appliedCL/interseek-35m

InterSeek 35M

InterSeek predicts whether two proteins interact from their amino-acid sequences. This release contains the complete pair predictor and the fine-tuned ESM-2 backbone from a completed, single-seed experiment.

The model starts from Meta ESM-2 35M. All 12 encoder layers and a CLS interaction head were fine-tuned. This is the validation-selected checkpoint at update 18,000, from a three-epoch run that completed 30,600 updates. It is separate from the ongoing 650M run.

Use the pair predictor

Requires Python 3.10 or later. Download this repository and install the included InterSeek wheel. Use a CUDA-enabled PyTorch build for NVIDIA GPUs.

pip install huggingface_hub
hf download appliedCL/interseek-35m --local-dir interseek-35m
cd interseek-35m
pip install -r requirements-inference.txt ./interseek-0.3.0-py3-none-any.whl
python predict.py --data example_pairs.csv --predictions predictions.csv --device cpu

Use --device cuda:0 for a GPU or --device mps for Apple Silicon. Input CSV columns are sequence_a and sequence_b; pair_id is optional. The example contains illustrative sequences, with no claimed interaction label. Output contains pair_id, label (empty for unlabelled inputs) and score.

Scores lie between zero and one. The threshold 0.48556941747665405 was selected using human validation MCC and fixed before both final tests. Scores are not established as calibrated probabilities of physical binding.

The root weights require the included pair-model implementation; they are not a standard Transformers sequence-classification model. The esm2/ subfolder contains a standard Transformers masked-language-model backbone:

from transformers import EsmForMaskedLM

backbone = EsmForMaskedLM.from_pretrained(
    "appliedCL/interseek-35m", subfolder="esm2"
)

The backbone alone does not produce interaction scores. interaction_head.safetensors contains the separate head; the complete predictor is in model.safetensors. No additional pretrained-weight download is needed for predict.py after the repository has been downloaded.

Training and architecture

Protein pairs are tokenised jointly as [CLS] A [EOS] B [EOS], with at most 1,024 tokens and two sequence-window views per pair. View logits are pooled using log-mean-exp. Inference averages the two protein-order logits before applying sigmoid.

Training combines binary cross-entropy with masked language modelling, weighted by 0.1, and masks 15% of eligible residues. The encoder learning rate was 2e-5; the head learning rate was 1e-4. AdamW used weight decay 0.01, 100 warm-up updates and an effective batch of 16 pairs. The run used seed 17, BF16 and an RTX 4090. Torch was 2.8.0+cu128 and Transformers was 5.12.1. The checkpoint was selected by validation AUROC.

Data and evaluation

The human data preserve the published Bernett et al. gold-standard split, using the authors' corrected Figshare version 4:

Role	Published partition	Pairs	Positive prevalence
Training	HIPPIE-derived Intra1	163,192	50%
Validation	HIPPIE-derived Intra0	59,260	50%
Human final test	HIPPIE-derived Intra2	52,048	50%
Independent yeast final test	D-SCRIPT S. cerevisiae	55,000	9.09%

The yeast data come from D-SCRIPT, independently derived from STRING experimental annotations. Its source repository was pinned to commit 23cbb0fbbb454d09ee41ff749c1454d8b3c8a4b8. Both final tests used the same fixed checkpoint and human-validation threshold.

Final test	AUROC	Average precision	MCC	Precision	Recall
Human	0.6800	0.6700	0.2487	0.6091	0.6882
Yeast	0.7207	0.2269	0.1792	0.1441	0.7674

Full metrics, including the alternative fixed threshold of 0.5, appear in evaluation.json. Validation history, split fingerprints, data audit and training provenance are included. The data themselves and optimiser checkpoints are not included in this inference release.

Intended use and limitations

Use this model for protein-interaction research, candidate ranking and follow-up experiments. A prediction does not establish physical binding or causation. These single-seed results do not establish superiority over other methods.

The published human split controls protein and homology leakage; local checks found no shared identifiers or exact sequences across the four data roles. No new all-against-all homology audit or ESM pretraining-exposure audit was run. Dataset curation, hub-protein biases and sampled negative pairs limit conclusions. The yeast source preserves 32 repeated unordered-pair rows, four label-conflicting pairs and five self-pairs, as recorded in the audit.

Two bounded views may omit relevant regions of long proteins. Transfer across species is imperfect: at the fixed threshold, yeast precision is 14.4%. Additional seeds, matched baselines and experimental validation remain needed.

Files and provenance

• model.safetensors: complete backbone and interaction head, with original tensors unchanged.
• config.json: model settings, threshold, original training provenance and release inference identity.
• esm2/: fine-tuned ESM-2 backbone in Transformers format.
• interaction_head.safetensors: separate classification head.
• predict.py and the InterSeek wheel: matching inference implementation.
• evaluation.json, validation-history.json, study.json, data-audit.json and training-provenance.json: records from the completed experiment.
• SHA256SUMS: file-integrity checks for this release.

The original trained tensor hash is 26531c43768486b217a0995bd721b08eb90d1a368986babdf265bc9f76ef074d. Original experiment metadata retain their historical internal format identifiers. The public inference package was renamed to InterSeek, with prediction equivalence checked against the original package; no weights were retrained or converted.

Source: appliedCL/interseek. This release uses the MIT licence; the upstream ESM-2 model also uses MIT.