LINKER

Knowing that a compound binds a protein is not the same as knowing how it binds. Medicinal chemists reason in terms of named non-covalent interactions — a hydrogen bond to a backbone carbonyl, π-stacking against a tyrosine, a salt bridge to a buried aspartate. Deep learning models that claim interpretability for protein-ligand binding generally fall short of that vocabulary: they either require a resolved 3D complex at inference time, or they explain themselves with distance-based contact maps, which record spatial proximity rather than chemistry.

LINKER, developed by Truong-Son Hy's group at the University of Alabama at Birmingham with a collaborator at the Ho Chi Minh City University of Technology, predicts biologically defined interaction types directly. For every pairing of a protein residue with a ligand functional group, it estimates independent probabilities across seven classes: hydrogen bonds, hydrophobic contacts, π-stacking, π-cation interactions, salt bridges, water bridges, and halogen bonds. Ground-truth labels come from applying the Protein-Ligand Interaction Profiler (PLIP) to experimentally resolved complexes, so 3D structure supervises training but never enters at inference: the deployed model needs only an amino acid sequence and a ligand SMILES string.

The work first appeared as an arXiv preprint in September 2025, was presented at the NeurIPS 2025 AI for Science workshop, and was published in the Journal of Chemical Information and Modeling in 2026. Sequence-only inference separates it from structure-based interpretability models such as DDMuffin, which fuses residue, ligand, and interaction graphs from a supplied complex.

#Key Features

• Interaction types rather than contact maps: The output is a residue × functional-group × 7 probability tensor over PLIP-defined chemistry, not a proximity matrix, so a prediction states which bond type is expected where.
• Functional groups as the ligand unit: A parser built on RDKit and PyCheckmol decomposes each SMILES into chemically meaningful substructures, assigning leftover atoms to the nearest group so coverage is complete.
• Position-aware chemical embeddings: Unlike Morgan/ECFP4 fingerprints, which cannot distinguish two occurrences of the same group, the FINGER-ID module combines learnable group embeddings, positional encodings, and graph-level context.
• Structure-supervised, structure-free at deployment: 3D complexes shape the attention during training only, making the model usable on targets with no experimental structure.
• Leak-proof evaluation: Training and testing use the Leak-Proof PDBBind split, which clusters complexes by binding-site similarity so memorized patterns cannot inflate results.

#Technical Details

The protein branch encodes sequences with ESM Cambrian at the 300M scale, producing residue embeddings of hidden size 960. The ligand branch runs FGParser to build an atom-group matrix, then FINGER-ID to generate multiscale functional-group embeddings via a graph convolutional network. A Self and Cross Attention Transformer (SCAT) applies self-attention within each modality and bidirectional cross-attention between them, and PairwiseUNet — a 2D U-Net adapted to the residue-by-functional-group grid — emits the interaction probability maps. Training uses focal loss (α = 0.85, γ = 1.0) against extreme class imbalance, for 30 epochs with Adam at learning rate 2 × 10⁻⁵ and batch size 2, on a single NVIDIA Tesla P100.

On residue-level interaction detection, LINKER reaches average precision 0.4073 and ROC AUC 0.9369, against 0.2938 and 0.8688 for ArkDTA, the closest sequence-based baseline, with a positive prevalence of only 0.0243. On the finer residue-functional-group task the model attains AUC 0.9753 and up to 174× enrichment over a prevalence baseline of 0.000613. A downstream affinity head — combining pooled embeddings with interaction probabilities under an MSE plus InfoNCE-and-uniformity objective — reaches a test RMSE of 1.47 on Leak-Proof PDBBind, ahead of ArkDTA (1.48), MPRL (1.55), RF-Score (2.10), InteractionGraphNet (2.16), DeepDTA (2.29), and AutoDock Vina (2.56), despite not being trained for affinity regression.

#Applications

The practical draw is annotating binding chemistry at a scale where docking or cofolding is impractical and structures are often missing: triaging virtual screening hits by the interactions they are predicted to form, generating structure-activity hypotheses for analogue series, and nominating pocket residues for mutagenesis. Because outputs are indexed by functional group rather than by atom, the predictions map onto the substructures chemists actually modify. The public repository ships preprocessing, training, and evaluation pipelines for Leak-Proof PDBBind alongside the BindingDB and Davis drug-target affinity benchmarks.

#Impact

LINKER defines a task prior work handled only indirectly — fine-grained, chemically typed interaction prediction from sequence — and shows that structural supervision can be distilled into a model that never sees a structure again. The competitive affinity RMSE, obtained by transfer rather than direct training, indicates that interaction-type supervision yields representations carrying genuine binding information. Limitations are real: evaluation rests on a single benchmark family derived from PDBBind, ArkDTA is the only directly comparable baseline for the interpretability task, and predictions inherit PLIP's geometric cutoffs as ground truth. The source code is public, but no trained weights have been released and the repository carries no license file, so reuse terms are undefined.