CoNCISE

Commercial compound catalogues have outrun the methods meant to search them. Enamine lists billions of make-on-demand molecules, but structural docking needs minutes per pair and even fast sequence-based predictors need milliseconds, turning a single-target screen into weeks of compute. The bottleneck is representational: methods score each ligand independently, so cost scales with catalogue size.

CoNCISE — Compact Novel Codebook Interaction Sequence Embeddings — attacks that by asking how much of a molecule actually matters for binding. The answer the model arrives at is 15 bits. A residual-learning stack of three finite scalar quantisation layers, each with 32 states, assigns every compound a hierarchical code a.b.c, sorting all of chemical space into 32,768 binding categories. Crucially the codebook is supervised on binding data rather than derived from unsupervised clustering of chemical features, so codes group molecules that behave alike against proteins rather than merely looking alike.

The work comes from Rohit Singh's group at Duke University with Lenore Cowen's group at Tufts University and a collaborator at the University of Guelph, extending the same lineage as their earlier ConPLex predictor. Its protein side is built on Raygun's fixed-length embeddings.

#Key Features

• 15-bit ligand codes: Any SMILES string maps in constant time to one of 32,768 hierarchical codes, replacing a 2048-bit Morgan fingerprint with something 0.7% its size.
• Screening decoupled from library size: Once a catalogue is indexed, querying it means scanning codes rather than compounds — 13.8 seconds against all 6.4 billion Enamine compounds, versus a projected 72 days for ConPLex.
• Interpretable hierarchy: Child codes are proper subsets of their parents, and within-group chemical similarity rises down the tree, so a code names a coherent binding class rather than an opaque cluster id.
• Out-of-distribution robustness: On held-out ligands whose codes were never populated during training, accuracy holds near the in-distribution level rather than collapsing.
• Binding-aware protein embeddings: Cross-attention training sharpens the protein side too, separating kinases, phosphatases, GPCRs, and proteases more cleanly than ESM-2 or Raygun embeddings alone.

#Technical Details

A protein sequence becomes a Raygun fixed-length embedding, softmax-pooled across its 50 positions; a ligand SMILES becomes a Morgan fingerprint, quantised through the three-tier codebook; the two are related by cross-attention and scored by cosine similarity in a shared space. Training used AdamW at learning rate 3e-4, batch size 32, for 20 epochs. The authors also assembled MooDengDB, a merged and deduplicated benchmark drawing on PLINDER, BindingDB, BIOSNAP, Davis, and LCIdb — 111,210 ligands, 12,549 proteins, 766K interactions, split 80:10:10 with under 50% protein sequence identity across splits — plus a deliberately large 186,057-ligand out-of-distribution test set.

The paper's tables qualify its abstract's claim of state-of-the-art accuracy. On MooDengDB, CoNCISE reaches AUPR 0.691 and AUROC 0.886, ahead of ConPLex at 0.601 and Komet at 0.591, and far ahead of a k-means-on-fingerprints baseline at 0.435. On the smaller legacy benchmarks it does not lead: AUPR 0.617 on BindingDB and 0.917 on BIOSNAP, slightly below ConPLex and Komet, which the authors attribute to limited data constraining the cross-attention. Out-of-distribution AUPR is 0.74 for ligands in seen codes and 0.67 for entirely unseen codes. Ablations favour the balanced three-tier tree (0.691) over a flat 32,768-way layer (0.641) or a 15-layer binary tree (0.675). Indexing ZINC took 3 hours and Enamine 16 hours on one A100; a whole-proteome sweep of roughly 650 million protein-code predictions finished in 5 hours.

#Applications

The obvious use is virtual screening at a scale previously reserved for well-resourced groups: identify the codes a target binds, sample compounds from those codes, and pass a short list to docking. The paper demonstrates this on KRAS, CXCR4, and p53, three targets long considered difficult to drug, with the pipeline taking under two hours per target and SwissDock reporting binding energies around -7 kcal/mol. Because codes are shared across proteins, the same index supports the inverse query — which other human proteins bind a drug's code — giving a first-pass off-target list, and it supports compact perturbation panels covering many binding classes with few compounds.

#Impact

CoNCISE reframes chemical space as a small discrete vocabulary learned for a task, rather than a continuous high-dimensional embedding — echoing the way structural tokenisation reshaped protein modelling in ESM-3 and SaProt. Its proteome-wide sweep also produced a testable claim about druggability: 1,552 human proteins matched no code confidently, but that count shrinks when unexplored regions of chemical space are included, suggesting some "undruggable" targets are unexplored rather than intractable. The honest limits are that all validation is computational — no compound was synthesised or assayed — and that docking follow-up used a rigid-receptor method. Code, the MooDengDB splits, and pretrained checkpoints are public — installable from PyPI, with weights pulled from Zenodo — under a dual arrangement: CC BY-NC-SA 4.0 for non-commercial academic use, with commercial terms negotiated separately through Duke's technology transfer office.