AlphaDTA: integrating AlphaFold3 embeddings and 3D complex structures for drug–target binding affinity prediction

Abstract

Accurate prediction of drug–target binding affinity plays an important role in structure-based drug discovery, yet existing approaches are constrained by their reliance on scarce experimentally determined protein–ligand complex structures. In this context, recent advances in biomolecular structure prediction models, such as AlphaFold3, have emerged as a promising approach to alleviate the scarcity of experimentally determined protein–ligand complex structures. In this study, we introduce AlphaDTA, a framework that integrates AlphaFold3-predicted structures and embeddings for drug–target binding affinity prediction. AlphaDTA processes two types of AlphaFold3 embeddings, single and pair embeddings. We utilized the single embeddings at both fine-grained and coarse-grained levels to capture local interaction patterns and global binding context, and used the pair embeddings to encode cross-molecular relational features. The AlphaFold3-predicted structures are further processed by a three-dimensional structure-based geometric encoder to produce corresponding structural embedding representations. The resulting single, pair, and structural embeddings are then integrated through adaptive fusion for affinity prediction. When we evaluated AlphaDTA using recently proposed PDBbind data splits designed to reduce train–test structural overlap, AlphaDTA achieved state-of-the-art or competitive performance across multiple independent benchmarks. In a case study on the cystic fibrosis transmembrane conductance regulator, AlphaDTA correctly identifies a clinically approved potentiator and suggests a potential repurposing candidate among FDA-approved drugs.

Scientific contribution

AlphaDTA enables accurate drug–target binding affinity prediction without relying on experimentally determined protein–ligand complex structures. By integrating AlphaFold3-derived single, pair, and structure-based embeddings through adaptive fusion, AlphaDTA improves generalization on structurally nonredundant benchmarks and demonstrates utility for target-specific drug repurposing.

Graphical Abstract

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s13321-026-01259-6.

Introduction

Identifying small molecules (ligands) that bind effectively to disease-relevant target proteins is a central challenge in drug discovery, making a thorough understanding of protein–ligand interactions (PLIs) essential for therapeutic development [1, 2]. However, the vast chemical space of potential ligands makes it impractical to characterize all candidate protein–ligand pairs using experimental methods, given the technical complexity and substantial resources involved [2–4]. Computational approaches therefore offer a scalable alternative for predicting PLIs across large compound libraries [5, 6]. Among these tasks, drug–target affinity (DTA) prediction plays a crucial role in computational drug discovery, as it quantifies binding strength to systematically rank and prioritize candidate ligands for a given target protein [7, 8].

Existing methods for binding affinity prediction can be broadly divided into free-energy–based approaches and scoring-function–based approaches [9, 10]. Free-energy–based approaches estimate binding affinity by computing the binding free energy ($\mathrm{\Delta }G$) using explicit physical models together with conformational sampling, but remain computationally expensive and impractical for large-scale screening [9–11]. Scoring-function–based methods approximate affinity using predefined interaction terms and are commonly formulated as force-field–based, empirical, or knowledge-based scoring functions [9, 10, 12], yet their fixed functional forms constrain their capacity to capture complex or atypical interaction patterns [10]. More recently, deep learning-based scoring models have gained considerable attention for their ability to learn data-driven representations that capture complex binding patterns from data [5, 9, 13, 14]. These deep learning approaches can be further categorized into sequence-based and complex-based methods depending on their use of three-dimensional (3D) structural information [15, 16].

Sequence-based methods [17–19] take proteins as amino acid sequences and ligands as SMILES (Simplified Molecular Input Line Entry System) [20] strings as input, enabling scalable learning from large datasets without requiring 3D structural information [15, 17–19]. However, because binding affinity is governed by spatial interactions within the binding site, the lack of explicit 3D structural context can limit predictive accuracy [5, 14, 21–23]. In contrast, complex-based methods [16, 24–26] leverage 3D protein–ligand complex structures to directly model binding geometry. However, their applicability is constrained by the limited availability of high-quality 3D structural data [16, 27, 28].

Recently, significant advances in 3D biomolecular complex structure prediction models such as NeuralPLexer [29], RoseTTAFold All-Atom [30], UMol [31], and AlphaFold3 [32] have alleviated these limitations. Among these, AlphaFold3 has achieved unprecedented performance in predicting biological complex structures including protein–ligand, protein–nucleic acid, and other molecular interactions, establishing a new standard in this domain [33, 34]. These advances now enable complex-based binding affinity prediction methods to access high-quality complex structures without relying on experimental determination.

In parallel, large-scale foundation models pretrained with self-supervised objectives such as ChemBERTa-2 [35], MolFormer [36], ProtTrans [37], and ESM2 [38] have demonstrated strong transfer across biochemical domains. Consequently, recent approaches for protein–ligand binding prediction, including both drug–target interaction and binding affinity prediction tasks, increasingly leverage contextual embeddings extracted from pretrained encoders [39–43]. These approaches primarily employ protein language models trained on protein sequences and molecular language models trained on SMILES strings to extract rich biochemical representations from sequence-level inputs, thereby enhancing generalization under limited labeled data. Recent studies have further extended this paradigm to complex-based methods by additionally incorporating pretrained sequence-level protein and ligand representations, demonstrating improved binding affinity prediction accuracy [44–46]. However, in these approaches, the protein and ligand encoders are pretrained independently on their respective modalities, without exposure to joint protein–ligand interaction context.

In this regard, beyond its role as a structure predictor, AlphaFold3 also generates two types of embeddings during the structure prediction process: single embeddings, which capture per-atom and per-residue features, and pair embeddings, which encode pairwise relational information. In the protein–ligand setting considered here, these representations are produced while modeling the protein and ligand together. Thus, AlphaFold3 embeddings differ from representations obtained from independently pretrained protein language models and molecular language models. Specifically, AlphaFold3 embeddings can represent proteins and ligands in a shared interaction-aware structural context, providing rich structural and interaction information learned from large-scale molecular data. However, despite their quality and potential relevance, such embeddings have not been effectively leveraged for binding affinity prediction.

In this work, we propose AlphaDTA, a framework for drug–target affinity prediction that integrates AlphaFold3-predicted complex structures and embeddings. AlphaDTA consists of two complementary encoders: a Multi-Granularity Encoder and a Geometric Encoder. The Multi-Granularity Encoder processes two types of AlphaFold3 embeddings—single and pair embeddings. Specifically, single embeddings are utilized at both fine-grained and coarse-grained levels to capture local interaction patterns and global binding context, while pair embeddings encode cross-molecular relational features between atoms and residues. The Geometric Encoder operates directly on the predicted 3D complex structure to extract structure-aware interaction features that explicitly model spatial binding geometry. For affinity prediction, the representations learned by the two encoders are adaptively fused to form a unified feature representation. AlphaDTA achieves competitive or superior performance across multiple benchmark datasets compared to existing methods. In addition, a case study on the cystic fibrosis transmembrane conductance regulator highlights its practical applicability, where the model suggests a clinically approved potentiator and a potential repurposing candidate among FDA-approved drugs.

Methods

Overview

AlphaDTA is a deep learning framework for drug–target binding affinity prediction that leverages predicted complex structures and two types of embeddings generated by AlphaFold3 from protein sequences and ligand SMILES strings (Fig. 1). From the full AlphaFold3-generated embeddings, pocket-level representations are constructed by retaining protein residues within 10 Å of any ligand heavy atom together with all ligand tokens, forming the input to the subsequent encoders. AlphaDTA is composed of two main encoders: a Multi-Granularity Encoder and a Geometric Encoder. The Multi-Granularity Encoder derives fine-grained and coarse-grained feature vectors from AlphaFold3 single and pair embeddings, capturing interaction patterns at multiple levels of resolution. In parallel, the Geometric Encoder extracts structure-aware geometric feature vectors from the predicted 3D protein–ligand complex. The resulting representations are then integrated through a Triple Adaptive Fusion module to produce the final binding affinity prediction. Detailed descriptions of AlphaFold3 inference and data preprocessing are provided in Additional file 1: Sections S1 and S2, respectively.

Fig. 1.

Overview of the AlphaDTA framework. Given a protein sequence and ligand SMILES, AlphaFold3 (AF3) generates a predicted complex structure and embeddings (single and pair). Pocket-level embeddings are extracted based on residues within 10 Å of the ligand. The Multi-Granularity Encoder processes these embeddings through Fine-grained, Coarse-grained, and Pair Extractors, while the Geometric Encoder extracts structural features via InteractionGraphNet (IGN). The resulting representations are integrated through Triple Adaptive Fusion to predict binding affinity (pK)

Model architecture

Multi-Granularity Encoder

Multi-Granularity Encoder (MGE) extracts features from AlphaFold3 embeddings at multiple levels of granularity to capture both detailed molecular interactions and global binding context. The AlphaFold3 embeddings consist of a single embedding $S\in {\mathbb{R}}^{N\times 384}$ and a pair embedding $Z\in {\mathbb{R}}^{N\times N\times 128}$, where $N=N_p+N_l$, with $N_p$ protein residues and $N_l$ ligand heavy atoms. The single embedding encodes individual characteristics of each token, including local structural and chemical properties, while the pair embedding captures relational information between token pairs, representing potential interactions and spatial dependencies across the complex. To process these two embeddings, MGE consists of four components: a Residual Token Projection (RTP) that adapts the single embedding, a Single Extractor that encodes it at fine-grained and coarse-grained levels, a Pair Extractor that extracts interaction-specific features from the pair embedding, and a Fusion layer that combines these representations.

Residual Token Projection

To adapt the AlphaFold3 single embedding to our architecture while preserving its pretrained knowledge, we use Residual Token Projection (RTP):

$$\begin{array}{c}\hfill S_{\text{proj}}=\text{RTP}(S)\end{array}$$ 1

$$\begin{array}{c}\hfill \text{RTP}(x)=x+\text{MLP}(\text{LN}(x)).\end{array}$$ 2

The residual formulation allows the model to refine the pretrained embeddings while maintaining the original representation as a stable reference. We adopt a pre-normalization design, applying layer normalization $\text{LN}(·)$ prior to the MLP to improve optimization stability.

Single Extractor

To capture binding information at different granularities, the projected single embedding $S_{\text{proj}}$ is processed through two complementary extractors: a Fine-grained Single Extractor for independent protein–ligand encoding with bidirectional cross-attention, and a Coarse-grained Single Extractor for joint protein–ligand encoding and global feature extraction.

The Fine-grained Single Extractor first separates the AlphaFold3 single embeddings into protein and ligand tokens according to their positional indices, and processes them independently to obtain protein- and ligand-specific representations. It then handles detailed protein–ligand dependencies by capturing their mutual interactions through bidirectional cross-attention.

We denote the protein and ligand embeddings extracted from $S_{\text{proj}}$ as $P^{(0)}$ and $L^{(0)}$, respectively. These embeddings are then encoded using stacked Transformer [47] blocks:

$$\begin{array}{c}\hfill \begin{array}{cc}\hfill {\tilde{X}}^{(l)}& =X^{(l)}+\text{SelfAttn}(\text{LN}(X^{(l)})),\hfill \\ \hfill X^{(l+1)}& ={\tilde{X}}^{(l)}+\text{FFN}(\text{LN}({\tilde{X}}^{(l)})).\hfill \end{array}\end{array}$$ 3

where $X\in \{P,L\}$. After separate encoding, bidirectional cross-attention exchanges information between protein and ligand representations:

$$\begin{array}{c}\hfill P_{\text{cross}},L_{\text{cross}}=\text{CrossAttn}(P,L),\end{array}$$ 4

where bidirectional cross-attention is implemented with pre-normalization:

$$\begin{array}{c}\hfill \begin{array}{cc}\hfill P_{\text{cross}}& =\text{Attn}(Q=\text{LN}(P),K,V=\text{LN}(L)),\hfill \\ \hfill L_{\text{cross}}& =\text{Attn}(Q=\text{LN}(L),K,V=\text{LN}(P)).\hfill \end{array}\end{array}$$ 5

The refined features are then obtained via residual connections:

$$\begin{array}{c}\hfill \begin{array}{cc}\hfill P_{\text{fine}}& =P+P_{\text{cross}},\hfill \\ \hfill L_{\text{fine}}& =L+L_{\text{cross}}.\hfill \end{array}\end{array}$$ 6

Coarse-grained Single Extractor encodes the full token sequence using the same Transformer block architecture (Eq. 3) to produce context-dependent token representations:

$$\begin{array}{c}\hfill S_{\text{coarse}}=\text{TransformerBlock}(S_{\text{proj}}).\end{array}$$ 7

This extractor provides a global summary that complements the fine-grained modeling in the Fine-grained Single Extractor.

Gated multi-scale pooling

Inspired by InteractionGraphNet (IGN) [26], we aggregate token-level features using a gated multi-scale pooling (GMP) mechanism. GMP combines three complementary pooling strategies—attention pooling, max pooling, and mean pooling—each capturing different aspects of protein–ligand interactions. Attention pooling emphasizes informative tokens through learned importance weights, max pooling preserves the most salient feature responses, and mean pooling summarizes global contextual information across all tokens. Formally, for a token sequence $X\in {\mathbb{R}}^{N\times D}$, the three pooling operations are defined as $X_{\text{attn}}$, $X_{\text{max}}$, and $X_{\text{mean}}$:

$$\begin{array}{c}\hfill X_{\text{attn}}=\sum _i{\alpha }_{i}X[i],\end{array}$$ 8

$$\begin{array}{c}\hfill X_{\text{max}}=\underset{i}{max}X[i],\end{array}$$ 9

$$\begin{array}{c}\hfill X_{\text{mean}}=\frac{1}{N}\sum _{i}X[i],\end{array}$$ 10

where $s_i=\text{MLP}(X[i])$ and $\alpha =\text{softmax}(s)$.

The three pooled representations are concatenated and processed through a gating network to obtain adaptive combination weights:

$$\begin{array}{cc}& (g_{\text{attn}},g_{\text{max}},g_{\text{mean}})\hfill \\ \hfill & =\text{Softmax}(\text{MLP}([X_{\text{attn}};X_{\text{max}};X_{\text{mean}}])),\hfill \end{array}$$ 11

where $[;]$ denotes concatenation.

The final pooled representation is computed as a weighted combination followed by output projection:

$$\begin{array}{c}\hfill \text{GMP}(X)=\text{Proj}(g_{\text{attn}}·X_{\text{attn}}+g_{\text{max}}·X_{\text{max}}+g_{\text{mean}}·X_{\text{mean}}).\end{array}$$ 12

This gating mechanism enables the model to adaptively emphasize different pooling strategies based on input characteristics. Using GMP, pooled representations are obtained for both Single Extractors:

$$\begin{array}{c}\hfill \begin{array}{cc}\hfill P_{\text{pooled}}& =\text{GMP}(P_{\text{fine}}),\hfill \\ \hfill L_{\text{pooled}}& =\text{GMP}(L_{\text{fine}}),\hfill \\ \hfill S_{\text{pooled}}& =\text{GMP}(S_{\text{coarse}}).\hfill \end{array}\end{array}$$ 13

Pair Extractor

To focus on protein–ligand interactions, we extract features only from the interaction region of the pair embedding, consistent with interaction-focused approaches such as Uni-Mol [48]. The full pair embedding Z contains four sections: protein–protein, protein–ligand ($Z_{\mathit{PL}}$), ligand–protein ($Z_{\mathit{LP}}$), and ligand–ligand. We combine the two off-diagonal blocks to obtain a symmetric cross-molecular interaction representation:

$$\begin{array}{c}\hfill Z_{\text{interaction}}=\frac{Z_{\mathit{PL}}+Z_{\mathit{LP}}^{\top }}{2}.\end{array}$$ 14

We flatten $Z_{\text{interaction}}\in {\mathbb{R}}^{N_p\times N_l\times D}$ into $Z_{\text{flat}}\in {\mathbb{R}}^{(N_p{N}_l)\times D}$ and apply the same three-way pooling strategy (attention, max, and mean) as in GMP to obtain $Z_{\text{attn}}$, $Z_{\text{max}}$, and $Z_{\text{mean}}$. The final pair representation is obtained by concatenating and projecting these three pooled vectors:

$$\begin{array}{c}\hfill Z_{\text{pair}}=\text{MLP}([Z_{\text{attn}};Z_{\text{max}};Z_{\text{mean}}]).\end{array}$$ 15

Fusion layer

The features from the Fine-grained and Coarse-grained Single Extractors are combined with pair interaction features from the Pair Extractor to produce two complementary fusion representations at different granularities. Fine-grained Fusion combines pair features with token-level protein and ligand representations to capture local interaction patterns, whereas Coarse-grained Fusion combines them with a global protein–ligand representation to capture context-level patterns. Thus, the model can adaptively use the pair information according to the abstraction level of the single features.

The Fine-grained Fusion feature is computed by combining pooled protein, pooled ligand, and pair interaction features:

$$\begin{array}{c}\hfill F_{\text{fine}}=\text{MLP}([P_{\text{pooled}};L_{\text{pooled}};Z_{\text{pair}}]),\end{array}$$ 16

The Coarse-grained Fusion feature is computed by combining the pooled global representation and pair interaction features:

$$\begin{array}{c}\hfill F_{\text{coarse}}=\text{MLP}([S_{\text{pooled}};Z_{\text{pair}}]).\end{array}$$ 17

Geometric Encoder

Geometric Encoder (GE) leverages the IGN [26] architecture to encode 3D protein–ligand complexes as molecular graphs and handle intra- and intermolecular interactions. Specifically, it first learns atom-level representations from covalent structures and then captures cross-molecular atom–atom interactions through message passing on a bipartite interaction graph. The resulting features are pooled into a graph-level representation of the complex. Within AlphaDTA, GE provides explicit geometry-aware interaction features that complement the embedding-based representations.

Triple Adaptive Fusion

Triple Adaptive Fusion integrates three complementary representations: the fine-grained fusion feature $F_{\text{fine}}$, the coarse-grained fusion feature $F_{\text{coarse}}$, and the geometric feature $F_{\text{geometric}}$. To combine these heterogeneous features, each representation is first projected into a shared latent space. The final fused feature is computed as:

$$\begin{array}{c}\hfill \begin{array}{cc}\hfill F_{\text{fused}}& =w_{\text{fine}}·{\text{Proj}}_{\text{fine}}(F_{\text{fine}})\hfill \\ \hfill & +w_{\text{coarse}}·{\text{Proj}}_{\text{coarse}}(F_{\text{coarse}})\hfill \\ \hfill & +w_{\text{geometric}}·{\text{Proj}}_{\text{geometric}}(F_{\text{geometric}}),\hfill \end{array}\end{array}$$ 18

where $(w_{\text{fine}},w_{\text{coarse}},w_{\text{geometric}})=\text{Softmax}(\mathit{\theta })$ are learnable fusion weights that adaptively balance the contribution of each feature source.

Prediction layer

The fused feature is passed to a fully connected (FC) layer to predict binding affinity:

$$\begin{array}{c}\hfill \widehat{y}=\text{FC}(F_{\text{fused}}).\end{array}$$ 19

The model is trained to minimize the mean squared error (MSE) loss between predicted and experimental binding affinities.

$$\begin{array}{c}\hfill {\mathcal{L}}_{\text{MSE}}=\frac{1}{n}\sum _{i=1}^n{(y_i-{\widehat{y}}_i)}^2,\end{array}$$ 20

where $y_i$ and ${\widehat{y}}_i$ denote the experimental and predicted binding affinities, respectively.

A detailed time complexity analysis of AlphaDTA is provided in Additional file 1: Section S4.4.2. The full AlphaDTA model contains 2,203,413 trainable parameters.

Datasets

PDBbind database

The PDBbind database [27] is a widely used benchmark for protein–ligand binding affinity prediction, providing experimentally measured binding affinities paired with 3D complex structures from the Protein Data Bank (PDB) [49]. In this study, we use PDBbind v2020, which is organized into three hierarchical subsets: the General Set, the Refined Set, and the Core Set.

The General Set is the largest subset, containing 19,443 protein–ligand complexes with experimentally reported binding affinities expressed as $p{K}_d$ ($-log{K}_d$), $p{K}_i$ ($-log{K}_i$), or $pI{C}_{50}$ ($-logI{C}_{50}$), collectively denoted as pK in this work. The Refined Set is a high-quality subset derived from the General Set through stringent filtering criteria. It comprises 5,316 complexes with crystal structure resolution better than 2.5 Å and reliable $K_d$ or $K_i$ measurements, excluding $I{C}_{50}$ values. The Core Set is a carefully selected, non-redundant subset of 285 complexes drawn from the Refined Set via cluster analysis. This subset forms the CASF2016 (Comparative Assessment of Scoring Functions) benchmark [50], which evaluates the generalization of scoring functions across diverse protein families. CASF2016 consists of 57 target-based clusters, each containing five complexes sharing the same protein target but different ligands and binding affinities.

External benchmark datasets

Three additional external benchmark datasets from Li et al. [51]—BDB2020+ (a BindingDB-derived dataset), Mpro (SARS-CoV-2 main protease complexes), and EGFR (epidermal growth factor receptor complexes)—are employed to assess generalization across varying similarity profiles. Detailed descriptions are provided in the Experimental Setup section.

Experimental setup

A common practice is to use the PDBbind General Set (excluding CASF2016) for training and validation and to evaluate performance on CASF2016 as an independent test set [16, 21, 24–26]. However, recent studies [26, 45, 51, 52] have reported a critical data leakage issue in this conventional split. High structural and sequence similarities between training and test complexes may lead to overly optimistic performance estimates, as models can rely on memorization rather than learning generalizable binding principles.

To mitigate data leakage and enable a more rigorous evaluation of generalization, we construct our experiments using two reorganized versions of PDBbind v2020: Leak-Proof PDBbind [51] and PDBbind CleanSplit [45]. Although both are derived from the same PDBbind release, they differ in their splitting strategies. Leak-Proof PDBbind controls protein sequence similarity across subsets, whereas PDBbind CleanSplit removes structurally similar complexes based on 3D protein–ligand alignment to prevent overlap with CASF2016. Together, these datasets enable evaluation under sequence-level and structure-level dissimilarity.

In addition, we evaluate performance using both experimentally determined crystal structures and AlphaFold3-predicted structures, allowing assessment under validated geometries as well as robustness to structural prediction uncertainty. In all experiments, model performance is evaluated in terms of the Pearson correlation coefficient (PCC) and root mean squared error (RMSE) (Additional file 1: Section S3).

During preprocessing and AlphaFold3-based structure generation, a subset of complexes could not be processed due to molecule-parsing or graph-construction errors. These failures were excluded from the corresponding analyses, and all performance comparisons were conducted on the intersection of complexes successfully processed by all compared methods. A detailed analysis of the failure types, excluded complexes, and potential mitigation strategies is provided in Additional file 1: Section S6.

LP-PDBbind

We first evaluate AlphaDTA on the Leak-Proof PDBbind (LP-PDBbind) dataset [51], which is designed to mitigate data leakage through two approaches: quality-based filtering and similarity-controlled splitting. For quality-based filtering, LP-PDBbind defines two cleaning levels: Clean Level 1 (CL1) and Clean Level 2 (CL2). CL1 removes covalent binders, complexes with structural defects (e.g., steric clashes or distant ligands), and ligands with uncommon elements or poor drug-likeness, whereas CL2 further restricts affinity measurements to high-confidence $K_d$ or $K_i$ values, excluding $I{C}_{50}$. For similarity-controlled splitting, a maximum protein sequence similarity of 50% and ligand similarity of 99% are enforced between training and validation/test sets. Based on these criteria, LP-PDBbind provides an independent train–validation–test split comprising 7,393 non-covalent CL1 complexes for training, 958 CL2 complexes for validation, and 2,171 CL2 complexes for testing.

To further assess generalization under different similarity regimes, we additionally evaluate on three external benchmarks provided by Li et al. [51]:

• BDB2020+ (115 complexes): A benchmark dataset compiled from BindingDB [53] entries deposited after 2020 and matched with PDB crystal structures, providing temporal separation from PDBbind v2020. The same similarity thresholds used in LP-PDBbind splitting are applied to ensure no overlap with the train set.

• Mpro (40 complexes): A moderate-similarity benchmark dataset targeting the SARS-CoV-2 main protease. The train set contains a related SARS-CoV-1 main protease (PDB ID: 3V3M), yielding partial protein and ligand similarity to the training data.

• EGFR (23 complexes): A challenging low-similarity benchmark dataset in which both the target protein and ligands exhibit minimal similarity to the train set, testing generalization to new protein–ligand systems.

Prediction models are trained using the LP-PDBbind training and validation sets and evaluated on the LP-PDBbind test set as well as the three external benchmarks (BDB2020+, Mpro, and EGFR). All experiments are repeated with five random seeds to ensure generalization across varying similarity profiles.

PDBbind CleanSplit

Experiments are additionally conducted using the PDBbind CleanSplit dataset proposed by Graber et al. [45]. Whereas LP-PDBbind constructs an independent train–validation–test split with sequence-based similarity control across subsets, PDBbind CleanSplit fixes CASF2016 as the external test set and systematically removes training complexes that structurally resemble any CASF2016 entry.

Similarity is assessed using three complementary metrics: protein structural alignment via TM-align scores, ligand chemical similarity via Tanimoto scores, and binding conformation similarity via pocket-aligned ligand RMSD (root mean squared deviation). Training complexes with high combined similarity across these dimensions are excluded to eliminate train–test overlap. Beyond mitigating external leakage, PDBbind CleanSplit also reduces internal redundancies by removing similarity clusters within the training data, ensuring learning from a diverse, non-redundant set.

Prediction models are trained on PDBbind CleanSplit using five-fold cross-validation and evaluated on the fixed CASF2016 benchmark, providing a stringent assessment of generalization to structurally dissimilar complexes.

Cluster-based PDBbind split

To further evaluate generalization under strict data separation, we constructed an additional cluster-based PDBbind split inspired by recent cluster-based evaluation protocols for drug–target interaction prediction [43, 54]. Unlike random or pair-level splitting, this setting separates complexes at both ligand- and protein-cluster levels, thereby reducing overlap in chemical and protein sequence spaces across training, validation, and test sets.

For ligand clustering, compounds were represented using ECFP4 (4-bond extended-connectivity fingerprint) [55] fingerprints, and pairwise ligand distances were computed using the Jaccard distance. For protein clustering, target sequences were represented using pseudo-amino acid composition (PSC) [56] descriptors, and pairwise protein distances were computed using the cosine distance. Ligands and proteins were then independently clustered using single-linkage hierarchical clustering, resulting in 5,649 ligand clusters and 3,761 protein clusters.

For each random seed, ligand clusters and protein clusters were independently partitioned into training, validation, and test groups at a ratio of 60:20:20. A complex was assigned to a split only when both its ligand cluster and protein cluster belonged to the corresponding partition; complexes spanning different partitions were excluded. This procedure ensured that the training, validation, and test sets shared no ligand clusters or protein clusters.

We generated five independent cluster-based splits using different random seeds. For each split, models were trained on the training set, selected using the cluster-disjoint validation set, and evaluated on the held-out test set. Final performance was reported as the mean and standard deviation across the five test sets.

Benchmark methods

To comprehensively evaluate the performance of AlphaDTA, we compare against ten established benchmark methods with diverse architectures and input modalities:

Sequence-based methods

These methods use only protein sequences and ligand SMILES as input:

• DeepDTA [17]: Encodes protein amino acid sequences and ligand SMILES using separate 1D-CNNs and combining the representations via fully connected layers.

• BAPULM [39]: Encodes protein amino acid sequences and ligand SMILES using pretrained language models (ProtT5-XL-U50 [37] for proteins; MolFormer [36] for ligands), then applies a prediction head to the concatenated embeddings.

• PMMR [40]: Fuses pretrained sequence embeddings (ESM2 [38] for proteins; ChemBERTa-2 [35] for ligands) with 2D ligand graph features, integrating sequence semantics and graph topology in a unified predictor.

Complex-based methods

These methods leverage 3D protein–ligand complex structures:

• RF-Score [10]: A Random Forest that uses predefined atom-pair interaction features based on occurrence counts of nine atomic types within a 12 Å distance cutoff.

• IGN [26]: Utilizes 3D molecular interaction graphs using two stacked graph convolution modules to sequentially capture intra-molecular and inter-molecular interactions.

• Search top-5 complexes [45]: A nearest-neighbor baseline that retrieves the five most similar training complexes and predicts affinity by averaging their labels.

• GenScore [57]: Combines graph-based geometric modeling with a probabilistic scoring formulation to estimate interaction strength.

• Pafnucy [25]: Converts 3D protein–ligand structures into voxel grids and learns spatial interaction patterns with 3D-CNNs.

• GEMS [45]: Constructs spatial interaction graphs by integrating 3D proximity-based interactions with pretrained protein/ligand language-model embeddings to capture sequence-level semantics.

• AutoDock Vina [58]: A traditional empirical scoring function widely used for protein–ligand docking and scoring. In this study, AutoDock Vina was used in a scoring-only setting, where complex structures were directly scored without docking or model training (Additional file 1: Section S4.7).

To determine whether the performance gains of AlphaDTA arise solely from the use of AlphaFold3 embeddings or from the proposed MGE, we introduce a simplified variant:

• AlphaDTA (Simplified MGE): A simplified version of AlphaDTA where the MGE is replaced by a simplified embedding processing module (Additional file 1: Section S5.1).

The baseline comparisons followed the benchmark sets used in the original studies that introduced each evaluation setting. For the LP-PDBbind setting, we followed Li et al. [51], including DeepDTA, RF-Score, IGN, and AutoDock Vina, and additionally included AlphaDTA (Simplified MGE) as our model variant. For the PDBbind CleanSplit/CASF2016 setting, we followed Graber et al. [45], including Search top-5 complexes, GenScore, Pafnucy, and GEMS. BAPULM and PMMR were included separately for the pretrained-representation comparison. For the cluster-based PDBbind split, we included all baselines evaluated across both settings: DeepDTA, RF-Score, IGN, AutoDock Vina, Search top-5 complexes, GenScore, Pafnucy, and GEMS. Detailed hyperparameter configurations and search procedures are provided in Additional file 1: Section S4.

Results

Performance under LP-PDBbind

All models except AutoDock Vina were trained on the LP-PDBbind training and validation sets across five random seeds. AutoDock Vina was instead used as a scoring-only baseline by directly scoring the corresponding complex structures without docking. All models were evaluated on the LP-PDBbind test set and three external benchmarks (BDB2020+, Mpro, and EGFR). Complex-based models were evaluated using either experimentally determined crystal or AF3-predicted structures; sequence-based baselines are included in both blocks for reference.

With crystal structures

On the LP-PDBbind test set with crystal structures, AlphaDTA achieved the best performance (RMSE $=1.529$, PCC $=0.578$), outperforming all baselines (Table 1, Crystal). Similar trends were observed on external benchmarks, including BDB2020+ (RMSE $=0.981$, PCC $=0.604$) and EGFR (RMSE $=0.613$, PCC $=0.714$), where AlphaDTA achieved the best performance in both metrics. On Mpro, RF-Score achieved the lowest RMSE (RMSE $=0.756$), followed by AlphaDTA with the second-lowest RMSE (RMSE $=0.808$); notably, AlphaDTA obtained the highest PCC (PCC $=0.721$). AlphaDTA also consistently outperformed AlphaDTA (Simplified MGE) across all benchmarks in both RMSE and PCC. Overall, AlphaDTA demonstrated robust generalization across diverse benchmarks with varying protein sequence similarity profiles, achieving the best results on LP-PDBbind, BDB2020+, and EGFR, while remaining competitive on Mpro.

Table 1.

Comparison of AlphaDTA and baseline models on LP-PDBbind and external benchmarks

Model	LP-PDBbind (n=2005)		BDB2020+ (n=105)		EGFR (n=13)		Mpro (n=27)
	RMSE $\downarrow$	PCC $\uparrow$	RMSE $\downarrow$	PCC $\uparrow$	RMSE $\downarrow$	PCC $\uparrow$	RMSE $\downarrow$	PCC $\uparrow$
Crystal
AlphaDTA	1.529 (0.039)	0.578 (0.034)	0.981 (0.048)	0.604 (0.019)	0.613 (0.059)	0.714 (0.106)	0.808 (0.122)	0.721 (0.025)
Simplified MGE	1.603 (0.068)	0.549 (0.033)	1.028 (0.060)	0.558 (0.033)	0.671 (0.113)	0.580 (0.194)	0.934 (0.133)	0.706 (0.039)
DeepDTA	1.691 (0.056)	0.469 (0.029)	1.258 (0.114)	0.413 (0.046)	0.862 (0.166)	0.522 (0.072)	0.821 (0.159)	0.568 (0.104)
IGN	1.549 (0.031)	0.562 (0.015)	1.050 (0.063)	0.524 (0.031)	0.701 (0.158)	0.613 (0.091)	0.881 (0.184)	0.680 (0.117)
RF-Score	1.567 (0.009)	0.568 (0.006)	1.157 (0.015)	0.520 (0.010)	0.632 (0.033)	0.701 (0.043)	0.756 (0.034)	0.685 (0.018)
AutoDock Vina	2.341 (0.000)	0.394 (0.000)	2.640 (0.000)	0.237 (0.000)	1.938 (0.000)	0.162 (0.000)	0.884 (0.000)	0.339 (0.000)
AF3
AlphaDTA	1.561 (0.028)	0.555 (0.027)	0.996 (0.045)	0.574 (0.017)	0.586 (0.019)	0.729 (0.033)	0.800 (0.119)	0.715 (0.022)
Simplified MGE	1.653 (0.082)	0.511 (0.045)	1.140 (0.156)	0.492 (0.021)	0.646 (0.145)	0.606 (0.158)	0.919 (0.101)	0.682 (0.040)
DeepDTA	1.691 (0.056)	0.469 (0.029)	1.258 (0.114)	0.413 (0.046)	0.862 (0.166)	0.522 (0.072)	0.821 (0.159)	0.568 (0.104)
IGN	1.587 (0.024)	0.529 (0.015)	1.075 (0.094)	0.473 (0.062)	0.769 (0.328)	0.507 (0.409)	0.883 (0.144)	0.608 (0.093)
RF-Score	1.567 (0.007)	0.562 (0.005)	1.080 (0.018)	0.526 (0.013)	0.682 (0.036)	0.580 (0.087)	0.770 (0.051)	0.672 (0.026)
AutoDock Vina	4.867 (0.000)	0.127 (0.000)	4.321 (0.000)	−0.033 (0.000)	6.843 (0.000)	−0.257 (0.000)	0.799 (0.000)	0.571 (0.000)

Results are evaluated on experimentally determined (Crystal) and AlphaFold3-predicted (AF3) complex structures and are reported as the mean (standard deviation) over five independent runs with different random seeds. AutoDock Vina is deterministic, yielding a standard deviation of 0.000. n denotes the number of complexes successfully processed by all methods. Bold and underline indicate the best and second-best results, respectively

With AlphaFold3-predicted structures

On the LP-PDBbind test set with AlphaFold3-predicted structures, AlphaDTA achieved the lowest RMSE (RMSE $=$ 1.561) and second-best PCC (PCC $=0.555$) comparable to RF-Score (PCC $=0.562$) (Table 1, AF3). AlphaDTA achieved the best performance on both BDB2020+ (RMSE $=0.996$, PCC $=0.574$) and EGFR (RMSE $=0.586$, PCC $=0.729$). On Mpro, AlphaDTA achieved the highest PCC (PCC $=$ 0.715), while its RMSE (RMSE $=$ 0.800) was comparable to AutoDock Vina (RMSE $=$ 0.799) and remained close to the best RMSE achieved by RF-Score (RMSE $=$ 0.770). AlphaDTA also outperformed AlphaDTA (Simplified MGE) across all benchmarks in both RMSE and PCC.

Compared with the crystal structure setting, the performance gap between AlphaDTA with AlphaFold3-predicted structures and AlphaDTA with crystal structures was small across most benchmarks. On LP-PDBbind, the difference was $\mathrm{\Delta }$RMSE $=+0.032$ and $\mathrm{\Delta }$PCC $=-0.023$. On BDB2020+, the gap was similarly modest ($\mathrm{\Delta }$RMSE $=+0.015$, $\mathrm{\Delta }$PCC $=-0.030$). On EGFR, AlphaDTA with AlphaFold3-predicted structures even outperformed the crystal structure counterpart, suggesting that AlphaFold3-predicted structures can provide complementary structural information in some settings. On Mpro, the gap remained marginal ($\mathrm{\Delta }$RMSE $=-0.008$, $\mathrm{\Delta }$PCC $=-0.006$). Overall, these results indicate that AlphaDTA with AlphaFold3-predicted structures achieves performance closely comparable to its crystal structure counterpart, supporting the practical utility of the framework when experimentally determined complex structures are unavailable.

Performance under PDBbind CleanSplit

All models were trained on the PDBbind CleanSplit via 5-fold cross-validation and evaluated on the fixed CASF2016 test set using either experimentally determined crystal structures or AlphaFold3-predicted structures.

With crystal structures

On CASF2016, AlphaDTA achieved the best overall performance with an average RMSE of 1.289 and PCC of 0.816 across five independently trained models from 5-fold cross-validation ($n=284$; one complex was excluded due to ligand sanitization failure during IGN graph construction), outperforming GEMS (RMSE $=1.308$, PCC $=0.803$) and the remaining baselines (Table 2). Overall, AlphaDTA demonstrated robust generalization to structurally dissimilar protein–ligand complexes across protein structure, ligand chemistry, and binding conformation.

Table 2.

Comparison of AlphaDTA and baseline models on the CASF2016 benchmark

Model	Crystal
	CASF2016 (n=285)
	RMSE $\downarrow$	PCC $\uparrow$
AlphaDTA*	1.289 (0.026)	0.816 (0.011)
GEMS	1.308	0.803
GenScore	1.362	0.780
Pafnucy	1.484	0.746
Search top-5 complexes	1.648	0.653

AlphaDTA results are reported as the mean (standard deviation) across five independently trained models from 5-fold cross-validation. Baseline results were taken from Graber et al. [45], where standard deviations were not reported. Bold and underline indicate the best and second-best results, respectively. ^*Results are reported for $n=284$ complexes

With AlphaFold3-predicted structures

For the CASF2016 evaluation using AlphaFold3-predicted structures, we focused on GEMS as the primary comparator because it was the strongest baseline in the CleanSplit/CASF2016 benchmark using crystal structures (Fig. 2). Figure 2a shows that AlphaDTA achieved lower RMSE and higher PCC than GEMS (AlphaDTA: RMSE $=1.306$, PCC $=0.824$; GEMS: RMSE $=1.350$, PCC $=0.803$) on CASF2016 using AlphaFold3-predicted structures, with ensemble predictions averaged over the five cross-validation models. Figure 2b compares per-cluster PCC on the subset of CASF2016 clusters for which all five complexes were successfully processed under AlphaFold3 ($n=44$ clusters). The remaining 13 clusters were excluded from the per-cluster analysis because at least one complex in each cluster failed during IGN graph construction for AlphaFold3-predicted structures (Additional file 1: Section S6), preventing complete five-complex cluster-level PCC calculation. AlphaDTA achieved higher PCC than GEMS in 32 clusters, whereas GEMS achieved higher PCC than AlphaDTA in 12 clusters. These results demonstrate that AlphaDTA maintains robust performance even under the structural noise in AlphaFold3-predicted structures, outperforming GEMS, the best-performing baseline. Furthermore, the cluster-level comparison shows that AlphaDTA’s gains are broadly distributed across diverse target families (32 out of 44 clusters), rather than being driven by a small number of favorable cases.

Fig. 2.

Comparison of AlphaDTA and GEMS on CASF2016 using AlphaFold3-predicted structures. Results are computed from ensemble predictions by averaging five independently trained models via 5-fold cross-validation. a Overall PCC and RMSE on the intersection set of complexes successfully processed by both methods ($n=263$ out of 285). Bold indicates the better result. b Per-cluster PCC comparison on CASF2016 clusters for which all five complexes were successfully processed under AF3 ($n=44$ out of 57). Each point represents one cluster; points above the diagonal favor AlphaDTA

Performance under cluster-based splitting

We further evaluated AlphaDTA using the cluster-based PDBbind split, where training, validation, and test complexes were separated at both ligand- and protein-cluster levels. This setting provides a strict assessment of generalization because model selection is also performed on a validation set composed of clusters unseen during training. All complex-based models in this comparison were evaluated using experimentally determined crystal structures, while AlphaDTA was additionally evaluated using AlphaFold3-predicted structures to assess its applicability when crystal complex structures are unavailable. The results are summarized in Table 3.

Table 3.

Comparison of AlphaDTA and baseline models under the cluster-based PDBbind split

Model	RMSE $\downarrow$	PCC $\uparrow$
AlphaDTA (Crystal)	1.4871 (0.1226)	0.5789 (0.0710)
AlphaDTA (AF3)	1.5056 (0.1658)	0.5646 (0.0777)
Pafnucy	1.4986 (0.1186)	0.5570 (0.0720)
IGN	1.5140 (0.1610)	0.5274 (0.0843)
GEMS	1.5531 (0.1997)	0.5223 (0.1171)
GenScore	1.5555 (0.1288)	0.5165 (0.0587)
RF-Score	1.5486 (0.1769)	0.5075 (0.1028)
DeepDTA	1.6722 (0.1260)	0.4130 (0.0818)
AutoDock Vina	2.6052 (0.3559)	0.3474 (0.1164)
Search top-5 complexes	1.7902 (0.1083)	0.3409 (0.0665)

Results are reported as the mean (standard deviation) across five independent cluster-based train–validation–test splits. AlphaDTA (Crystal) and all complex-based baselines were evaluated using experimentally determined crystal structures, whereas AlphaDTA (AF3) was evaluated using AlphaFold3-predicted structures. Bold and underline indicate the best and second-best results, respectively

Under experimentally determined crystal structures, AlphaDTA achieved the best overall performance among the evaluated models (RMSE $=1.487$, PCC $=0.579$). AlphaDTA outperformed the second-best model, Pafnucy, in both RMSE (RMSE $=1.499$) and PCC (PCC $=0.557$), indicating that AlphaDTA remains robust even under cluster-disjoint train–validation–test separation. While AlphaDTA using AlphaFold3-predicted structures showed slightly lower performance than its crystal-structure counterpart, it still achieved competitive performance (RMSE $=1.506$, PCC $=0.565$). Notably, AlphaDTA with AlphaFold3-predicted structures achieved a higher PCC than all non-AlphaDTA baseline models, including Pafnucy, IGN, GEMS, GenScore, RF-Score, DeepDTA, AutoDock Vina, and Search top-5 complexes. These results suggest that AlphaDTA can retain strong ranking performance even when experimentally determined complex structures are unavailable, supporting the practical advantage of integrating AlphaFold3-derived structures and embeddings for binding affinity prediction.

Comparison of pretrained representation approaches

AlphaFold3 embeddings encode structural and relational information learned from diverse biomolecular complexes, unlike independently pretrained language models that process protein sequences and ligand SMILES separately. To assess whether this distinction translates to improved binding affinity prediction, AlphaDTA (w/o GE)—a variant of AlphaDTA that relies solely on AlphaFold3’s embeddings without the Geometric Encoder—is compared against BAPULM and PMMR, which employ independently pretrained language models (ProtT5, ESM2 for proteins; MolFormer, ChemBERTa-2 for ligands).

Across benchmarks, AlphaDTA (w/o GE) consistently showed strong performance (Table 4). On the LP-PDBbind test set, it achieved the best overall results (RMSE $=1.573$, PCC $=0.559$), outperforming BAPULM (RMSE $=1.580$, PCC $=0.530$) and PMMR (RMSE $=1.650$, PCC $=0.494$). On BDB2020+, AlphaDTA (w/o GE) achieved a competitive PCC (0.467) relative to BAPULM (0.504). However, AlphaDTA (w/o GE) demonstrated superior predictive accuracy with a significantly lower RMSE (1.131) compared to BAPULM (1.261). Moreover, on EGFR, BAPULM failed to generalize (PCC $=-0.051$), whereas AlphaDTA (w/o GE) showed best performance (RMSE $=0.896$, PCC $=0.270$). On Mpro, AlphaDTA (w/o GE) achieved the highest PCC (PCC $=0.586$), outperforming BAPULM (PCC $=0.388$) and PMMR (PCC $=0.421$), with a competitive RMSE that ranked second (RMSE $=0.970$). Overall, AlphaDTA (w/o GE) demonstrated competitive or superior performance relative to BAPULM and PMMR across benchmarks, suggesting that AlphaFold3’s embeddings provide effective representations for binding affinity prediction.

Table 4.

Comparison of AlphaDTA without GE and models with pretrained encoders on LP-PDBbind and external benchmarks

Model	LP-PDBbind (n=2171)		BDB2020+ (n=115)		EGFR (n=23)		Mpro (n=40)
	RMSE $\downarrow$	PCC $\uparrow$	RMSE $\downarrow$	PCC $\uparrow$	RMSE $\downarrow$	PCC $\uparrow$	RMSE $\downarrow$	PCC $\uparrow$
AlphaDTA (w/o GE)	1.573 (0.020)	0.559 (0.022)	1.131 (0.058)	0.467 (0.017)	0.896 (0.097)	0.270 (0.054)	0.970 (0.169)	0.586 (0.053)
BAPULM	1.580 (0.022)	0.530 (0.016)	1.261 (0.154)	0.504 (0.053)	1.185 (0.053)	−0.051 (0.103)	0.941 (0.226)	0.388 (0.080)
PMMR	1.650 (0.031)	0.494 (0.031)	1.222 (0.083)	0.406 (0.057)	1.162 (0.098)	0.237 (0.135)	1.205 (0.346)	0.421 (0.136)

Results are reported as the mean (standard deviation) over five independent runs with different random seeds. n denotes the number of complexes successfully processed by all methods. Bold and underline indicate the best and second-best results, respectively

Ablation study

To systematically evaluate the contributions of each component, we designed an ablation study to examine (i) the complementarity between AlphaFold3 embeddings and geometric representations, (ii) the relative contributions of single and pair embeddings, and (iii) the effectiveness of the Fine-grained and Coarse-grained components. As shown in Table 5, all variants are evaluated using experimentally determined crystal structures and averaged over five independent runs. Since removing individual components changes model capacity and optimization behavior, each ablation variant was independently tuned using the same predefined hyperparameter search space to avoid penalizing any variant due to hyperparameter mismatch. In addition, to examine whether the simultaneous use of attention, max, and mean pooling in Eq. (12) is necessary, we conducted additional ablation studies on pooling strategies. These results are reported in Additional file 1: Section S5.2.

Table 5.

Ablation study of AlphaDTA on LP-PDBbind and external benchmarks

Crystal
Model	LP-PDBbind (n=2153)		BDB2020+ (n=113)		EGFR (n=23)		Mpro (n=27)
	RMSE $\downarrow$	PCC $\uparrow$	RMSE $\downarrow$	PCC $\uparrow$	RMSE $\downarrow$	PCC $\uparrow$	RMSE $\downarrow$	PCC $\uparrow$
AlphaDTA	1.505 (0.036)	0.584 (0.031)	1.003 (0.046)	0.574 (0.025)	0.716 (0.033)	0.530 (0.089)	0.808 (0.122)	0.721 (0.025)
w/o Single&Pair	1.524 (0.030)	0.569 (0.014)	1.055 (0.063)	0.507 (0.032)	0.848 (0.166)	0.526 (0.079)	0.881 (0.184)	0.680 (0.117)
w/o Single	1.516 (0.021)	0.583 (0.010)	1.102 (0.030)	0.482 (0.032)	0.710 (0.063)	0.574 (0.073)	0.916 (0.194)	0.688 (0.065)
w/o Pair	1.583 (0.053)	0.548 (0.035)	1.061 (0.090)	0.481 (0.062)	0.776 (0.097)	0.527 (0.061)	0.950 (0.173)	0.689 (0.018)
w/o Fine-grained	1.556 (0.028)	0.570 (0.017)	1.151 (0.049)	0.465 (0.054)	0.748 (0.090)	0.530 (0.107)	1.010 (0.297)	0.587 (0.106)
w/o Coarse-grained	1.570 (0.024)	0.557 (0.010)	1.190 (0.111)	0.508 (0.053)	0.763 (0.047)	0.505 (0.096)	0.752 (0.059)	0.594 (0.099)

Results are evaluated on experimentally determined (crystal) structures and are reported as the mean (standard deviation) over five independent runs with different random seeds. For fair comparison, each model variant was independently optimized through hyperparameter tuning. n denotes the number of complexes successfully processed by all methods. Bold and underline indicate the best and second-best results, respectively

Complementarity of AlphaFold3 embeddings and geometric representations

Removing AlphaFold3 single and pair embeddings consistently reduces performance. Compared to the full model, the w/o Single&Pair variant shows degraded RMSE (1.505 $\to$ 1.524) and PCC (0.584 $\to$ 0.569) on LP-PDBbind, as well as across all external benchmarks (BDB2020+: RMSE 1.003 $\to$ 1.055, PCC 0.574 $\to$ 0.507; Mpro: RMSE 0.808 $\to$ 0.881, PCC 0.721 $\to$ 0.680). On EGFR, the variant also exhibits degraded RMSE (0.716 $\to$ 0.848) with a slight decrease in PCC (0.530 $\to$ 0.526). These findings indicate that AlphaFold3 embeddings provide complementary information beyond geometric features.

Pair vs. Single embeddings

Across benchmarks, removing pair embeddings (w/o Pair) generally led to larger degradations than removing single embeddings (w/o Single). Compared to the full model, the performance changes were as follows: LP-PDBbind test set ($\mathrm{\Delta }$RMSE $=+0.078$ (w/o Pair), $+0.011$ (w/o Single); $\mathrm{\Delta }$PCC $=-0.036$ (w/o Pair), $-0.001$ (w/o Single)), BDB2020+ ($\mathrm{\Delta }$RMSE $=+0.058$ (w/o Pair), $+0.099$ (w/o Single); $\mathrm{\Delta }$PCC $=-0.093$ (w/o Pair), $-0.092$ (w/o Single)), EGFR ($\mathrm{\Delta }$RMSE $=+0.060$ (w/o Pair), $-0.006$ (w/o Single); $\mathrm{\Delta }$PCC $=-0.003$ (w/o Pair), $+0.044$ (w/o Single)), and Mpro ($\mathrm{\Delta }$RMSE $=+0.142$ (w/o Pair), $+0.108$ (w/o Single); $\mathrm{\Delta }$PCC $=-0.032$ (w/o Pair), $-0.033$ (w/o Single)). This suggests a stronger contribution from pair embeddings in most settings. Furthermore, the full model outperformed both w/o Pair and w/o Single on LP-PDBbind, BDB2020+, and Mpro, indicating that combining pair and single embeddings is typically beneficial. EGFR represents a target-specific exception, where w/o Single marginally outperformed the full model in PCC.

Fine-grained and Coarse-grained components

Removing the Fine-grained component (w/o Fine-grained) consistently degraded performance on LP-PDBbind test set (RMSE 1.505 $\to$ 1.556; PCC 0.584 $\to$ 0.570), BDB2020+ (RMSE 1.003 $\to$ 1.151; PCC 0.574 $\to$ 0.465), and Mpro (RMSE 0.808 $\to$ 1.010; PCC 0.721 $\to$ 0.587), and also increased RMSE on EGFR (0.716 $\to$ 0.748), with no change in PCC. Removing the Coarse-grained component (w/o Coarse-grained) likewise reduced performance on LP-PDBbind test set (RMSE 1.505 $\to$ 1.570; PCC 0.584 $\to$ 0.557) and BDB2020+ (RMSE 1.003 $\to$ 1.190; PCC 0.574 $\to$ 0.508), and also degraded EGFR performance (RMSE 0.716 $\to$ 0.763; PCC 0.530 $\to$ 0.505). On Mpro, w/o Coarse-grained showed a lower RMSE (RMSE $=0.752$ vs. 0.808) but a lower PCC (PCC $=0.594$ vs. 0.721). Overall, both the Fine-grained and Coarse-grained components of MGE are essential to AlphaDTA, with the Fine-grained component capturing detailed protein–ligand cross-interaction patterns and the Coarse-grained component capturing global binding context.

Case study: CFTR potentiator identification

To demonstrate the practical applicability of AlphaDTA, we conducted a case study targeting cystic fibrosis transmembrane conductance regulator (CFTR) to identify candidate potentiators among FDA-approved drugs. A total of 979 FDA-approved drugs (excluding cancer drugs) were screened. Ivacaftor, a clinically approved CFTR potentiator with well-characterized binding interactions [61], was included as a benchmark compound to validate our computational pipeline.

For each drug, AlphaFold3 was used to generate drug–target complex structures and embeddings from the CFTR amino acid sequence and drug SMILES. The resulting outputs served as input for AlphaDTA (trained on the PDBbind CleanSplit) to predict ensemble pK values. Binding affinity prediction was successfully completed for 904 of the 979 FDA-approved drugs considered; the remaining 75 drugs were excluded because ligand sanitization failed during IGN graph construction from AlphaFold3-predicted structures (Additional file 1: Section S6). To prioritize candidates binding to the therapeutically relevant site, we established the CFTR–ivacaftor crystal structure (PDB ID: 6O2P) as our reference. The binding site is composed of ten residues: A309, G930, L233, F236, F305, F312, F931, F932, S308, and Y304. Drugs were selected as candidates if their AlphaFold3-predicted binding sites overlapped with the reference binding site by at least five residues. Among 904 drugs for which binding affinity prediction was successfully completed, 50 drugs met this criterion (Additional file 1: Section S7).

For ivacaftor, the AlphaFold3-predicted binding site overlapped with the reference site at nine residues and achieved a predicted pK of 7.1, the highest value among all candidates. This demonstrates that AlphaDTA successfully identified ivacaftor, a clinically validated potentiator, as the top-ranked candidate within the FDA-approved drug library without relying on experimentally determined complex structures.

Among the 50 candidates, closantel emerged as the most promising compound, achieving the highest predicted pK (6.736) after ivacaftor while exhibiting eight common residues with the reference binding site. Analysis of protein–ligand interactions using PLIP (Protein–Ligand Interaction Profiler) [59] revealed that ivacaftor formed hydrogen bonds with S308, $\pi$–$\pi$ stacking with F312, and hydrophobic contacts with Y304, F305, F236, and F931 (Fig. 3a). Closantel similarly interacted with multiple binding site residues critical for potentiator function (Fig. 3b), forming hydrophobic interactions with F316, F305, and F931, $\pi$–$\pi$ stacking with F312, and hydrogen bonds with S308 and Y304. The importance of these interactions is supported by experimental evidence. Crystallographic investigations of CFTR–potentiator complexes demonstrated that S308 forms hydrogen bonds with bound potentiators, while F312 forms critical $\pi$–$\pi$ stacking interactions [61]. Experimental studies using site-directed mutagenesis showed that disrupting these interactions by replacing S308 or Y304 with alanine dramatically reduced potentiator efficacy, confirming these residues as critical determinants of drug binding [61].

Fig. 3.

Protein–ligand interaction profiles of a ivacaftor, b closantel, and c xipamide with CFTR. Grey dashed lines indicate hydrophobic contacts, blue dashed lines indicate hydrogen bonds, and green dashed lines indicate $\pi$–$\pi$ stacking interactions. 2D interaction diagrams were generated using PLIP [59] and plip-to-2D [60] software. 3D binding poses of the three compounds are shown in Additional file 1: Figure S2

Not all candidates with high common residue counts exhibited favorable binding profiles. For instance, xipamide shared eight common residues with the reference site (equal to closantel), yet its predicted pK (6.003) was substantially lower than closantel’s. PLIP analysis revealed that xipamide formed hydrophobic interactions, $\pi$–$\pi$ stacking, and hydrogen bonds centered on F931, along with additional hydrophobic contacts involving F312, F316, and L233 (Fig. 3c). Notably, xipamide lacked the critical hydrogen bonding interactions with residues experimentally shown to be essential for potentiator binding (S308 and Y304) [61].

Within this computational screening of FDA-approved drugs, closantel was identified as a potential CFTR potentiator candidate. This assessment was based on its relatively high predicted binding affinity and predicted interactions with multiple residues previously implicated in potentiator activity [61], suggesting that closantel could be considered for experimental validation.

Discussion

Under leakage-controlled configurations derived from PDBbind v2020, AlphaDTA achieved state-of-the-art or competitive performance across all benchmarks, including external test sets with varying similarity profiles to the training data, using both experimentally determined crystal structures and AlphaFold3-predicted structures. Under a strict cluster-based split, where training, validation, and test sets were separated at both ligand- and protein-cluster levels, AlphaDTA continued to achieve the best overall performance among all evaluated models. This consistency across leakage-controlled, cluster-disjoint, and external evaluation settings suggests that AlphaDTA learns representations that generalize beyond sequence or structural overlap with the training data.

Beyond prediction accuracy, the consistent performance on AlphaFold3-predicted structures has practical implications for drug discovery. By leveraging AlphaFold3-predicted complex structures and their single and pair embeddings, AlphaDTA enables binding affinity prediction using only protein sequences and ligand SMILES, without requiring experimentally resolved complex structures. Additionally, the CFTR case study demonstrates the practical utility of the pipeline. Using only sequence-derived inputs, the model ranked ivacaftor—a clinically approved CFTR potentiator—as the top candidate among FDA-approved drugs and suggested closantel as a potential drug repurposing candidate.

The performance of AlphaDTA critically depends on how representations derived from AlphaFold3 are utilized. A variant relying solely on AlphaFold3 embeddings (without GE) achieved superior or competitive performance compared to methods based on independently pretrained protein and ligand encoders (BAPULM and PMMR), suggesting that AlphaFold3 embeddings provide sufficient representational capacity for affinity prediction. Ablation studies further suggest that pair embeddings are more informative than single embeddings, as their removal leads to greater performance degradation across benchmarks. Comparisons with the simplified MGE variant suggest that these gains cannot be attributed to the embeddings alone. Instead, they arise from how MGE decomposes and integrates these representations across multiple levels of granularity, together with how Triple Adaptive Fusion leverages geometric features from predicted structures. Specifically, the Fine-grained component captures detailed protein–ligand interaction patterns through bidirectional cross-attention, whereas the Coarse-grained component provides global binding context through joint sequence encoding. Therefore, ablation experiments show that removing either component consistently degrades performance, highlighting the benefit of multi-granularity processing.

Despite these strengths, several limitations exist. A central limitation of AlphaDTA is its dependence on AlphaFold3 for protein–ligand complex prediction and representation extraction. First, this dependence introduces substantial computational cost, making AlphaDTA unsuitable for initial screening of ultra-large compound libraries containing millions of molecules. Because AlphaDTA requires AlphaFold3 inference for each protein–ligand pair, the main computational overhead arises from MSA search and template retrieval. Therefore, AlphaDTA is better suited for later-stage affinity prediction and re-ranking after the initial library has been reduced to a smaller candidate set, such as hundreds to a few thousand compounds. Second, limitations in AlphaFold3-derived representations can propagate to downstream affinity predictions. Subtle mutation-induced structural effects, local binding-site rearrangements, or alternative conformational states may not be fully captured by a single AlphaFold3-predicted complex representation. Consistent with this limitation, our additional mutation-response evaluation showed reduced sensitivity to fine-grained affinity changes between closely related protein variants (Additional file 1: Section S8).

In future work, AlphaDTA could be further improved through architectural extensions. The current architecture processes AlphaFold3 embeddings and geometric features through separate encoders before fusion. Integrating AlphaFold3 embeddings more directly into the graph construction process could enable richer interactions between structural and embedding information. In addition, the Geometric Encoder may benefit from recent advances in geometric deep learning, such as equivariant graph neural networks, which are designed to better capture the symmetries of molecular systems. We anticipate that early embedding–geometry integration, modern geometric architectures, and continued advances in protein–ligand structure prediction will further strengthen AlphaDTA and related affinity prediction frameworks.

Conclusion

AlphaDTA provides a framework for drug–target binding affinity prediction that leverages AlphaFold3-predicted complex structures and embeddings, reducing dependence on experimentally resolved protein–ligand complex structures. Across leakage-controlled evaluation settings derived from PDBbind, AlphaDTA achieved state-of-the-art or competitive performance on benchmarks designed to reduce sequence- and structure-level train–test overlap, using both crystal and AlphaFold3-predicted structures. The CFTR case study further demonstrated the practical utility of the framework by prioritizing ivacaftor, a clinically approved potentiator, and identifying closantel as a promising repurposing candidate among FDA-approved drugs. Together, these results support AlphaDTA as an effective approach for affinity prediction and structure-guided drug discovery when experimentally determined complex structures are unavailable.

Abbreviations

AF3

AlphaFold3

CASF2016

Comparative Assessment of Scoring Functions 2016

CFTR

Cystic fibrosis transmembrane conductance regulator

CL1

Clean level 1

CL2

Clean level 2

DTA

Drug–target affinity

ECFP4

4-bond extended-connectivity fingerprint

EGFR

Epidermal growth factor receptor

FC

Fully connected

FDA

U.S. Food and Drug Administration

FFN

Feed-forward network

GE

Geometric Encoder

GMP

Gated multi-scale pooling

IGN

InteractionGraphNet

LN

Layer normalization

LP-PDBbind

Leak-Proof PDBbind

MGE

Multi-Granularity Encoder

MLP

Multilayer perceptron

Mpro

SARS-CoV-2 main protease

MSE

Mean squared error

PCC

Pearson correlation coefficient

PDB

Protein Data Bank

pK

Collective notation for pKd, pKi, pIC50

PLI

Protein–ligand interaction

PLIP

Protein–ligand interaction profiler

PSC

Pseudo-amino acid composition

RMSE

Root mean squared error

RMSD

Root mean squared deviation

RTP

Residual Token Projection

SMILES

Simplified molecular-input line-entry system

3D

Three-dimensional

Author contributions

MC developed and implemented the proposed framework and conducted the experiments. MC, SP, and HL analyzed and interpreted the results. MC drafted the manuscript. SP contributed to result interpretation and manuscript revision. HL conceived and supervised the study and contributed to the study design and manuscript revision. All authors read and approved the final manuscript.

Funding

This work was supported by a National Research Foundation of Korea (NRF) grant funded by the Korean government (MSIT) (RS-2025-00519063), an Institute of Information & Communications Technology Planning & Evaluation (IITP) Grant funded by the Korean government (MSIT) (No. 2019-0-00567, Development of Intelligent SW Systems for Uncovering Genetic Variation and Developing Personalized Medicine for Cancer Patients with Unknown Molecular Genetic Mechanisms and No.2019-0-01842, Artificial Intelligence Graduate School Program (GIST)), and InnoCORE program of the Ministry of Science and ICT(GIST InnoCORE KH0860).

Data availability

The PDBbind v2020 dataset used in this study is publicly available at https://www.pdbbind-plus.org.cn/. The preprocessed datasets supporting the conclusions of this article are available in the Hugging Face repository at https://huggingface.co/datasets/DMCB-GIST/alphadta. This repository includes the preprocessed interaction graphs and embeddings required to reproduce the reported experiments; therefore, rerunning AlphaFold3 is not required. The source code for AlphaDTA is available on GitHub at https://github.com/DMCB-GIST/AlphaDTA.

Declarations

Competing interests

The authors declare no competing interests.