CausalBind: Causal Modeling and Learning for Protein-Molecule Virtual Screening (NeurIPS 2026 Oral)

Trained checkpoints of CausalBind.

• Paper: arXiv:2610.07340
• Code, training and evaluation: https://github.com/lokali/CausalBind

CausalBind is a dual-tower retrieval model for zero-shot protein-molecule virtual screening. A Perceiver-style concept extractor decomposes the pocket structure, molecule structure and protein sequence into latent concepts, and a learnable cross-modal mask routes binding evidence through a small number of concept pairs, following identifiability results for a V-structure causal model under binder-only selection.

Files

Each file holds the model weights (fp16) of the last checkpoint of the seed-1 run.

File	Variant	Interaction structure	DUD-E AUROC / BEDROC / EF@1%	LIT-PCBA AUROC / BEDROC / EF@1%
causalbind_sp.pt	CausalBind-SP	learnable sparse concept-pair mask	0.935 / 0.744 / 47.69	0.639 / 0.097 / 8.56
causalbind_lr.pt	CausalBind-LR	low-rank concept-pair mask	0.935 / 0.704 / 44.87	0.632 / 0.103 / 8.84
causalbind_emb.pt	CausalBind-EMB	mask in the pooled embedding space	0.938 / 0.744 / 48.00	0.623 / 0.089 / 7.61
causalbind_atom.pt	CausalBind-ATOM	SP with per-atom concept attention	0.943 / 0.768 / 49.66	0.621 / 0.079 / 6.25

SHA-256 checksums are in SHA256SUMS.

Usage

The checkpoints are loaded by the code in the GitHub repository (Uni-Core plugin); they are not transformers models.

git clone https://github.com/lokali/CausalBind.git && cd CausalBind
# set up the environment and data as described in the GitHub README, then
hf download LokaLi/CausalBind --local-dir checkpoints
bash scripts/test.sh sp DUDE checkpoints/causalbind_sp.pt results/sp

or, in Python:

from huggingface_hub import hf_hub_download
path = hf_hub_download("LokaLi/CausalBind", "causalbind_sp.pt")

Training data and evaluation

Trained on the LigUnity / HypSeek assay-level corpus (ChEMBL, BindingDB), with every assay whose target appears in DUD-E, LIT-PCBA or DEKOIS 2.0 removed. Evaluated zero-shot on DUD-E (102 targets) and LIT-PCBA (15 targets). See DATA.md for sources and protocol.

License

The weights are released under CC BY-NC 4.0, because the training data (LigUnity resources) are CC BY-NC 4.0 and part of the model code derives from MERU (CC BY-NC 4.0). The rest of the code is MIT; see the GitHub repository's LICENSE and NOTICE.

Citation

@inproceedings{li2026causalbind,
  title     = {CausalBind: Causal Modeling and Learning for Protein-Molecule Virtual Screening},
  author    = {Li, Loka and Tian, Jin and Zhang, Kun},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026},
  url       = {https://arxiv.org/abs/2610.07340}
}