Deep Binding: Neural Modeling of Interactions

• Deep binding is a research term describing the use of deep neural architectures to learn binding relationships—from protein-ligand affinity to feature conjunctions—directly from raw or minimally processed data.
• It leverages diverse data modalities such as sequences, graphs, and 3D structures, enhancing prediction accuracy while integrating biochemical and genomic domain knowledge.
• Recent advances extend deep binding from predictive affinity estimation to generative ligand design, addressing interpretability, structural dynamics, and multi-domain challenges.

“Deep binding” is a polysemous research term that most commonly denotes the use of deep neural architectures to model binding relations directly from raw or minimally processed inputs. In molecular and chemical biology, it usually refers to prediction or generation of protein–ligand or drug–target binding, binding affinity, or binding sites from sequence, graph, docking, or 3D structural data; in genomics, it refers to predicting transcription factor binding from DNA sequence; and in representation learning, it denotes the capacity of a model to encode which features belong to the same object or role. Across these settings, the shared idea is that a deep model learns a binding-relevant mapping—such as ligand structure to docking-derived affinity, protein–ligand complex geometry to affinity, DNA sequence to TFBS probability, or internal representation to object-level feature conjunctions—without relying exclusively on hand-crafted rules (Aman, 9 Jan 2025).

1. Conceptual scope of the term

In the molecular setting, “deep binding” is explicitly used for deep neural prediction of binding affinity or related binding quantities. One formulation is a ligand-based surrogate of docking: Morgan fingerprints derived from SMILES are mapped to AutoDock Vina docking scores for ErbB inhibitors, where the regression target is the docking-derived binding affinity in kcal/mol and more negative values denote stronger predicted binding (Aman, 9 Jan 2025). A broader drug–target formulation treats binding affinity prediction as regression from protein and ligand representations to $pK_d$, $KIBA$, or related continuous labels, as in DeepDTA, WideDTA, DeepGS, ResDTA, and KEPLA (Öztürk et al., 2018).

The term also appears in sequence biology. TFBS-Finder frames transcription factor binding site prediction as a binary classification task on 101 bp DNA sequences, outputting $P(\text{TFBS present} \mid s)$, and is described as a “deep binding” model because it maps raw sequence to a binding or non-binding decision using DNABERT together with CNN and attention modules (Ghosh et al., 3 Feb 2025). This suggests that, outside small-molecule discovery, the term extends naturally to any deep model that predicts molecular binding events from sequence.

A distinct but related use appears in vision and structured representation learning. “Formalizing the Binding Problem” defines binding information as mutual information between a representation $Z$ and an object code $O$, with conditional binding information $I(O;Z\mid F)$ quantifying what the representation knows about which features go together beyond merely knowing which features are present (Huang et al., 2 Jun 2026). In symbolic memory architectures, “Recursive Binding on a Budget” studies role–filler binding in tensor memories and proposes Orthogonal Subspace Carving as a way to support deep recursive binding while decoupling tensor order from structural depth (Pence et al., 9 Jun 2026). In this usage, deep binding concerns representational structure rather than biochemical interaction.

2. Molecular deep binding as affinity prediction

A large body of work uses “deep binding” for continuous prediction of protein–ligand or drug–target affinity. The earliest sequence-only models emphasized direct learning from 1D representations. DeepDTA uses CNNs on SMILES and amino acid sequences and trains with MSE to predict continuous affinity values on Davis and KIBA, achieving its best reported KIBA performance with CI $= 0.863$ and MSE $= 0.194$ (Öztürk et al., 2018). WideDTA retains the sequence-only setting but moves from character-level encodings to word-level representations: protein sequence words, protein domains and motifs, ligand SMILES words, and ligand maximum common substructure words are processed by CNN blocks and fused in dense layers, with the best KIBA configuration reported as CI $= 0.875$ and MSE $= 0.179$ (Öztürk et al., 2019).

Subsequent models enriched the representation side. DeepGS combines a drug graph branch based on GAT, a SMILES branch using Smi2Vec with BiGRU, and a protein sequence branch using Prot2Vec with CNN, and reports on Davis a CI of $KIBA$0, MSE of $KIBA$1, $KIBA$2, and AUPR $KIBA$3 (Lin, 2020). ResDTA stays in the sequence-only regime but adds residual skip connections and a third “combined” stream over joint drug–protein features; on KIBA it reports CI $KIBA$4 for the full model, compared with $KIBA$5 without residual connections (Ghosh et al., 2023). These studies collectively support a recurring pattern: deep binding models often improve by learning richer latent representations of ligands and targets before affinity regression.

Ligand-only surrogates of docking represent a narrower but practically important variant. For ErbB inhibitors, Morgan fingerprints with radius $KIBA$6 and $KIBA$7 bits are used as features, while the target is the AutoDock Vina binding affinity on an ErbB4 homology model. The model is a fully connected DNN with hidden layers Dense(128), Dense(64), Dense(64), trained for 50 epochs with Adam and MSE on scaled labels; it reports training MSE $KIBA$8, MAE $KIBA$9, $P(\text{TFBS present} \mid s)$0, and test $P(\text{TFBS present} \mid s)$1 (Aman, 9 Jan 2025). The paper describes this as a classic ligand-based deep binding model because it predicts how tightly a molecule will bind to a specific protein from chemical structure alone.

More recent interaction-free systems incorporate external biochemical priors. KEPLA combines an ESM2-based protein encoder, a GCN-based ligand encoder, a knowledge graph built from Gene Ontology and ligand properties, and a cross-attention interaction module. Its total loss is

$P(\text{TFBS present} \mid s)$2

where $P(\text{TFBS present} \mid s)$3 is MAE on affinity labels and $P(\text{TFBS present} \mid s)$4 aligns global protein and ligand representations with GO and ligand-property relations using RotatE and TransE scoring functions. On PDBbind core, KEPLA reports RMSE $P(\text{TFBS present} \mid s)$5, MAE $P(\text{TFBS present} \mid s)$6, and $P(\text{TFBS present} \mid s)$7, and on CSAR-HiQ RMSE $P(\text{TFBS present} \mid s)$8, MAE $P(\text{TFBS present} \mid s)$9, and $Z$0 (Liu et al., 16 Jun 2025). This suggests a second major trend in deep binding research: coupling learned structural representations with domain knowledge to improve both generalization and interpretability.

3. Structure-based deep binding: complexes, dynamics, and geometric learning

A separate line of work defines deep binding through explicit 3D structure. Pafnucy represents each protein–ligand complex as a $Z$1 voxel grid with 19 channels per voxel and applies a 3D CNN with convolutional layers of 64, 128, and 256 filters followed by dense layers of 1000, 500, and 200 units. Trained on PDBbind v2016, it reports on the 2016 core set RMSE $Z$2, MAE $Z$3, SD $Z$4, and $Z$5, outperforming classical scoring functions on the tested benchmarks (Stepniewska-Dziubinska et al., 2017). DeepSurf addresses a related but distinct task—binding site prediction rather than affinity prediction—by placing $Z$6 voxel grids on solvent-accessible surface points and applying 3D CNNs. On COACH420, its ResNet-18 variant reports DCA Top-$Z$7 success of $Z$8, and on HOLO4K $Z$9, exceeding earlier deep learning competitors on those datasets (Mylonas et al., 2020).

Graph-based rescoring of docked complexes is another structure-based interpretation. A dual-graph GCNN for binding mode prediction uses one graph for ligand bonded topology and a second for ligand–protein contacts. On PDBbind pose classification, the interaction-sensitive LP branch reaches AUC $O$0, combined L+LP reaches AUC $O$1, and inclusion of docking pose rank in the consensus L+LP+R model raises AUC to $O$2 and top-ranked pose correctness to $O$3, versus $O$4 for Vina (Morrone et al., 2019). The same study argues that dataset bias dominates virtual screening results on DUD-E, whereas binding mode prediction provides a cleaner setting in which deep models genuinely learn structural interaction patterns.

Dynamics-aware variants move beyond a single static structure. Dynaformer is pretrained on an MD dataset of 3,218 protein–ligand complexes, each simulated for 10 ns with 100 snapshots, and then finetuned on static PDBbind data. It encodes atom-level graphs with geometry-aware attention biases based on Gaussian basis functions of distances and angles and augments them with interaction fingerprints. On CASF-2016, it reports Pearson $O$5 and RMSE $O$6, while its HSP90 case study identifies 12 hit compounds out of 20 tested, including two submicromolar hits, after screening a 50,000-compound library (Min et al., 2022). The paper’s central claim is that affinity is an ensemble property and that MD-sampled thermodynamic information improves deep binding.

Geometric deep learning has also been used to improve cross-target generalization. IPBind models binding affinity as the energy difference between bound and unbound states: $O$7 using a shared SE(3)-invariant graph encoder with frame averaging and atom-level contributions (Li et al., 22 Apr 2025). On Atom3D LBA60 it reports RMSE $O$8 and Pearson $O$9, and on the harder LBA30 RMSE $I(O;Z\mid F)$0 and Pearson $I(O;Z\mid F)$1, outperforming several strong baselines in the unseen-protein regime (Li et al., 22 Apr 2025). HCLBind extends structure-based deep binding to multi-domain proteins via self-supervised hierarchical contrastive learning, domain-gated graph attention, LoRA-adapted foundation models, and evidential regression; on PDBBind it reports RMSE $I(O;Z\mid F)$2, PCC $I(O;Z\mid F)$3, and C-index $I(O;Z\mid F)$4 (Zhang et al., 19 May 2026).

4. Generative deep binding and docking surrogates

Deep binding can also be generative rather than predictive. A conditional variational autoencoder trained on CrossDocked2020 learns $I(O;Z\mid F)$5 from 3D atomic density grids of complexes and reconstructs ligand densities conditioned on receptor densities (Ragoza et al., 2021). The model uses an input encoder on the full complex, a conditional encoder on the receptor, and a decoder that outputs ligand density, trained with reconstruction, KL, and steric-clash penalties. Discrete molecules are then obtained by atom fitting and bond inference. Reported validity is $I(O;Z\mid F)$6 for posterior sampling and $I(O;Z\mid F)$7 for prior sampling, uniqueness is $I(O;Z\mid F)$8 and $I(O;Z\mid F)$9, respectively, and both settings yield $= 0.863$0 novelty (Ragoza et al., 2021). Mutation-conditioning experiments on shikimate kinase show that generated ligands change in chemically intuitive ways when pocket residues are altered, which the paper interprets as evidence of receptor-specific deep binding.

At the other end of the abstraction spectrum, docking-derived labels can be distilled into fast ligand-only models. The ErbB docking surrogate described above illustrates this strategy clearly: docking provides a target-specific binding signal, and the DNN approximates that signal from fingerprints alone (Aman, 9 Jan 2025). A plausible implication is that deep binding models can serve as learned scoring surrogates within multi-stage pipelines, where inexpensive learned models pre-filter vast libraries before more expensive docking, MD, or free-energy methods.

5. Sequence-based binding in regulatory genomics

In genomics, deep binding usually denotes prediction of binding events from sequence. TFBS-Finder treats transcription factor binding site identification as binary classification on 101 bp windows derived from ENCODE ChIP-seq. Its architecture combines pre-trained DNABERT, a CNN module, a modified CBAM with spatial attention followed by channel attention, a multi-scale convolutions with attention module, and an output classifier trained with binary cross-entropy (Ghosh et al., 3 Feb 2025).

Across 165 ENCODE datasets, the full TFBS-Finder reports mean Accuracy $= 0.863$1, PR-AUC $= 0.863$2, and ROC-AUC $= 0.863$3, compared with $= 0.863$4, $= 0.863$5, and $= 0.863$6 for BERT-TFBS and substantially lower values for DeepBind, DanQ, CRPTS, DLBSS, D-SSCA, and DSAC (Ghosh et al., 3 Feb 2025). In cross-cell-line evaluation for CTCF, all train–test combinations yield ROC-AUC $= 0.863$7, including Train Gm12878 $= 0.863$8 Test Hepg2 $= 0.863$9 and Train K562 $= 0.194$0 Test Helas3 $= 0.194$1 (Ghosh et al., 3 Feb 2025). This literature uses the same phrase as molecular affinity studies, but here “binding” refers to TF–DNA occupancy rather than ligand–receptor affinity.

6. Binding as a representational property

Outside biochemistry, deep binding denotes a model’s ability to preserve feature conjunctions or recursive structure. “Formalizing the Binding Problem” defines binding information as

$= 0.194$2

where $= 0.194$3 is an object code and $= 0.194$4 a model representation, and conditional binding information as

$= 0.194$5

which measures what the representation knows about object identity given the feature code $= 0.194$6 (Huang et al., 2 Jun 2026). On a synthetic ColorShape $= 0.194$7 dataset, DINOv2-Large spatial tokens with an attention probe achieve $= 0.194$8 bits and $= 0.194$9 bits, corresponding to normalized binding measures of about $= 0.875$0 and $= 0.875$1, while the [CLS] token retains far less (Huang et al., 2 Jun 2026). The paper argues that binding is a key ingredient of strong visual recognition and reasoning.

“Recursive Binding on a Budget” studies deep recursive binding in tensor memories. It proposes Orthogonal Subspace Carving, where a context $= 0.875$2 defines a subspace with orthogonal-complement projector

$= 0.875$3

and a filler $= 0.875$4 is bound by projecting each component into that complement and forming a fixed order-$= 0.875$5 tensor (Pence et al., 9 Jun 2026). The paper shows that interference for a query matching a stored binding has standard deviation

$= 0.875$6

and reports, for example, about $= 0.875$7 retrieval accuracy with $= 0.875$8, $= 0.875$9, and $= 0.179$0 bindings (Pence et al., 9 Jun 2026). Here deep binding refers to role–filler composition under superposition, not to molecular recognition, but the shared concern is still how deep models represent and retrieve relational structure.

7. Interpretability, limitations, and research directions

Interpretability in deep binding varies by domain but often emerges from architecture-specific decompositions. In protein–ligand affinity models, IPBind exposes atom-level contributions $= 0.179$1, enabling visualization of interaction hotspots (Li et al., 22 Apr 2025). KEPLA provides both cross-attention maps over protein fragments and ligand atoms and KG-based explanations via GO and ligand-property relations (Liu et al., 16 Jun 2025). TFBS-Finder’s ablations support the idea that global transformer context and local motif-sensitive attention are complementary (Ghosh et al., 3 Feb 2025). In the visual binding literature, probing results indicate that most object-binding information resides in spatial tokens rather than global summary tokens (Huang et al., 2 Jun 2026).

Common limitations are also recurrent. Sequence- and graph-based molecular models lack explicit 3D interaction geometry (Öztürk et al., 2018). Docking-surrogate approaches inherit docking bias because the supervision target is a docking score rather than experimental $= 0.179$2 or $= 0.179$3 (Aman, 9 Jan 2025). Structure-based models depend on input pose or structure quality, even when they are more robust than earlier baselines (Li et al., 22 Apr 2025). Dynamics-aware models improve realism but at significant data-generation cost because MD is expensive (Min et al., 2022). Generative 3D models still rely on post hoc atom fitting and bond inference and do not directly optimize final chemical validity or binding energetics (Ragoza et al., 2021). Multi-domain protein systems remain difficult because inter-domain geometry and flexible linkers introduce aleatoric noise and challenge rigid, monolithic graph assumptions (Zhang et al., 19 May 2026).

The recent literature points toward several converging directions. One is tighter integration of foundation models with task-specific geometric or attention modules, as in TFBS-Finder, KEPLA, and HCLBind (Ghosh et al., 3 Feb 2025). Another is stronger physical grounding, either through bound–unbound energy differences, SE(3)-aware architectures, MD-informed pretraining, or evidential uncertainty (Li et al., 22 Apr 2025). A third is movement from prediction to generation: conditional models that design ligands directly from receptor binding sites extend deep binding from estimating compatibility to sampling it (Ragoza et al., 2021). Across domains, a plausible synthesis is that “deep binding” increasingly denotes not a single model family but a broader methodological program: learning binding relations, binding strengths, or binding structure directly from data-rich representations while progressively incorporating geometry, knowledge, uncertainty, and compositional structure.