DeepGGL: Graph Learning for Drug Affinity

• DeepGGL is a deep learning model that predicts drug–target binding affinity using multiscale geometric graph representations derived from 3D protein–ligand complexes.
• It transforms spatial interactions into image-like tensors and employs residual CNN blocks with Luong-style dot-product attention to capture detailed binding patterns.
• Benchmarked on datasets like CASF and PDBbind, DeepGGL demonstrates state-of-the-art performance and robust generalization across diverse structural complexes.

DeepGGL is a deep geometric graph learning model for regression of drug–target binding affinity from 3D structures of protein–ligand complexes. In structure-based drug design, it represents a protein–ligand complex as multiscale weighted colored bipartite subgraphs, aggregates centrality-based statistics into an image-like tensor, and processes that tensor with a 2D convolutional neural network containing residual blocks and a Luong-style dot-product self-attention layer to predict binding affinity in $\text{p}K$ units. The model was introduced for benchmarking on CASF-2013, CASF-2016, CSAR-NRC-HiQ, and a PDBbind v2019 hold-out set, where it was reported to achieve state-of-the-art performance and strong generalization (Rana et al., 15 Sep 2025).

1. Problem domain and intended role

DeepGGL is designed to predict drug–target binding affinity from 3D structures of protein–ligand complexes. In this setting, a ligand is a small molecule and a target is typically a protein; affinity is expressed as $K_d$, $K_i$, or $IC_{50}$ and standardized to $\text{p}K$, the negative logarithm of the affinity constant. The task is central to structure-based drug design because it supports virtual screening, lead optimization, and high-throughput filtering that can reduce reliance on experimental assays such as ITC and SPR (Rana et al., 15 Sep 2025).

The model is positioned against two classes of prior methods. Traditional scoring functions include physics-based, empirical, and knowledge-based approaches. Physics-based methods can be accurate but are computationally expensive and sensitive to force-field parameters and structural noise. Empirical and knowledge-based methods depend on hand-crafted terms or statistical potentials and may miss complex non-linear interactions. Earlier machine-learning and deep-learning methods improved performance, but sequence-based models such as DeepDTA, DeepDTAF, and sequence-mode CAPLA do not explicitly encode 3D geometry, while many structure-based methods such as RF-Score, Pafnucy, K$_\text{DEEP}$, OnionNet, OnionNet-2, EGNA, and BAPA rely on voxel grids or coarse distance-based descriptors, use relatively limited feature sets or single-scale representations, or show limitations in generalization and sensitivity to structural noise. DeepGGL is introduced as a response to those limitations through multiscale geometric graphs combined with residual CNN blocks and attention (Rana et al., 15 Sep 2025).

A common misconception is to treat DeepGGL as a generic graph neural network over molecular structures. In the reported formulation, it is more specific: it is a structure-based affinity regressor whose intermediate representation is not an arbitrary message-passing graph embedding, but a fixed multiscale tensor assembled from weighted colored bipartite subgraph statistics. This design choice makes the model simultaneously graph-based and tensor-based in its downstream computation (Rana et al., 15 Sep 2025).

2. Geometric graph representation

Each protein–ligand complex is modeled as a geometric graph $G=(V,E)$ whose vertices are atoms with 3D coordinates and atom-type labels. Protein atoms and ligand atoms are treated as distinct colored vertex sets,

$$V_p = \left\{ (\mathbf{r}_i, t^{(p)}_k) \mid \mathbf{r}_i \in \mathbb{R}^3,\ t^{(p)}_k \in \mathcal{T}_p \right\}, \qquad V_l = \left\{ (\mathbf{r}_j, t^{(l)}_{k'}) \mid \mathbf{r}_j \in \mathbb{R}^3,\ t^{(l)}_{k'} \in \mathcal{T}_l \right\}.$$

Protein atom types are residue-level labels such as CA, CB, and CD1, whereas ligand atom types follow the SYBYL scheme, including labels such as C.3, C.ar, N.ar, and N.am (Rana et al., 15 Sep 2025).

Interactions are encoded as weighted bipartite edges between protein and ligand atoms whose Euclidean distance is below a cutoff $c$. The characteristic interaction distance is

$\eta_{kk'} = \eta (r_k + r_{k'}),$

where $K_d$0 and $K_d$1 are van der Waals radii and $K_d$2 is a scaling factor. Edge weights are defined by radial basis functions that satisfy $K_d$3 as $K_d$4 and $K_d$5 as $K_d$6. Two forms are used:

$K_d$7

and

$K_d$8

These weights are described as rotation/translation-invariant interaction strengths that decay with distance and capture non-covalent contacts (Rana et al., 15 Sep 2025).

For each protein atom type $K_d$9 and ligand atom type $K_i$0, DeepGGL computes a centrality score

$K_i$1

which functions as a density-like measure of how strongly a given protein atom interacts with ligand atoms of a specified type. For the subgraph associated with an atom-type pair, the resulting centrality values are summarized by eight statistical descriptors: count, sum, mean, median, variance, standard deviation, minimum, and maximum. With $K_i$2 protein atom types, $K_i$3 ligand atom types, and $K_i$4 statistics, a single-scale representation contains $K_i$5 features (Rana et al., 15 Sep 2025).

DeepGGL extends this representation across three distance cutoffs: 5 Å, 10 Å, and 15 Å. Each cutoff produces a tensor of shape $K_i$6, and the three tensors are reorganized into a final input tensor of shape $K_i$7. The three channels correspond analogously to RGB channels, but they encode short-range, medium-range, and long-range interaction structure. The 5 Å channel is described as encoding local contacts such as hydrogen bonds and hydrophobic packing; the 10 Å channel encodes medium-range structural context and secondary interactions; and the 15 Å channel captures longer-range structural proximity and environment (Rana et al., 15 Sep 2025).

3. Network architecture and learning objective

Before entering the network, the $K_i$8 tensor is standardized with StandardScaler from scikit-learn to zero mean and unit variance. The convolutional backbone begins with an initial 2D convolutional layer containing 32 filters and a $K_i$9 kernel, where the kernel size was selected by cross-validation. It is followed by two residual blocks. Residual Block 1 contains two convolutional layers with 64 filters each, batch normalization after each convolution, ReLU activations, and a skip connection. Residual Block 2 follows the same pattern with 128 filters in both convolutional layers. The residual formulation is written as

$IC_{50}$0

with the stated benefits of improved gradient flow, mitigation of vanishing gradients, deeper feature extraction, and feature reuse (Rana et al., 15 Sep 2025).

After the convolutional stages, the final feature map is flattened into a 2D matrix and passed through a Luong-style dot-product attention layer implemented with the TensorFlow Keras Attention layer:

$IC_{50}$1

The stated role of this component is to dynamically reweight spatial patterns, focus the model on the most informative interactions, and capture long-range dependencies and non-local patterns in the learned feature space. Post-attention processing uses two dense layers: first 128 neurons, then 64 neurons, both with batch normalization, dropout rate $IC_{50}$2, and L2 regularization with penalty $IC_{50}$3. The output layer is a single linear neuron producing $IC_{50}$4 (Rana et al., 15 Sep 2025).

The target value is defined by converting experimental affinities to

$IC_{50}$5

Training uses a composite loss that combines Pearson correlation coefficient and RMSE:

$IC_{50}$6

with $IC_{50}$7 selected by cross-validation. This weighting emphasizes correlation while retaining numerical error minimization. Optimization uses stochastic gradient descent with Nesterov momentum $IC_{50}$8, learning rate $IC_{50}$9, decay $\text{p}K$0, gradient clipping with clipvalue = 0.01, batch size $\text{p}K$1, and up to $\text{p}K$2 epochs. Early stopping halts training if validation loss does not improve by at least $\text{p}K$3 over $\text{p}K$4 consecutive epochs. Hyperparameter tuning tests kernel sizes from 3 to 9 and $\text{p}K$5 values from 0.1 to 0.9 via 5-fold cross-validation using RepeatedKFold. Implementation uses the Keras API with a TensorFlow backend (Rana et al., 15 Sep 2025).

This architecture is not a generic end-to-end atom message-passing network. Instead, raw 3D structures are first transformed into multiscale centrality-statistics tensors, then processed by a residual CNN with attention. A plausible implication is that DeepGGL occupies an intermediate design space between handcrafted geometric descriptors and fully learned graph propagation (Rana et al., 15 Sep 2025).

4. Datasets, evaluation protocol, and reported results

Training and validation are based on PDBbind v2016. The reported split removes all complexes appearing in CASF-2016 and CASF-2013 to avoid leakage, samples 1,000 complexes from the refined set as validation data, and uses the remaining complexes from the general set for training. CASF-2016 consists of 285 high-quality complexes; CASF-2013 contains 195 complexes across 65 clusters and is described as structurally diverse with affinities spanning approximately 10 orders of magnitude. External generalization tests use CSAR-NRC-HiQ, reduced to 49 unique complexes after merging subsets and removing overlaps with PDBbind v2016 training data, and a PDBbind v2019 hold-out set of 222 complexes from refined v2019 deposited after the v2016 release (Rana et al., 15 Sep 2025).

Evaluation uses Pearson correlation coefficient $\text{p}K$6, RMSE, and MAE, all in $\text{p}K$7 units. The model is trained independently five times and predictions are averaged across runs as an ensemble to reduce variance and improve robustness (Rana et al., 15 Sep 2025).

Dataset	$\text{p}K$8	RMSE / MAE
Training set	0.996	0.212 / 0.165
Validation set	0.810	1.158 / 0.882
CASF-2016	0.868	1.150 / 0.900
CASF-2013	0.844	1.288 / 1.071
CSAR-NRC-HiQ	0.764	1.512 / 1.211
PDBbind v2019 hold-out	0.557	1.267 / 1.001

On CASF-2016, DeepGGL is reported with the highest $\text{p}K$9 and the lowest RMSE among the listed methods: $_\text{DEEP}$0, RMSE $_\text{DEEP}$1, and MAE $_\text{DEEP}$2. The comparison table includes PerSpect, AGL-Score, RF-Score v3, PLEC-Linear, OnionNet-2, EGNA, K$_\text{DEEP}$3, BAPA, OnionNet, Pafnucy, CAPLA, and DeepDTAF. On CASF-2013, DeepGGL is again reported as best among the listed methods with $_\text{DEEP}$4, RMSE $_\text{DEEP}$5, and MAE $_\text{DEEP}$6 (Rana et al., 15 Sep 2025).

The external benchmarks are intended to test robustness and temporal or structural disjointness. On CSAR-NRC-HiQ, DeepGGL achieves the highest correlation and the lowest RMSE among the listed methods, with MAE described as competitive because it is slightly higher than EGNA’s $_\text{DEEP}$7 despite a better overall metric combination. On the PDBbind v2019 hold-out set, DeepGGL leads the listed methods DeepFusion, Pafnucy, and K$_\text{DEEP}$8 on all three reported metrics (Rana et al., 15 Sep 2025).

5. Ablation findings, interpretability, strengths, and limitations

Ablation experiments separate the contributions of multiscale interaction modeling, residual connections, and attention. Single-range variants are defined as DeepGGL$_\text{DEEP}$9 for 5 Å only, DeepGGL$G=(V,E)$0 for 10 Å only, and DeepGGL$G=(V,E)$1 for 15 Å only. On CASF-2016, DeepGGL$G=(V,E)$2 reports $G=(V,E)$3, RMSE $G=(V,E)$4, MAE $G=(V,E)$5; DeepGGL$G=(V,E)$6 reports $G=(V,E)$7, RMSE $G=(V,E)$8, MAE $G=(V,E)$9; DeepGGL$$V_p = \left\{ (\mathbf{r}_i, t^{(p)}_k) \mid \mathbf{r}_i \in \mathbb{R}^3,\ t^{(p)}_k \in \mathcal{T}_p \right\}, \qquad V_l = \left\{ (\mathbf{r}_j, t^{(l)}_{k'}) \mid \mathbf{r}_j \in \mathbb{R}^3,\ t^{(l)}_{k'} \in \mathcal{T}_l \right\}.$$0 reports $$V_p = \left\{ (\mathbf{r}_i, t^{(p)}_k) \mid \mathbf{r}_i \in \mathbb{R}^3,\ t^{(p)}_k \in \mathcal{T}_p \right\}, \qquad V_l = \left\{ (\mathbf{r}_j, t^{(l)}_{k'}) \mid \mathbf{r}_j \in \mathbb{R}^3,\ t^{(l)}_{k'} \in \mathcal{T}_l \right\}.$$1, RMSE $$V_p = \left\{ (\mathbf{r}_i, t^{(p)}_k) \mid \mathbf{r}_i \in \mathbb{R}^3,\ t^{(p)}_k \in \mathcal{T}_p \right\}, \qquad V_l = \left\{ (\mathbf{r}_j, t^{(l)}_{k'}) \mid \mathbf{r}_j \in \mathbb{R}^3,\ t^{(l)}_{k'} \in \mathcal{T}_l \right\}.$$2; and the full multiscale model reports $$V_p = \left\{ (\mathbf{r}_i, t^{(p)}_k) \mid \mathbf{r}_i \in \mathbb{R}^3,\ t^{(p)}_k \in \mathcal{T}_p \right\}, \qquad V_l = \left\{ (\mathbf{r}_j, t^{(l)}_{k'}) \mid \mathbf{r}_j \in \mathbb{R}^3,\ t^{(l)}_{k'} \in \mathcal{T}_l \right\}.$$3, RMSE $$V_p = \left\{ (\mathbf{r}_i, t^{(p)}_k) \mid \mathbf{r}_i \in \mathbb{R}^3,\ t^{(p)}_k \in \mathcal{T}_p \right\}, \qquad V_l = \left\{ (\mathbf{r}_j, t^{(l)}_{k'}) \mid \mathbf{r}_j \in \mathbb{R}^3,\ t^{(l)}_{k'} \in \mathcal{T}_l \right\}.$$4, MAE $$V_p = \left\{ (\mathbf{r}_i, t^{(p)}_k) \mid \mathbf{r}_i \in \mathbb{R}^3,\ t^{(p)}_k \in \mathcal{T}_p \right\}, \qquad V_l = \left\{ (\mathbf{r}_j, t^{(l)}_{k'}) \mid \mathbf{r}_j \in \mathbb{R}^3,\ t^{(l)}_{k'} \in \mathcal{T}_l \right\}.$$5. On CASF-2013, the same pattern holds: medium-range is individually most informative, short-range is close, long-range is weakest, and combining all three ranges yields the best performance. The paper summarizes this as evidence that multi-scale interactions are complementary (Rana et al., 15 Sep 2025).

Architectural ablations compare the full model to variants without residual blocks, without attention, and without both. The full model reports $$V_p = \left\{ (\mathbf{r}_i, t^{(p)}_k) \mid \mathbf{r}_i \in \mathbb{R}^3,\ t^{(p)}_k \in \mathcal{T}_p \right\}, \qquad V_l = \left\{ (\mathbf{r}_j, t^{(l)}_{k'}) \mid \mathbf{r}_j \in \mathbb{R}^3,\ t^{(l)}_{k'} \in \mathcal{T}_l \right\}.$$6 values of $$V_p = \left\{ (\mathbf{r}_i, t^{(p)}_k) \mid \mathbf{r}_i \in \mathbb{R}^3,\ t^{(p)}_k \in \mathcal{T}_p \right\}, \qquad V_l = \left\{ (\mathbf{r}_j, t^{(l)}_{k'}) \mid \mathbf{r}_j \in \mathbb{R}^3,\ t^{(l)}_{k'} \in \mathcal{T}_l \right\}.$$7 on CASF-2016, $$V_p = \left\{ (\mathbf{r}_i, t^{(p)}_k) \mid \mathbf{r}_i \in \mathbb{R}^3,\ t^{(p)}_k \in \mathcal{T}_p \right\}, \qquad V_l = \left\{ (\mathbf{r}_j, t^{(l)}_{k'}) \mid \mathbf{r}_j \in \mathbb{R}^3,\ t^{(l)}_{k'} \in \mathcal{T}_l \right\}.$$8 on CASF-2013, $$V_p = \left\{ (\mathbf{r}_i, t^{(p)}_k) \mid \mathbf{r}_i \in \mathbb{R}^3,\ t^{(p)}_k \in \mathcal{T}_p \right\}, \qquad V_l = \left\{ (\mathbf{r}_j, t^{(l)}_{k'}) \mid \mathbf{r}_j \in \mathbb{R}^3,\ t^{(l)}_{k'} \in \mathcal{T}_l \right\}.$$9 on CSAR-HiQ, and $c$0 on the v2019 hold-out. Removing the residual block lowers performance on CASF-2016, CASF-2013, and v2019 hold-out, although CSAR-HiQ shows a slight increase to $c$1. Removing attention yields a slight decrease on CASF-2016 and v2019 hold-out but slight increases on CASF-2013 and CSAR-HiQ. Removing both residual block and attention produces broader degradation, including a drop from $c$2 to $c$3 on the v2019 hold-out. The paper interprets the residual and attention components as synergistic overall, while also describing attention as dataset-dependent in utility (Rana et al., 15 Sep 2025).

The structured subgraph descriptors themselves are identified as interpretable because their centrality statistics can be linked to specific atom-type pairs and distance ranges. At the same time, the paper states that it does not provide detailed interpretability visualizations such as attention maps. This suggests that interpretability in DeepGGL derives primarily from the descriptor construction rather than from a post hoc analysis of latent features (Rana et al., 15 Sep 2025).

The reported strengths fall into three groups. Conceptually, the model uses a geometric graph-based representation that is bipartite, multiscale, colored, and weighted; it captures fine-grained atom-type interactions with physically meaningful distance weighting and enjoys rotation/translation invariance through radial basis functions. Architecturally, it combines a residual CNN, attention, and a composite Pearson-plus-RMSE loss. Empirically, it reports state-of-the-art performance on CASF-2013 and CASF-2016 and superior generalization on CSAR-NRC-HiQ and the PDBbind v2019 hold-out (Rana et al., 15 Sep 2025).

The reported limitations are equally specific. DeepGGL depends on high-quality 3D structures and may degrade when complexes are poorly resolved or docked rather than crystallographic. It uses static structures and does not explicitly model protein flexibility, induced fit, or dynamics. It is trained heavily on PDBbind v2016 and therefore inherits dataset coverage limitations. Although atom-type aggregation constrains dimensionality, preprocessing still requires substantial computation for centrality and tensor construction. Suggested future directions include extension to other biomolecular interaction tasks such as protein–protein interactions, incorporation of richer physical priors and possibly topological features such as persistent homology, stronger interpretability via visualization of subgraph contributions or attention weights, and multi-task learning for affinity together with properties such as selectivity and kinetics (Rana et al., 15 Sep 2025).

6. Terminological scope and related usages

As a formal model name, DeepGGL refers to the affinity-prediction architecture introduced in “A Geometric Graph-Based Deep Learning Model for Drug-Target Affinity Prediction” (Rana et al., 15 Sep 2025). However, the surrounding literature supplied with the term shows that the label is not globally unique. The infrared small target detection paper “Gradient-Guided Learning Network for Infrared Small Target Detection” introduces GGL-Net rather than DeepGGL, yet its implementation-oriented exposition explicitly frames the method as “DeepGGL” in the sense of gradient-guided learning (Zhao et al., 10 Dec 2025). This suggests that “DeepGGL” can also be used informally as shorthand for gradient-guided learning, even though that usage designates a different problem domain, architecture, loss, and benchmark suite.

Other neighboring usages are even broader. “Constructing an Interpretable Deep Denoiser by Unrolling Graph Laplacian Regularizer” presents a graph-based deep denoiser and explicitly describes it as a “DeepGGL”-type model in the sense of graph-Laplacian-guided unrolling, but not as a named method called DeepGGL (Hosseini et al., 2024). “Deep Graphs” proposes recurrent vertex updates for graph learning (Platanios et al., 2018), “Deep Graph Library” refers to DGL as a graph-centric framework for GNN implementation (Wang et al., 2019), and “Deep Feature Learning for Graphs” introduces DeepGL as a hierarchical relational feature-learning framework (Rossi et al., 2017). These are related by theme—deep learning on graphs, graph-based representation, or graph-guided modeling—but they are distinct from the drug–target affinity model named DeepGGL.

A practical implication is that references to “DeepGGL” require domain-specific disambiguation. In structure-based drug discovery, the term denotes the multiscale weighted colored bipartite subgraph model for binding affinity prediction (Rana et al., 15 Sep 2025). In other contexts, similar letter sequences may refer to gradient-guided infrared detection, graph Laplacian denoisers, graph-learning frameworks, or graph feature-learning methods, none of which are interchangeable with the affinity-prediction model (Zhao et al., 10 Dec 2025).