genomenet/twin-point-1024

Twin point-model, Resnik-contrastive, 1024-dim — aspect-specific checkpoints

Fine-tuned two-tower contrastive encoders trained on Resnik GO-semantic similarity, one per GO aspect. Trained on 2025-12-21 with std_ft_bs32ga4 configuration (standard fine-tuning, batch 32, grad-accum 4, dropout 0.25).

• Architecture: custom AA Transformer + ESM2-backbone (facebook/esm2_t33_650M_UR50D), both fine-tuned
• Projection heads: 2-layer MLPs (custom: 256→128→512, ESM: 1280→640→512)
• Output: concat(custom_proj, esm_proj) → 1024-dim
• Distance: L2 on L2-normalized embeddings (point model)

Files

File	GO Aspect	Training run
bp_cp_best.pt	Biological Process	train_point_BP_20251221_std_ft_bs32ga4
cc_cp_best.pt	Cellular Component	train_point_CC_20251221_std_ft_bs32ga4
mf_cp_best.pt	Molecular Function	train_point_MF_20251221_std_ft_bs32ga4

Each file is ~2.7 GB and contains the fine-tuned ESM2 backbone plus the custom transformer and projection heads.

Usage

import torch
from huggingface_hub import hf_hub_download
from twin_model import load_twin_model  # twin_application/scripts/twin_baseline/

ckpt = hf_hub_download("genomenet/twin-point-1024", "bp_cp_best.pt")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model, seq_len, emb_dim = load_twin_model(ckpt, device, "facebook/esm2_t33_650M_UR50D")
# emb_dim == 1024

Used by the genomenet/functional-distance Space with a runtime aspect switcher. Inference code: twin_application/scripts/twin_baseline/twin_model.py.