Trained model parameters for the manuscript TEMA-ENM: An Interpretable Attention-Topology Framework Separates Affinity- and Efficacy-Associated Residue Signals in the Ghrelin Receptor (RSC Digital Discovery, under review). This record contains parameters only: the BiLSTM-BAN backbone pre-trained on BindingDB (49,199 pairs over 2,623 targets, epoch 94), and the ten GHSR fine-tuned checkpoints (seeds 42-51) used for every stability analysis in the paper. README.md gives the per-seed epoch, test AUROC and AUPRC, md5 checksums, the architecture, and the commands to reproduce the attention read-out. Source code, the curated GHSR dataset and the leakage-controlled partitions are not in this record. They are on GitHub at https://github.com/CHIHX12/interpretable-attention-topology-drug-discovery and in the accompanying software record. The separation is deliberate: the GHSR dataset carries a ShareAlike obligation inherited from ChEMBL and cannot be released under a NonCommercial licence, whereas these parameters are our own work and are released for noncommercial use. Licence: CC BY-NC 4.0, with the additional terms in MODEL-WEIGHTS-TERMS.md. Access is granted for academic research, teaching, verification and reproduction of the published results, method development and open benchmarking. Commercial use requires a separate licence.
Source code for CD-BAN, a dual-branch graph neural network with a bilinear cross-attention layer that classifies drug-cyclodextrin inclusion-complex binding affinity (Strong: K > 10,000 M⁻¹ vs. Weak: K < 100 M⁻¹) from 2D SMILES alone, without 3D geometry or handcrafted descriptors. Trained only on the binary extremes, it orders 1,850 withheld intermediate-affinity compounds along a continuous gradient (Kendall τ = -0.289, p < 1e-75); a multi-task regression head converts this into a formula-free affinity estimate. This release contains the complete code, the hyperparameter configuration (configs/CDBAN.yaml), the pinned environment (requirements.txt), and all analysis scripts (no-attention ablation, classical baselines, PCA/UMAP embedding, atom-masking causal validation, host-derivative cold-start, calibration, external-guest similarity). The training data are derived from the OpenCycloDB (Tahil et al. 2023, doi:10.5281/zenodo.7575539) and are referenced, not redistributed, in this code release. Code: https://github.com/CHIHX12/CD-BAN
Source code for CD-BAN, a dual-branch graph neural network with a bilinear cross-attention layer that classifies drug-cyclodextrin inclusion-complex binding affinity (Strong: K > 10,000 M⁻¹ vs. Weak: K < 100 M⁻¹) from 2D SMILES alone, without 3D geometry or handcrafted descriptors. Trained only on the binary extremes, it orders 1,850 withheld intermediate-affinity compounds along a continuous gradient (Kendall τ = -0.289, p < 1e-75); a multi-task regression head converts this into a formula-free affinity estimate. This release contains the complete code, the hyperparameter configuration (configs/CDBAN.yaml), the pinned environment (requirements.txt), and all analysis scripts (no-attention ablation, classical baselines, PCA/UMAP embedding, atom-masking causal validation, host-derivative cold-start, calibration, external-guest similarity). The training data are derived from the OpenCycloDB (Tahil et al. 2023, doi:10.5281/zenodo.7575539) and are referenced, not redistributed, in this code release. Code: https://github.com/CHIHX12/CD-BAN
