export of the pretrained ProtT5-XL-UniRef50 backbone. The backbone. prefix was removed so that the file is directly loadable with T5EncoderModel.

ESM-C uses the native ESM checkpoint format (.pth), not a Transformers safetensors file. Load it with the ESM library or use the repository’s extraction script.

Loading the weights

Full fine-tuned model (PH/FH without LoRA)

Use the repository’s extraction scripts, pointing --model at the corresponding full/<model>/<config>/ directory. The scripts detect model.safetensors or pytorch_model.bin, read custom_args.json, build the matching classifier, and load the state dict.

python src/evaluation/extract_embeddings_esmc.py \
  --model /path/to/ss-supervision-rhd-weights/full/esmc_600m/ph_no_lora \
  --layer pre_classifier

python src/evaluation/extract_embeddings_t5.py \
  --model /path/to/ss-supervision-rhd-weights/full/prott5/fh_no_lora \
  --layer pre_classifier

Full LoRA model

Point --model at the adapter directory. The script loads the base model named in custom_args.json/adapter_config.json and applies the adapter with PEFT.

python src/evaluation/extract_embeddings_esmc.py \
  --model /path/to/ss-supervision-rhd-weights/full/esmc_600m/lora_ph \
  --layer backbone

Compact head-only weights

Load the pretrained base model, build the same classifier as in src/finetuning/lora_esmc.py or src/finetuning/lora_t5.py, and load the compact state dict with strict=False:

from safetensors.torch import load_file

head_state = load_file("compact/prott5/fh_no_lora/head.safetensors")
missing, unexpected = classifier.load_state_dict(head_state, strict=False)

The 18 tensors are exactly the non-backbone parameters; backbone weights come from the Hugging Face base model.

Compact LoRA adapters

from peft import PeftModel

classifier = ESMCSecondaryStructureClassifier(...)  # or T5 equivalent
model = PeftModel.from_pretrained(classifier, "compact/esmc_600m/lora_ph")

Baseline

Load the model directly from the Hugging Face model ID in the table above, or use the full baseline directory for an offline run.

File formats and sizes

Artifact	Format	Size
ESM-C full PH/FH checkpoint	.safetensors	1.28 GB each
ESM-C full baseline	.pth	2.30 GB
T5 full PH/FH checkpoint	.bin	~4.93 GB each
ProtT5 full baseline (encoder-only)	.bin	4.83 GB
ProstT5 full baseline	.bin	11.28 GB
ESM-C compact head	.safetensors	126.1 MB each
T5 compact head	.safetensors	99.6 MB each
ESM-C LoRA-PH / LoRA-FH	.safetensors	14.7 MB / 136.8 MB
T5 LoRA-PH / LoRA-FH	.bin	24.5 MB / 121.0 MB

Total: ~35 GB for full/ and ~1.1 GB for compact/.

Validation

• Every compact head file contains 18 tensors.
• The compact head tensors have the same names and shapes as the non-backbone tensors in the corresponding full checkpoint.
• The ProtT5 baseline state dict contains 196 keys with only encoder.* and shared.* prefixes; the backbone. prefix has been removed.
• No symbolic links remain in full/; all files are regular files.

Not included

• 6B model weights (the 6B experiments are documented through their metrics in the repository).
• Optimizer states, scheduler states, RNG states, and trainer states.
• Sample-size sweep checkpoints, HPO trials, 300M runs, and ws20/CATH runs.
• Full merged models; LoRA adapters are kept unmerged.

License and citation

The weights retain the licenses of their upstream models:

• ESM-C: Biohub
• ProtT5: Rostlab
• ProstT5: Rostlab

Please cite the paper when using these weights. See the repository CITATION.cff and the paper for the full citation.