Research map / Probabilistic and scientific modeling / Causal discovery

CausalBind: Causal Modeling and Learning for Protein-Molecule Virtual Screening

NeurIPS 2026 · Probabilistic and scientific modeling › Causal discovery · approach: Constraint-based structure learning with latents

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes CausalBind, a retrieval-based virtual screening model whose contrastive scores pass through a learnable sparse (or low-rank) concept-pair interaction mask derived from a V-structure causal model under Heckman selection, improving DUD-E and LIT-PCBA early enrichment.

Problem

Retrieval-based protein-molecule virtual screening uses dense holistic alignment in a shared embedding space, which entangles binding determinants with nuisance correlations and limits generalization to new targets and scaffolds, especially when training data contain only observed binding pairs.

Approach

Model binding as a selected collider in a V-structure causal model and constrain cross-modal interactions to sparse antichain or low-rank concept-pair masks over Perceiver-extracted concepts, instead of dense alignment; three identifiability theorems justify the constraints.

Key results

• On LIT-PCBA, CausalBind-SP reaches EF1% 8.56 vs 6.50 for LigUnity and 4.62 for HypSeek, with AUROC 0.639 (Table 1).
• On DUD-E, CausalBind-SP reports AUROC 0.935 and EF1% 47.69, while CausalBind-ATOM reaches the best DUD-E AUROC 0.943 and EF1% 49.66.
• Against S$^2$Drug (paper-reported), S$^2$Drug keeps the best DUD-E BEDROC (0.793 vs 0.744 for CausalBind-SP), while CausalBind variants are higher on DUD-E AUROC and EF1%.

Surprising findings

• Removing the sparsity penalty from CausalBind-LR or adding it has limited effect, so the gain is attributed to the low-rank parameterization rather than sparsity regularization.
• Hard-zero sparsity (lambda=1e-3 with 20.5% zeros) gives no gain in BEDROC and lower LIT-PCBA performance, indicating a soft interaction bottleneck works better than hard masking.

Limitations

• Evidence centered on retrieval-based virtual screening under cross-dataset generalization; no uniform gains shown on sequence-family or time-stratified splits.
• Interpretability analysis gives model-level attribution rather than proof of a physical causal mechanism.
• S$^2$Drug results are paper-reported rather than reproduced, and it beats CausalBind on DUD-E BEDROC.

Experimental setup

Datasets: ChEMBL BindingDB PDBBind DUD-E LIT-PCBA DEKOIS 2.0 Merck-FEP-8

Benchmarks: DUD-E LIT-PCBA DEKOIS 2.0 Merck-FEP-8

Metrics: AUROC BEDROC_80.5 EF_1%

Baselines: Glide-SP DeepDTA Gnina RTMScore TankBind BigBind DrugCLIP GenScore EquiScore DrugHash LigUnity HypSeek

Builds on: HypSeek LigUnity DrugCLIP Perceiver Uni-Mol

Closest papers by meaning

• AANet: Virtual Screening under Structural Uncertainty via Alignment and Aggregation (NeurIPS 2025 · similarity 0.86)
• PPDiff: Diffusing in Hybrid Sequence-Structure Space for Protein-Protein Complex Design (ICML 2025 · similarity 0.86)
• Generalist Equivariant Transformer Towards 3D Molecular Interaction Learning (ICML 2024 · similarity 0.86)
• Knowledge Enhanced Representation Learning for Drug Discovery (AAAI 2024 · similarity 0.86)
• CrossBind: Collaborative Cross-Modal Identification of Protein Nucleic-Acid-Binding Residues (AAAI 2024 · similarity 0.86)
• Latent Retrieval Augmented Generation of Cross-Domain Protein Binders (NeurIPS 2025 · similarity 0.85)