LiBRe: A Ligand-Aware Sequence-Based Binding Residue Prediction Model for Virtual Screening

Abstract

Identifying protein–ligand binding residues is fundamental to unlocking molecular recognition and advancing therapeutic development. Sequence-based deep learning models for predicting protein–ligand binding residues have gained attention due to their scalability and ability to operate without relying on structural information. However, most existing methods primarily focus on protein sequence information without considering ligand information, even though binding residues are inherently defined through interactions with specific ligands. To address this, we propose a ligand-aware sequence-based binding residue prediction model that explicitly incorporates both residue-level information from protein sequences and ligand information. The proposed model achieved significant improvements in the prediction of ligand-binding residues, outperforming both existing sequence-based and structure-based baselines. Furthermore, pockets defined by the ligand-binding residues predicted by our model led to a stronger and more stable binding affinity compared to existing tools. These results demonstrate that our model shows significant potential for applications in virtual screening and drug discovery. Our source code is publicly available at https://github.com/GoldRiver0/LiBRe.

Introduction

Precise prediction of protein–ligand interactions is crucial for effective virtual screening and the rational design of novel therapeutics. , Extending this, identifying key interaction residues can enhance the efficiency of screening by enabling the exclusion of ligands that do not engage with the functionally relevant regions of the target protein, i.e., by focusing on compounds with a higher likelihood of interacting with biologically relevant regions of the target protein. − This focused approach, integrated with large-scale chemical library screening, facilitates the rapid identification of potential drug-like compounds. −

To predict protein–ligand binding residues, dozens of machine learning approaches have been developed, which are typically classified into two categories: structure-based and sequence-based approaches. With the advent of deep learning, both approaches have achieved significant improvements through more effective modeling of residue–ligand interaction patterns. Despite their accuracy, their practical application in large-scale drug discovery is limited due to their intensive computational requirements and limited performance across diverse targets.

Structure-based approaches generally utilize three-dimensional protein structures to examine the geometric and physicochemical characteristics of protein surfaces in order to identify potential binding residues. Fpocket is a geometry-based technique that detects ligand-binding pockets by clustering α spheres obtained through Voronoi tessellation. P2Rank predicts ligand-binding sites by learning ligandability patterns on the protein surface using a random forest classifier that analyzes solvent-accessible points. GraphPLBR treats the prediction as a graph-based node classification task, leveraging a deep graph convolutional network to model the spatial relationships among residues. Although these approaches show solid predictive performance, their reliance on high-quality protein structures restricts their use when structural data are lacking or incomplete. Additionally, many require extensive molecular docking or energy calculations, which are computationally intensive and less feasible for large-scale data sets.

Sequence-based approaches have recently gained attention as an alternative. These methods infer ligand-binding residues using residue-level features derived only from protein sequences. Each residue is typically encoded as a multidimensional feature vector capturing evolutionary, structural, and biochemical properties, such as position-specific scoring matrices (PSSMs), secondary structure (SS), and evolutionary conservation scores. , Deep learning models then transform these feature vectors into residue-level embeddings, which are processed through convolutional or residual architectures to learn contextual sequence patterns. Representative models such as DeepCSeqSite and BiRDS follow this modeling strategy by encoding residue features and learning binding patterns from sequence context alone. Another sequence-based model is Pseq2sites, which leverages protein language models to construct residue embeddings directly from the amino acid sequence. While sequence-based models have shown comparable predictive performance, they still face a critical limitation in that they do not consider ligand information. Since binding residues are defined by their interactions with ligands, incorporating ligand information is essential for accurate prediction. However, most existing sequence-based models rely solely on residue-level features derived from the protein sequence, without incorporating any ligand information. In addition to this limitation, sequence-based models are affected by the class imbalance in binding residue data sets, where binding residues constitute only a small portion of the entire sequence. This class imbalance affects the behavior of sequence-based models in different ways. Some models achieve high recall by predicting many residues as binding sites, which can lower overall accuracy and increase false positives. Other models predict the majority of residues as nonbinding, which results in high overall accuracy but a failure to detect actual binding sites. This makes it difficult to trust the predictions in practice.

To overcome the limitations of existing sequence-based models, we propose LiBRe, a novel deep learning model for predicting protein–ligand binding residues that incorporates both protein sequence and ligand information. Unlike most previous approaches that rely on protein-derived features, LiBRe integrates information from both the protein and the ligand, enabling ligand-specific prediction of binding residues for a given protein target. The model constructs residue representations using a pretrained protein language model, which eliminates the need for tools such as PSSM or multiple sequence alignment. For ligands, we build graph-based embeddings using a graph attention network that captures both structural and chemical features at the atom level. The combined representations allow the model to predict binding residues in a ligand-aware context. To improve robustness in the presence of imbalanced data, we apply contrastive learning, which helps the model distinguish binding residues from nonbinding ones more effectively during training. We demonstrated the feasibility of this approach by constructing binding pockets centered on the predicted ligand-binding residues.

Methods

Data Sets

We adopted two publicly available data sets, PDBbind­(v.2020) and sc-PDB (v.2017) to construct our training data set. Although our model is sequence-based and does not use structural information as input, these structure-based data sets were employed to obtain reliable residue–ligand binding annotations. PDBbind­(v.2020) provides a comprehensive collection of protein–ligand complexes with experimentally determined binding affinities and associated 3D structural information derived from the Protein Data Bank (PDB). Similarly, sc-PDB is a curated resource that focuses on druggable binding sites within protein–ligand complexes, also extracted from the PDB. For each complex, it provides detailed structural information, including separate MOL2 files for the ligand, the corresponding binding site and the associated protein chain(s). Binding residue information for the training set was obtained from these annotated binding site structures as provided in the data sets. For evaluation, we used COACH420 and HOLO4K, two benchmark test data sets first introduced in P2Rank, and their binding residue information was obtained according to the binding residue definition used in that study.

Data Preprocessing

To ensure the quality and reliability of the data sets used for model training and evaluation, we applied several preprocessing strategies to remove uninformative or redundant samples. Based on the findings of a previous study that highlighted the adverse effects of excessive sequence padding on model performance and computational efficiency, we excluded protein sequences longer than 1500 residues. Such long sequences accounted for less than 1% of the entire data set, and including them would require excessive padding, resulting in inefficient batching and increased memory usage during training. Additionally, similar to the filtering strategy reported in another study, we removed samples containing fewer than 10 binding residues. Since our objective is to accurately identify protein–ligand binding sites, samples with very few binding residues provide limited information for the model to learn meaningful patterns associated with binding, making it more challenging to distinguish between binding and nonbinding residues. We then applied CD-HIT, with a sequence identity threshold of 0.9 to process protein sequences in the training set. Because ligand-binding residue prediction is highly class-imbalanced, preserving sufficient positive binding residue signals during training is critical. In bioinformatics prediction tasks with similar data imbalance, − a 90% sequence identity cutoff is commonly used to remove near-duplicate sequences while maintaining sufficient training data volume and diversity. Accordingly, we adopted a 0.9 threshold to balance redundancy reduction with effective binding-residue signal preservation during model training. After constructing the training set, we further minimized potential redundancy between the training and test sets by ensuring that no proteins shared the same PDB IDs across the two splits. In addition, to enable a rigorous evaluation of ligand-binding residue prediction performance, we followed prior work such as Pseq2sites, which constructed unseen test data sets by removing structurally similar proteins based on TM-score. Similarly, we used Foldseek to assess structural similarity and excluded from the test set any proteins exhibiting high structural similarity (Foldseek similarity >0.4) to the training data. Finally, because our model is designed to incorporate ligand information, we included only those protein–ligand complexes whose ligand structures could be successfully converted into canonical SMILES format using Open Babel and the RDKit Python package. Details of the number of binding and nonbinding residues in each data set after preprocessing are provided in Table .

1. Statistics of the Datasets Used in This Study .

data set	binding residues	nonbinding residues	% of binding residues
train	1,256,721	9,376,609	11.82
test_HOLO4K	52,846	807,221	6.14
test_COACH420	2646	38,936	6.36

^aThe columns give, in order, the dataset name, the number of binding residues, the number of nonbinding residues, and the percentage of the binding residues out of the total.

Residue Representation

Building on previous studies that employed protein language models for residue representation, , which showed clear advantages over traditional handcrafted features, we also adopted this strategy and encoded each residue using the tokenizer of ESM-2. This approach eliminates the need for external tools such as PSI-BLAST, HHblits, and DSSP, reducing the time and cost associated with preprocessing.

Ligand Representation

Since our model explicitly incorporates ligand information for predicting protein–ligand binding residues, accurate and meaningful representation of ligands is essential. In previous compound–protein interaction (CPI) studies, compounds have often been represented as molecular graphs to encapsulate their structural and chemical properties. For example, TransformerCPI and BACPI converted SMILES strings into molecular graphs and applied graph neural networks to construct compound embeddings. Motivated by these approaches, each ligand was converted from its SMILES representation into a molecular graph. Specifically, we employed the RDKit Python package to encode each atom as a 40-dimensional vector based on its chemical properties. A description of these atomic features is provided in Table .

2. List of Ligand Atom Features.

atom type	C, N, O, F, P, S, Cl, Br, I, other
degree of atom	0, 1, 2, 3, 4, 5, 6, 8, 9, 10, 11, 12
formal charge	0 or 1
number of radical electrons	0 or 1
hybridization type	sp, sp2, sp3, sp3d, sp3d2, other
aromatic	0 or 1
number of hydrogen atoms attached	0, 1, 2, 3, 4
chirality	0(false) or 1(true)
configuration	R or S

The Proposed Model Architecture

The proposed model, LiBRe, is a deep learning-based model designed to predict ligand-binding residues within a protein sequence. It consists of four main stages: (i) Residue Encoder, for constructing embeddings of each residue within the protein sequence; (ii) Ligand Encoder, for constructing ligand embeddings; (iii) Residue-Ligand Interaction, for modeling bidirectional interactions between each residue in a protein sequence and the ligand that binds to it; (iv) Ligand-Binding Residue Prediction, for predicting ligand-binding residues based on the residue–ligand interaction embeddings produced by Residue-Ligand Interaction module. The overall model architecture is illustrated in Figure .

1.

Overview of the LiBRe model architecture. (A) Residue Encoder takes a protein sequence as input and constructs residue-level embeddings using ESM-2, 1D-CNN, and BiLSTM. (B) Ligand Encoder processes a SMILES string into a molecular graph and produces a ligand embedding through GAT layers and global mean pooling. (C) Residue–Ligand Interaction module takes the residue and ligand embeddings from (A) and (B), and constructs residue–ligand interaction embeddings through cross-attention between residues and the ligand, followed by element-wise multiplication. Finally, residue–ligand interaction embeddings are passed to a binding residue predictor to predict per-residue binding probabilities.

Residue Encoder

The residue encoder constructs residue-level embeddings from protein sequences. Unlike existing sequence-based methods that construct residue feature map through multiple handcrafted stages, our model directly obtains initial residue embeddings using ESM-2, a pretrained protein language model. Protein language models (PLMs), trained on large-scale protein sequence data sets, are capable of capturing rich contextual, evolutionary, and structural information. Their effectiveness in downstream tasks such as structure prediction, contact map estimation, and functional annotation has been well demonstrated. , By leveraging PLMs, our model obtains biologically meaningful initial embeddings.

Given a protein sequence X = {x ₁, x ₂,..., x _L }, where L is the sequence length and x _i is the ith amino acid residue, ESM-2 produces an initial embedding matrix $R_{\mathrm{init}}\in {\mathbb{R}}^{L\times d}$ , where each row corresponds to a residue-level embedding and d denotes the embedding dimension, as shown in eq .

$$R_{\mathrm{init}}=\mathrm{ESM}-2(X),R_{\mathrm{init}}\in {\mathbb{R}}^{L\times d}$$ 1

Next, to enrich the initial embeddings, we apply a combination of a one-dimensional convolutional neural network (1D-CNN) and a bidirectional long short-term memory network (BiLSTM). This hybrid architecture, widely used in NLP and time-series tasks, is effective for modeling both local and global dependencies in protein sequences. , As a first step, we apply the 1D-CNN to R _init. The 1D-CNN effectively captures local contextual patterns by sliding convolutional filters over the residue embeddings, allowing each residue to reflect local contextual information. We denote the output of the 1D-CNN applied to R _init as R _cnn, where d′ represents the reduced embedding dimension after passing through the 1D-CNN defined in eq .

$$R_{\mathrm{cnn}}=\mathrm{GELU}(\mathrm{Conv}1\mathrm{D}(R_{\mathrm{init}})),R_{\mathrm{cnn}}\in {\mathbb{R}}^{L\times d\prime }$$ 2

To model long-range dependencies, we apply a BiLSTM to R _cnn. The BiLSTM processes the sequence in both forward and backward directions, allowing each residue to incorporate information from the full sequence context. This global awareness is particularly important for identifying ligand-binding residues that are distant in the primary sequence but spatially close in the tertiary structure. The final enhanced residue embedding matrix is computed as follows

$$R=\mathrm{LayerNorm}(\mathrm{BiLSTM}(R_{\mathrm{cnn}})),R\in {\mathbb{R}}^{L\times d″}$$ 3

$$R={\{{\mathrm{v⃗}}_{\mathrm{residue},i}\}}_{i=1}^L,{\mathrm{v⃗}}_{\mathrm{residue},i}\in {\mathbb{R}}^{d″}$$ 4

In eq , R denotes the final residue embedding matrix, where each row v⃗ _residue, i corresponds to the embedding of residue x _i . The matrix R can thus be viewed as the collection {v⃗ _residue, i } _i = 1 , as in eq , and this is used as input to the Residue–Ligand Interaction module.

Ligand Encoder

The ligand encoder processes molecular graphs derived from SMILES representations using a graph attention network (GAT). The GAT allows each atom to assign varying importance to its neighbors and iteratively update its representation based on local chemical context. This enables the model to extract atom-level representations that reflect substructure-specific properties of the ligand. As described in the Ligand representation section, each atom in the molecular graph is represented as a node with a 40-dimensional initial feature vector. Let G = (V, E) denote the molecular graph of a ligand, where V is the set of atom nodes and E is the set of chemical bonds. Each atom v _i ∈V is associated with an initial feature vector ${\mathrm{v⃗}}_i^{\mathrm{init}}\in {\mathbb{R}}^{H_c}$ , where H _c = 40. To project these initial features into a higher-dimensional latent space, we apply a linear transformation as shown in eq , where H _c ′ is the hidden dimension.

$${\mathrm{v⃗}}_i=\mathbf{W}{\mathrm{v⃗}}_i^{\mathrm{init}},\mathrm{where}\mathbf{W}\in {\mathbb{R}}^{H_c\prime \times H_c}$$ 5

The GAT computes attention weights to adaptively aggregate information from neighboring nodes. Specifically, the attention coefficient α _ij between atom i and neighbor j is defined as

$${\alpha }_{\mathit{ij}}=\frac{\exp (\mathrm{LeakyReLU}({\mathrm{a⃗}}^{\tau }[{\mathrm{v⃗}}_i\parallel {\mathrm{v⃗}}_j]))}{\sum _{k\in {\mathcal{N}}_i}\exp (\mathrm{LeakyReLU}({\mathrm{a⃗}}^{\tau }[{\mathrm{v⃗}}_i\parallel {\mathrm{v⃗}}_k]))}$$ 6

In eq , ${\mathcal{N}}_i$ denotes the neighbors of node i, a⃗ is a learnable attention vector, and ∥ represents vector concatenation. Here, the index k in the denominator serves as a summation index iterating over all neighbors $k\in {\mathcal{N}}_i$ of node i for normalization. Using these coefficients, each atom updates its embedding by aggregating its neighbors’ features

$${\mathrm{v⃗}}_i^{\prime }=\sigma \left(\sum _{j\in {\mathcal{N}}_i}{\alpha }_{\mathit{ij}}{\mathrm{v⃗}}_j\right)$$ 7

The resulting vector v⃗ _i in eq represents the updated embedding for atom i, computed by aggregating the features of its neighbors, weighted by the learned attention coefficients. In addition, we apply a multihead attention to capture diverse aspects of neighborhood interactions and improve the expressive capacity of node embeddings.

$${\mathrm{v⃗}}_i^{\mathrm{cat}}={\parallel }_{k=1}^K\sigma \left(\sum _{j\in {\mathcal{N}}_i}{\alpha }_{\mathit{ij}}^{(k)}{\mathrm{v⃗}}_j^{(k)}\right)$$ 8

$${\mathrm{v⃗}}_i^{\mathrm{final}}=\mathrm{LeakyReLU}({\mathbf{W}}_{\mathrm{out}}{\mathrm{v⃗}}_i^{\mathrm{cat}}),{\mathbf{W}}_{\mathrm{out}}\in {\mathbb{R}}^{H_c\times K{H}_c\prime }$$ 9

In the above eqs and , K denotes the number of attention heads, v⃗ _j is the transformed feature of atom j under head k, and α _ij is the corresponding attention weight. The operator ∥ denotes concatenation across heads, and v⃗ _i is the resulting vector. The final embedding v⃗ _i is obtained by applying a shared linear projection and LeakyReLU activation. Through these equations, each atom aggregates information from its neighbors via multihead attention, producing embeddings that reflect multiple relational perspectives. This process is repeated across several GAT layers. Finally, the ligand encoder constructs the ligand-level embedding v⃗ _ligand by aggregating the resulting atom-level embeddings through graph-level mean pooling.

$${\mathrm{v⃗}}_{\mathrm{ligand}}=\frac{1}{|V|}\sum _{i=1}^{|V|}{\mathrm{v⃗}}_i^{\mathrm{final}}$$ 10

Equation shows how the ligand-level embedding is computed as the average of all atom-level embeddings in the graph.

Residue-Ligand Interaction

The Residue–Ligand Interaction module is designed to enable mutual information exchange between the residues in the protein and the ligand through cross-attention, and to construct residue–ligand interaction embeddings by combining the updated residue and ligand embeddings. To achieve this, we apply two cross attention mechanisms: Residue-to-Ligand (R2L) and Ligand-to-Residue (L2R) attention.

The Residue-to-Ligand (R2L) attention mechanism incorporates ligand embeddings into each residue embedding, enabling the model to reflect how each residue interacts with the given ligand. Specifically, for each residue x _i , the embedding v⃗ _residue, i , obtained from the residue encoder, serves as the query. The ligand embedding v⃗ _ligand is used as both the key and value. The R2L attention mechanism is formulated in the following eq .

$${\widetilde{\mathit{v⃗}}}_{\mathrm{residue},i}=\mathrm{MultiHeadAttn}({\mathrm{v⃗}}_{\mathrm{residue},i},{\mathrm{v⃗}}_{\mathrm{ligand}},{\mathrm{v⃗}}_{\mathrm{ligand}})$$ 11

In eq , the MultiHeadAttn denotes a 4-head attention mechanism. The resulting vector ${\widetilde{\mathit{v⃗}}}_{\mathrm{residue},i}$ is added to the residue embedding v⃗ _residue, i through a residual connection to produce ligand-aware residue embeddings.

$${\mathit{v⃗}}_{\mathrm{residue},i}^{\prime }={\mathit{v⃗}}_{\mathrm{residue},i}+{\widetilde{\mathrm{v⃗}}}_{\mathrm{residue},i}$$ 12

The ligand-aware residue embeddings v⃗ _residue are defined in eq .

The Ligand-to-Residue (L2R) attention mechanism, in contrast to R2L, incorporates residue-level context into the ligand embedding, enabling the model to reflect how the ligand attends to the protein. While the R2L attention uses residue embeddings as the query and ligand embeddings as the key and value, L2R reverses thisthe ligand embedding serves as the query, and the residue embeddings are used as both the key and value. Let R = { v⃗ _residue, i } _i = 1 denote the residue embedding set. The L2R attention is formulated as

$${\widetilde{\mathit{v⃗}}}_{\mathrm{ligand}}=\mathrm{MultiHeadAttn}({\mathrm{v⃗}}_{\mathrm{ligand}},R,R)$$ 13

In eq , the resulting vector ${\widetilde{\mathit{v⃗}}}_{\mathrm{residue},i}$ is added to the ligand embedding v⃗ _ligand through a residual connection

$${\mathrm{v⃗}}_{\mathrm{ligand}}^{\prime }={\mathrm{v⃗}}_{\mathrm{ligand}}+{\widetilde{\mathrm{v⃗}}}_{\mathrm{ligand}}$$ 14

This produces the residue-aware ligand embedding v⃗ _ligand , as defined in eq which integrates information from all residues and is used for interaction modeling. Finally, to construct the residue–ligand interaction embeddings, we performed element-wise multiplication between each ligand-aware residue embedding and the residue-aware ligand embedding

$$I_{\mathrm{res}-\mathrm{lig},i}={\mathrm{v⃗}}_{\mathrm{residue},i}^{\prime }\odot {\mathrm{v⃗}}_{\mathrm{ligand}}^{\prime }$$ 15

In eq , ⊙ denotes element-wise multiplication. The resulting vector I _{res–lig, i} represents the interaction embedding between residue x _i and the ligand, and serves as input to the final binding residue prediction module.

Ligand-Binding Residue Prediction

The residue-ligand interaction embedding I _{res–lig, i} obtained from the Residue–Ligand Interaction module is used as input to the binding residue predictor. This predictor consists of three feedforward layers and estimates the binding probability between each residue in the protein sequence and the ligand. Based on this probability, each residue is classified as either binding (1) or nonbinding (0).

Model Training Configuration

Loss Function

Our model optimizes a weighted combination of two objectives: (1) a weighted binary cross-entropy (BCE) loss for classifying whether a residue is ligand-binding or not, and (2) a contrastive loss , that aims to maximize the distinction between the interaction embeddings of binding and nonbinding residues. The BCE loss trains the model to predict the binding probability of each residue, with a higher weight assigned to the binding class to address the strong imbalance between binding and nonbinding residues. Meanwhile, the contrastive loss is employed to learn discriminative residue–ligand interaction embeddings, encouraging embeddings of binding residues (positive pairs) to cluster closely while pushing those of nonbinding residues (negative pairs) apart in the embedding space. Specifically, the contrastive objective follows the normalized temperature-scaled cross-entropy (NT-Xent) formulation introduced in SimCLR, defined as

$${\mathcal{L}}_{\mathrm{con}}=-\frac{1}{N}\sum _{i=1}^N\log \frac{\sum _{j\ne i}{\mathbf{1}}_{[y_i=y_j]}\exp (\mathrm{sim}(z_i,z_j)/\tau )}{\sum _{k\ne i}\exp (\mathrm{sim}(z_i,z_k)/\tau )}$$ 16

where N denotes the number of residue–ligand interaction pairs in a mini-batch, z _i and z _j are the corresponding embeddings, and y _i , y _j ∈{0, 1} indicate whether each residue is ligand-binding. The indicator function 1 _{[y i  = y j ]} equals 1 if the two samples belong to the same class (positive pair) and 0 otherwise. sim­(z _i , z _j ) represents the cosine similarity, and τ is a temperature parameter controlling the sharpness of the similarity distribution. Therefore, the total loss is defined as

$$\mathcal{L}={\mathcal{L}}_{\mathrm{cls}}+{\lambda }_{\mathrm{con}}·{\mathcal{L}}_{\mathrm{con}}$$ 17

where ${\mathcal{L}}_{\mathrm{cls}}$ is the weighted binary cross-entropy loss and ${\mathcal{L}}_{\mathrm{con}}$ is the contrastive loss computed over residue–ligand interaction embeddings. The hyperparameter λ_con controls the influence of the contrastive objective. This joint loss improves robustness to class imbalance by applying weighted supervision, and helps the model learn interaction embeddings that clearly distinguish between binding and nonbinding residues.

Training Procedure

The model was trained with a batch size of 32 and evaluated with a batch size of 64. We used the AdamW optimizer with a learning rate of 1 × 10^–4 and a weight decay of 1 × 10^–5. A cosine annealing scheduler was applied to gradually reduce the learning rate over epochs. A dropout rate of 0.1 was applied during training to prevent overfitting. All training and inference were conducted on a single NVIDIA A100 GPU.

Evaluation Metrics

To evaluate models for ligand-binding residue prediction, where binding residues are much fewer than nonbinding ones, we used Precision, Recall, F1-score, and Matthews Correlation Coefficient (MCC). The metrics are defined as follows

$$\mathrm{Precision}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}$$ 18

$$\mathrm{Recall}(\mathrm{Sensitivity})=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}$$ 19

$$\mathrm{F}1-\mathrm{score}=\frac{2·\mathrm{Precision}·\mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}$$ 20

$$\mathrm{MCC}=\frac{\mathrm{TP}·\mathrm{TN}-\mathrm{FP}·\mathrm{FN}}{\sqrt{(\mathrm{TP}+\mathrm{FP})(\mathrm{TP}+\mathrm{FN})(\mathrm{TN}+\mathrm{FP})(\mathrm{TN}+\mathrm{FN})}}$$ 21

In all equations above, TP is the number of binding residues correctly predicted as binding, FP is the number of nonbinding residues incorrectly predicted as binding, TN is the number of nonbinding residues correctly predicted, and FN is binding residues that are incorrectly predicted as nonbinding. Also, to define ligand-binding residues from the model outputs, a classification threshold is required. Therefore, the classification threshold was determined using 5-fold cross-validation on the training set. Specifically, threshold values ranging from 0.1 to 0.9 with a step size of 0.05 were evaluated, and the threshold that maximized the average F2-score on the validation folds was selected. Based on this procedure, a threshold of 0.6 was chosen and fixed for all subsequent experiments.

Results and Discussion

Performance of LiBRe

We evaluated our LiBRe model relative to other major models with two test data sets (see Table ). All metrics are measured at the best model state, averaging the results over five different seeds (note that the results of structure-based methods such as Fpocket and P2Rank are reported as single values without standard deviation, as they are provided as executable tools that return identical results regardless of random seed settings). LiBRe outperformed all baseline methods across both test sets in all evaluation metrics except Recall. In terms of Recall, LiBRe ranked second across both test sets. Pseq2sites achieved the highest Recall, due to its tendency to broadly classify residues as binding. While this leads to a higher number of true positives, it also leads to a large number of false positives, resulting in low Precision and Accuracy. In addition, we analyzed ROC-AUC and PR-AUC for sequence-based methods. Since binding residues account for only a small fraction of all residues, these metrics are more informative for evaluating model performance under class imbalance. As shown in Figure , LiBRe achieved the highest ROC-AUC and PR-AUC on both COACH420 and HOLO4K data sets. These results demonstrate that LiBRe provides greater robustness and more balanced performance than both structure-based and existing sequence-based methods, particularly under the class imbalance between binding and nonbinding residues.

3. Performance Comparison between LiBRe and the Baselines on Two Test Sets .

test set	method	MCC	F1-score	precision	recall	accuracy
COACH420	Fpocket	0.0314	0.1261	0.0729	0.4652	0.5986
	P2Rank	0.1371	0.1795	0.1106	0.4758	0.7636
	HoTS	0.0532 ± 0.007	0.1357 ± 0.003	0.0763 ± 0.002	0.6137 ± 0.013	0.5027 ± 0.002
	BiRDS	0.3754 ± 0.091	0.4190 ± 0.010	0.3158 ± 0.021	0.6227 ± 0.001	0.8825 ± 0.016
	Pseq2sites	0.2360 ± 0.051	0.2316 ± 0.079	0.1338 ± 0.051	0.9517 ± 0.040	0.5310 ± 0.260
	LiBRe	0.5610 ± 0.009	0.5647 ± 0.013	0.4269 ± 0.018	0.8352 ± 0.015	0.9179 ± 0.006
HOLO4K	Fpocket	0.0231	0.1182	0.0683	0.4365	0.5997
	P2Rank	0.1478	0.1788	0.1074	0.5333	0.7466
	HoTS	0.0662 ± 0.003	0.1375 ± 0.001	0.0770 ± 0.001	0.6402 ± 0.007	0.5064 ± 0.001
	BiRDS	0.3947 ± 0.005	0.4338 ± 0.052	0.3276 ± 0.003	0.6419 ± 0.011	0.8937 ± 0.016
	Pseq2sites	0.2421 ± 0.128	0.2346 ± 0.085	0.1363 ± 0.055	0.9454 ± 0.037	0.5469 ± 0.269
	LiBRe	0.6203 ± 0.010	0.6225 ± 0.012	0.4945 ± 0.016	0.8516 ± 0.005	0.9373 ± 0.003

^aThe best results are highlighted in bold, while the second-best results are underlined.

2.

Comparison of AUC-ROC and PR-AUC of sequence-based ligand-binding residue prediction models on two benchmark data sets. (A) Shows ROC (left) and precision–recall (PR) curves (right) on the COACH420 data set. (B) Shows ROC (left) and PR curves (right) on the HOLO4K data set.

Residue-Ligand Interaction Embeddings Comparison Using UMAP

We visualized the residue-ligand interaction embeddings to evaluate the ligand-awareness of sequence-based models. Specifically, we analyzed whether the interaction embeddings differ when different ligands bind to the same protein sequence. To evaluate this, we selected samples from each test set in which the same protein sequence binds to multiple ligands and compared the UMAP visualizations of the interaction embeddings with those from the sequence-based models HoTS and Pseq2sites. Figure A,D shows the residue-ligand interaction embeddings produced by our model, including both binding and nonbinding residues for each ligand. Notably, the interaction embeddings vary depending on the bound ligand, despite identical protein sequences. Also, the embeddings of binding and nonbinding residues form clearly separated clusters. In contrast, the embeddings produced by Pseq2sites, shown in Figure B,E, and those produced by HoTS, shown in (C) and (F), do not form clearly separated clusters between binding and nonbinding residues and do not exhibit such ligand-aware patterns. Since these models produce the same residue-ligand interaction embeddings even when different ligands bind to the same protein sequence, we observe overlap in the embedding points for binding and nonbinding residues across different ligands in their UMAP visualizations. These results demonstrate that our model effectively recognizes ligand information and produces more discriminative representations between binding and nonbinding residues than existing sequence-based models. Additional UMAP visualizations for further samples from each test data set beyond those shown here are provided in the Supporting Information (Figures S1 and S2).

3.

UMAP visualizations of ligand-binding residue embeddings constructed by LiBRe, Pseq2sites, and HoTS. (A–C) Correspond to PDB ID 1ixn (chain A) in complex with ligands DXP and G3P from the COACH420 data set, and (D–F) correspond to PDB ID 2ukd (chain A) in complex with ligands ADP and C5P from the HOLO4K data set. Each point represents an individual residue, and colors indicate whether the residue is binding or nonbinding for each ligand. A clearer separation between binding and nonbinding residues for each ligand in the embedding space indicates that the model can better encode ligand-related residue representations.

Binding Affinity Analysis Based on Predicted Binding Residues

We performed docking-based binding affinity evaluation to assess the practical utility of binding residue prediction models in virtual screening. , To perform this evaluation, we used the ligand-binding residues predicted by our model, Pseq2sites, and P2Rank. AutoDock Vina was then used to estimate binding affinities, with the docking center defined by the geometric center of the predicted ligand-binding residues. The search space was set to a grid box of 20 Å × 20 Å × 20 Å centered at this point, and the exhaustiveness was set to 8. For each sample in the test sets, docking was performed with five random seeds, and the mean of the best affinity values was used for analysis. The resulting affinity values were aggregated over all samples, and their distributions were visualized using kernel density estimation (KDE) plots for both the COACH420 and HOLO4K test sets.

As shown in Figure , our model produced lower average binding affinity values compared to Pseq2sites and P2Rank on all test sets. This implies that our model provides more accurate binding centers for guiding ligand docking. Our model also showed the lowest standard deviation in affinity values on both test sets, indicating that it produces more stable and reliable results across a wide range of ligand types. These results demonstrate that the binding centers derived from our model’s predicted residues provide more accurate and reliable spatial guidance for virtual screening of protein–ligand interactions.

4.

Binding affinity distributions of docking results using predicted binding residues from different models on the COACH420 and HOLO4K test sets. Each subplot shows the kernel density estimate of binding affinity values (x-axis, in kcal/mol) across all protein–ligand samples in each data set, with density shown on the y-axis. The red dashed line indicates the mean binding affinity (μ), and the standard deviation (σ) is also provided to assess the spread of affinity values.

Comparison of Known and Predicted Binding Pockets Using PyMOL

To evaluate how binding pockets formed from predicted binding residues differ across methods and how closely they resemble pockets defined by known binding residues, we visualized four protein–ligand complexes from the test set using PyMOL. In Figure , each visualization shows a protein–ligand complex in which binding pockets formed from known or predicted binding residues are color-coded. Binding pockets based on known binding residues are shown in blue, while those based on predicted binding residues are shown in red.

5.

Comparison of known and predicted binding pockets for four protein–ligand complexes. The top two complexes (1 kgz–PRP and 3exs–5RP) are from the COACH420 data set, while the bottom two complexes (121P–GCP and 2fax–FMN) from the HOLO4K data set. Binding pockets shown to the left of the dashed line represent ground truth pockets derived from known binding residues (blue). Binding pockets to the right of the dashed line are predicted pockets (red), derived from residues predicted by LiBRe, Pseq2sites, and P2rank, respectively. Ligands are shown in green.

As illustrated by the visualizations, Pseq2sites tends to overpredict binding residues, leading to overly broad binding pocket regions. Furthermore, both Pseq2sites and P2Rank sometimes predict binding residues in regions that are structurally implausible for ligand binding. In contrast, although our model does not always predict binding residues that exactly coincide with the known sites, many of the incorrectly predicted residues are positioned close to the ligand. These results indicate that, compared to existing methods, our model effectively captures residues located near the ligand without structural information.

Ablation Study on Model Components

Our model applies several modeling techniques to enhance the prediction of ligand-binding residues, including the integration of ligand information and the contextual refinement of residue embeddings through a combination of 1D-CNN and BiLSTM. To examine how each of these techniques affects the performance of our model, we conducted ablation experiments.

The results of the ablation study are summarized in Table . Removing ligand information degraded performance across both COACH420 and HOLO4K benchmarks, confirming that incorporating ligand features improves the identification of binding residues. Similarly, excluding the 1D-CNN/BiLSTM module, which refines residue embeddings with local and sequential context, led to lower F1 and MCC scores, highlighting the importance of contextual refinement. In addition, contrastive learning also contributed to performance gains. Overall, the full model achieved the best results, demonstrating that ligand integration, residue-level contextual refinement, and contrastive learning collectively enhance binding residue prediction.

4. Ablation Study Results on the COACH420 and HOLO4K Benchmarks to Assess the Contribution of Each Model Component .

	COACH420			HOLO4K
model	F1-score	MCC	accuracy	F1-score	MCC	accuracy
w/o ligand, w/o 1D-CNN/BiLSTM	0.4439 ± 0.005	0.4627 ± 0.005	0.8582 ± 0.002	0.4554 ± 0.003	0.4821 ± 0.003	0.8643 ± 0.001
w/o ligand	0.4804 ± 0.016	0.4841 ± 0.012	0.8851 ± 0.009	0.4918 ± 0.009	0.5054 ± 0.008	0.8878 ± 0.003
w/o 1D-CNN/BiLSTM	0.4596 ± 0.005	0.4827 ± 0.006	0.8626 ± 0.002	0.5073 ± 0.005	0.5309 ± 0.004	0.8883 ± 0.005
w/o contrastive learning	0.5141 ± 0.005	0.5227 ± 0.004	0.8948 ± 0.002	0.5570 ± 0.005	0.5691 ± 0.004	0.9112 ± 0.002
full model	0.5647 ± 0.013	0.5610 ± 0.009	0.9179 ± 0.006	0.6225 ± 0.012	0.6203 ± 0.010	0.9373 ± 0.003

^aF1-score, MCC, and Accuracy scores were obtained from experiments conducted with five different seeds, and the mean and standard deviation are presented. Each row corresponds to a model variant with one or more components removed. The best performance is highlighted in bold.

Performance with Different Protein Language Models

Recently, a variety of protein language models have emerged and have been widely applied to diverse downstream tasks. Our model also employs ESM-2 to construct initial embeddings for individual residues from protein sequences. Therefore, we performed an analysis to examine how the performance of our model changes when different protein language models are used. To this end, in addition to the ESM-2 650 M model used in our main experiments, we conducted additional experiments using three alternative protein language models: ESM-C 300 M and ESM-C 600M, and ProtT5-XL (prot_t5_xl_half_uniref50-enc) from the ProtTrans.

Table summarizes the results obtained by replacing the protein language model originally used in our architecture (ESM-2) with alternative protein language models. All evaluated protein language models exhibit comparable performance across F1, MCC, and Accuracy on both COACH420 and HOLO4K data sets. Although the performance differences are modest, models with higher embedding dimensionality tend to show slightly better performance overall, which can be attributed to the richer residue representations provided by higher-dimensional embeddings. Nevertheless, the performance differences are not dramatic, indicating that our framework does not strongly depend on a specific protein language model. As a result, different protein language models can be flexibly selected based on practical considerations, such as computational efficiency, while maintaining competitive performance.

5. Performance Comparison Across Different Protein Language Models .

		COACH420			HOLO4K
model	emb. dim	F1-score	MCC	accuracy	F1-score	MCC	accuracy
ProtT5-XL	1024	0.5599 ± 0.008	0.5571 ± 0.008	0.9161 ± 0.003	0.6105 ± 0.006	0.6057 ± 0.006	0.9336 ± 0.002
ESM-C 300M	960	0.5549 ± 0.008	0.5487 ± 0.007	0.9166 ± 0.003	0.6199 ± 0.006	0.6135 ± 0.003	0.9355 ± 0.002
ESM-C 600M	1152	0.5633 ± 0.009	0.5579 ± 0.008	0.9186 ± 0.003	0.6184 ± 0.012	0.6134 ± 0.010	0.9356 ± 0.003
ESM-2 650M	1280	0.5647 ± 0.013	0.5610 ± 0.009	0.9179 ± 0.006	0.6225 ± 0.012	0.6203 ± 0.010	0.9373 ± 0.003

^aF1-score, MCC, and Accuracy scores are reported as the mean and standard deviation over five seeds.

Case Study of Virtual Screening Application

We conducted a case study to assess the practical utility of our model in virtual screening for the EGFR target (PDB ID: 2RGP), using the DUD-E benchmark data set. The data set includes 542 active ligands and 35,050 decoys. Active compounds refer to molecules that have been experimentally confirmed to bind to the target protein, while decoys are molecules with similar physicochemical properties but dissimilar topologies that are assumed, though not experimentally validated, to be nonbinders and are widely used to assess the specificity and robustness of virtual screening methods. , All actives were used, and following common practice in virtual screening benchmarks, we randomly sampled decoys at a 1:50 active-to-decoy ratio using five different random seeds to obtain five independent replicates. For each active and decoy ligand, our model predicted the ligand-binding residues of the EGFR target, and the geometric center of the predicted residues was then defined as the docking pocket center. Docking was performed using AutoDock Vina with a 20 Å × 20 Å × 20 Å search box centered at the predicted pocket. For each ligand, only the best binding affinity among the generated poses was used, and the resulting scores were used to rank both active and decoy ligands. Based on the docking scores, we evaluated the virtual screening performance of our model using the enrichment factor (EF) and the area under the receiver operating characteristic curve (ROC-AUC). , The enrichment factor (EF) measures how much more frequently active compounds appear in the top k percent of the ranked ligand list based on predicted binding affinities, compared to their proportion in the overall data set. We used EF@1%, EF@5%, and EF@10%. EF values greater than 1 indicate that the model is able to effectively select active compounds. Values close to 1 suggest no particular screening ability. As shown in Table , virtual screening using our model predicted binding centers achieved EF values of 1.49, 2.44, and 3.16 at the top 1, 5, and 10%, respectively. These results demonstrate that our model can reduce a large initial compound library to a smaller subset enriched with promising candidates in the early stage of screening. In addition, we analyzed the distribution of affinity scores for actives and decoys. Figure illustrates the distribution of binding affinity scores between the target protein and both actives and decoys. Active ligands show stronger average binding affinities, while the affinities of decoys are mostly concentrated around 0. This observation, along with a ROC-AUC of 0.8242, indicates that the binding centers predicted by our model help distinguish between actives and decoys in virtual screening. Collectively, these findings highlight the potential of our model to accelerate virtual screening workflows and improve hit discovery efficiency.

6. Virtual Screening Results for EGFR from the DUD-E Dataset Using Our Model’s Predicted Ligand-Binding Residues .

metric	EF@1%	EF@5%	EF@10%	ROC-AUC
our model	1.49	2.44	3.16	0.8242

^aEF@1%, EF@5%, and EF@10% represent the enrichment factors at the corresponding top-ranked percentages, indicating how much more likely active compounds are retrieved compared to random selection. A higher EF indicates that more active compounds are ranked near the top among all screened molecules.

6.

Binding affinity distribution of active and decoy compounds based on docking results using the predicted binding residues. The plot shows kernel density estimates for actives (blue) and decoys (red).

Conclusions

In this study, we proposed LiBRe, a ligand-aware sequence-based deep learning model for predicting protein–ligand binding residues. Unlike existing sequence-based methods that rely only on protein sequence information, LiBRe overcomes the limitations of ligand-unaware models by explicitly incorporating ligand information. Although binding residue prediction is a highly imbalanced classification task, our model demonstrates robustness by achieving higher performance than baseline methods across all evaluation metrics except Recall on the benchmark data sets. Embedding visualizations further indicate that our model is ligand-aware, as interaction embeddings vary depending on the bound ligand even for identical protein sequences. Additionally, structural visualizations reveal that the predicted binding residues form pockets that closely resemble known binding pockets, suggesting that our model can accurately identify ligand-binding regions even without structural information. Based on these findings, we conducted docking-based binding affinity evaluations using the predicted binding centers, and observed that our model enables stronger and more stable affinity predictions compared to baseline methods. To assess the practical applicability of our model, we conducted a case study on the EGFR target from the DUD-E database. Using the predicted binding residues from our model as docking centers, our model successfully prioritized active compounds over decoys in virtual screening, achieving meaningful enrichment factors and a high ROC-AUC. These results confirm that LiBRe is not only effective for binding residue prediction but also readily applicable to structure-free virtual screening pipelines.

While LiBRe achieves promising results, it still leaves room for further improvement. Currently, the model does not leverage structural information, which may limit its ability to capture detailed spatial interactions. Additionally, LiBRe is designed for single protein–ligand binding scenarios, which restricts its ability to model complex interactions such as multiple ligand binding events. The proposed framework could also be extended toward protein-type-aware ligand-binding residue prediction, enabling more specialized modeling of diverse binding patterns across different classes of proteins. Future work may address these limitations by developing a hybrid architecture that leverages both sequence and structural information, and by incorporating data reflecting multiple ligand binding scenarios. These advancements are expected to further improve virtual screening performance and contribute to the acceleration of drug discovery.

The data sets used in this study are publicly available. Protein–ligand complex data sets were obtained from the sc-PDB (v.2017) (http://bioinfo-pharma.u-strasbg.fr/scPDB/), the PDBbind­(v.2020) (https://www.pdbbind-plus.org.cn/), and benchmark sets COACH420 and HOLO4K (available at https://github.com/rdk/p2rank-datasets), as referenced in the manuscript. Virtual screening experiments employed the DUD-E data set (http://dude.docking.org). All source code for the LiBRe is available at https://github.com/GoldRiver0/LiBRe.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.jcim.5c02883.

• UMAP visualizations of ligand-binding residue embeddings constructed by LiBRe, Pseq2Sites, and HoTS on the COACH420 dataset (Figure S1); UMAP visualizations of ligand-binding residue embeddings constructed by LiBRe, Pseq2Sites, and HoTS on the HOLO4K dataset (Figure S2) (PDF)

K.K. and G.S. designed the work. K.K. developed the deep learning algorithm and associated tools, conducted the experiments, performed the analyses, and drafted the manuscript. M.K. contributed to the experiments. J.K. and S.S. contributed to the analyses and manuscript writing. G.S. initiated, supervised, and coordinated the project and revised the manuscript. All authors reviewed and approved the final manuscript.

The authors declare no competing financial interest.