Protein-ligand binding affinity model that tokenizes quantum electron-cloud density into discrete codes, plus a distilled cloud-free variant.
No providers recorded yet. Browse all providers
ECBind is a pretraining-and-finetuning framework for protein-ligand interaction prediction that pushes representation learning below the atomic scale. Nearly all deep-learning binding predictors — equivariant graph networks, geometric transformers, hierarchical message-passing models — treat atoms or molecular fragments as the finest unit of description. But the forces that actually determine binding, including electrostatics, van der Waals contacts, hydrogen bonding, and π-π stacking, arise from the distribution of electrons, not from the positions of nuclei. A ball-and-stick graph cannot express charge polarization or lone pairs, so models built on one must infer those effects indirectly from atom types and geometry.
ECBind instead takes an explicit electron cloud as input. Density is computed for the binding site with the semi-empirical GFN2-xTB method, then compressed by a structure-aware transformer and hierarchical vector-quantized codebooks into discrete tokens — one per atom rather than one per grid point. Those tokens are learned once during self-supervised pretraining and reused for downstream supervised tasks: relative binding affinity, absolute binding affinity, and ligand specificity classification. The work comes from Haitao Lin, Odin Zhang, Stan Z. Li and colleagues at the Westlake University AI Lab, Zhejiang University, and DP Technology, and was posted to arXiv in May 2025 as a preprint that has not yet been peer reviewed.
Because computing electron densities is expensive, ECBind also includes a knowledge-distilled student that reproduces the teacher's tokens from full-atom structure alone, making the representation usable where quantum-chemical preprocessing is unaffordable.
Pretraining combines three objectives: reconstruction of the full electron cloud from the compressed per-atom tokens, masked atom-type prediction at a 10% masking ratio, and a vector-quantization commitment loss with exponential-moving-average codebook updates. Repeated k-means re-initialization and spherical projection guard against codebook collapse. A parallel full-atom tokenizer captures longer-range structure through binned distance classification and an inter-molecular interaction loss; downstream heads attend over all four codebooks. Pockets are defined as residues within 6 Å of the ligand following the Atom3D protocol, hydrogens are placed with Reduce and MMFF, and complexes that fail to converge during density calculation are dropped.
Evaluation spans three benchmarks. On MISATO relative binding affinity (9,964/1,074/156 train/validation/test pairs), ECBind reaches a per-structure Pearson correlation of 0.431 and Spearman of 0.445, improvements of 6.42% and 15.58% over the strongest baseline, GET, alongside an overall Pearson of 0.692, RMSE of 3.892, and AUROC of 0.843. The distilled student scores 0.407 and 0.420 on the same per-structure metrics. On LBA absolute affinity with a 30% sequence-identity split, ECBind reports Pearson 0.626 and RMSE 1.298; on the LEP specificity task it reaches AUROC 0.775. Gains are largest on MISATO, where molecular-dynamics refinement yields cleaner geometry and therefore more reliable densities — on the noisier LBA set the full-atom tokenizer alone matches the complete model.
The model targets structure-based drug design, where ranking congeneric ligands against a single target matters more than ranking across targets. Per-structure metrics — correlation computed within one protein across many candidate molecules — are what a lead optimization campaign cares about, and they are where ECBind's advantage is largest. The distilled student is the practical entry point for virtual screening: it accepts ordinary prepared complexes and needs no semi-empirical calculation per compound, while the specificity head supports active/inactive triage in hit finding.
ECBind extends the biomolecular tokenization trend — Foldseek's structural alphabet, ESM-3's structure tokens — down to the electronic scale, showing that the resulting codes carry signal atom-level geometry alone does not. The distillation result matters more than the headline numbers: quantum-chemical supervision can be baked into a model during training and then discarded at inference, a pattern applicable well beyond affinity prediction. The limitations are concrete. Density quality bounds the benefit, so the approach helps most on curated, MD-refined datasets and least on raw crystallographic ones. No code or pretrained weights have been released, blocking independent reproduction and downstream reuse, and the paper remains an unreviewed preprint.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.