MM-DRPNet: A multimodal dynamic radial partitioning network for enhanced protein–ligand binding affinity prediction Authors: Dayan Liu, Tao Song, Shudong Wang Published: 2024 ABSTRACT Accurate prediction of drug-target binding affinity remains a fundamental challenge in contemporary drug discovery. Despite significant advances in computational methods for protein-ligand binding affinity prediction, current approaches still face substantial limitations in prediction accuracy. Moreover, the prevalent methodologies often overlook critical three-dimensional (3D) structural information, thereby constraining their practical utility in computer-aided drug design (CADD). Here we present MM-DRPNet, a multimodal deep learning framework that enhances binding affinity prediction by integrating protein-ligand structural information with interaction features and physicochemical properties. The core innovation lies in our dynamic radial partitioning (DRP) algorithm, which adaptively segments 3D space based on complex-specific interaction patterns, surpassing traditional fixed partitioning methods in capturing spatial interactions. MM-DRPNet further incorporates molecular topological features to comprehensively model both structural and spatial relationships. Extensive evaluations on benchmark datasets demonstrate that MM-DRPNet significantly outperforms state-of-the-art methods across multiple metrics, with ablation studies confirming the substantial contribution of each architectural component. Source code for MM-DRPNet is freely available for download at https://github.com/Bigrock-dd/MMDRPv1. Introduction Accurately predicting the binding affinity between drugs and targets is a vital aspect of drug discovery [1]. The strength of the interaction between a drug molecule and its protein target directly affects its therapeutic effectiveness. Therefore, understanding and predicting these interactions is essential for the rational design of new medications. In recent years, various computational models have been developed to estimate protein-ligand binding affinity, offering the potential to streamline the drug discovery process by reducing the reliance on costly and time-consuming experimental tests [2], [3]. Despite these advances, many current models struggle to achieve high accuracy, mainly due to difficulties in effectively capturing the three-dimensional structural details of protein-ligand interactions [4]. The complex spatial arrangement of atoms within both proteins and ligands, along with their dynamic interaction patterns, plays a crucial role in determining binding affinity [5]. Modern drug discovery has increasingly incorporated computational tools, such as virtual screening of large compound libraries and analysis of ligand binding modes, as integral components of the development pipeline [6]. They not only provide ligands that may bind to specific targets, but also explain observed phenomena, such as the conformational relationships of compounds. Thus, developing a model that accurately captures the intricate 3D interactions between proteins and ligands remains a significant challenge. Although various research approaches have yielded encouraging results, they have focused mainly on single-modal data and have neglected the types of interactions between proteins and molecules with different biological functions. Recent research has realized the limitations of single modal information (e.g. protein sequence or structure) and has begun to combine sequence and structure information for analysis [7], [8], [9], [10]. However, many studies still ignore the heterogeneity between different modal information and the complexity of proteins and ligands in the process of specific binding [11][12]. To address these limitations, we propose MM-DRPNet, a novel multimodal framework that integrates three key innovations: (1) A Dynamic Radial Partitioning (DRP) method that adaptively captures spatial interaction features in protein-ligand complexes; (2) A comprehensive feature extraction strategy combining structural information, interaction features, and physicochemical properties; and (3) A multimodal fusion architecture that effectively aligns and integrates diverse types of molecular information. Our framework consists of three core modules: protein-ligand interaction feature extraction, molecular structural feature extraction, and multimodal feature fusion. The interaction module employs DRP to process 3D structural information and physicochemical interactions, while the structural module utilizes Graph Attention Networks [13] to capture molecular topology. These features are then optimally combined through our fusion module for accurate affinity prediction. Related work Molecular docking is one of the main approaches to structure-based drug discovery [14]. In molecular docking, the conformation of favorable ligand-protein interactions can be determined, producing so-called ligand poses (poses, binding modes) [15]. Their interactions with proteins can be quantified by scoring functions. To achieve efficiency in molecular docking, ligand-protein interactions would be quantified by a simple scoring function. However, this simplified model of protein-ligand interactions is a major limitation of the method, manifested in particular in the inaccuracy of the ranking of poses and the poor performance in predicting absolute or relative binding free energies. In addition, the correct sampling of ligand binding patterns may be limited by induced fit effects as well as by the different conformational states of the proteins (proteins are often treated as rigid bodies), which lead to inaccuracies that outweigh the inaccuracies introduced by the scoring function itself [16]. In addition, sometimes intramolecular interactions are incorrectly modelled, such as the presence of distorted amide groups, structural impulses between ligands within proteins, or unrealistic three-dimensional (3D) structures [17]. And another key factor affecting molecular docking is the molecular mass of the input structure [18]. Despite the widespread use of molecular docking, its simplified scoring functions and rigid protein assumptions limit prediction accuracy. Sequence-based protein-ligand interaction prediction methods leverage the amino acid sequence of proteins to infer binding affinity [19]. These methods typically extract relevant features from protein sequences, such as amino acid composition, physicochemical properties, and evolutionary information, to train machine learning or deep learning models [20], [21]. Early approaches, such as support vector machines (SVM) and random forest models, have demonstrated promise by utilizing these sequence-derived features to predict protein-ligand interactions [22], [23]. However, these models often struggle to capture the complex nonlinear relationships present in protein-ligand binding, limiting their predictive power [24][25][26]. While traditional machine learning methods showed initial promise, they struggled to capture complex nonlinear relationships in protein-ligand binding. With the advent of deep learning, more sophisticated models, including convolutional neural networks (CNNs) and recurrent neural networks (RNNs), have been employed to improve prediction accuracy by capturing higher-level sequence patterns and long-range dependencies within protein sequences [27], [28], [29], [30]. CNNs are particularly adept at extracting local interaction patterns from sequences, while RNNs and their variants, such as long short-term memory (LSTM) networks, excel at modeling sequential data by accounting for the temporal relationships between amino acids in a protein chain [31], [32], [33], [34]. Additionally, transformer-based models, which utilize self-attention mechanisms, have recently demonstrated superior performance in capturing long-range dependencies and complex interactions within protein sequences, further advancing the field of sequence-based prediction [35], [36], [37], [38]. Some examples include, DeepDTA leverages CNN to predict drug-target binding affinities by extracting features directly from protein sequences and compound representations [21]. GraphDTA utilizes graph neural networks to model molecular structures as graphs, improving upon previous sequence-based methods like DeepDTA by more accurately capturing the complex topological interactions within compounds for drug-target binding affinity prediction [39]. TransformerCPI applies Transformer architectures to capture long-range dependencies in both protein sequences and compound representations, offering improvements over graph-based models like GraphDTA by enhancing the ability to model complex interactions with greater contextual understanding for compound-protein interaction prediction [40]. There are also some models for the 3D structure of complexes, such as the OnionNet [41], a CNN based on element-specific contacts between proteins and ligands dependent on distance and Pafnucy [42], a 3D CNN that employs some computer vision-derived strategies to encode the protein and the ligand. There is also KDeep [43], which employs 3D CNN to predict protein-ligand binding affinity. All of these models are, to some extent, characterized by a vectorized lattice within a ligand-centered cube representing the protein-ligand complex, and they demonstrate good performance in predicting protein-ligand binding [44], [45], [46]. Although deep learning approaches have significantly improved prediction accuracy, most existing methods still rely on single-modal information, overlooking the inherent multimodal nature of protein-ligand interactions. Materials and methods Data preparation We utilized the PDBBind v2016 database [47] as our primary data source, which comprises high-quality three-dimensional structures and experimentally determined binding affinity data for 13,285 protein-ligand complexes. The database is hierarchically organized into the General Set, Refined Set, and Core Set (CASF-2013 [48] and CASF-2016 [49]). To ensure high-quality training data, we implemented rigorous filtering criteria on the General Set (9,228 complexes). Only X-ray crystal structures with resolution better than 2.5 Å were retained, and complexes with missing atoms or residues were excluded. After applying these quality control measures, 8,873 complexes qualified for our training dataset. To prevent data leakage and ensure unbiased evaluation, we carefully partitioned the datasets. The 8,873 filtered complexes from the General Set were used as the training set. For the validation set, we first removed all complexes that overlapped with the CASF-2016 and CASF-2013 sets from the Refined Set (4,057 complexes), then randomly selected 1,000 complexes. For testing, we utilized all 285 complexes from the CASF-2016 core set as our primary test set and 108 complexes from the CASF-2013 core set as our secondary test set, ensuring no overlap with the training and validation sets to maintain evaluation integrity. We established a systematic preprocessing pipeline to standardize all protein-ligand complexes. Initially, water molecules and non-essential ions were removed from the structures, and alternative conformations were eliminated. For ligand processing, structures in mol2 format were converted to PDB format, with chemical correctness verified and proper protonation states assigned at pH 7.4. The ligands were then docked with their corresponding receptor PDB files to maintain consistent binding poses. Following complex preparation, we determined the elemental type of each atom and calculated relevant atomic properties including partial charges and hydrophobicity. To standardize the measurement units, we transformed the binding affinity data into the negative logarithmic form according to:(1)pKa=−log10⁡Kx where Kx represents IC50, Ki, or Kd. The final step involved dynamic spatial distance feature extraction on the processed complexes. All preprocessing steps were automated using custom Python scripts to ensure reproducibility. To maintain consistency with previous research, no additional alterations were made to the protein-ligand complexes beyond these standard preprocessing steps. Dynamic racial partitioning To effectively model protein–ligand interactions, we use the atomic distances between the protein surface and the ligand as key descriptors. The 3D spatial arrangement of these atoms significantly influences binding affinity. To capture this spatial relationship, we divide the atomic pairs into multiple layers based on specific distance thresholds. Each layer includes pairs within a defined range, determined by cutoff values, which represent the maximum distances between protein and ligand atoms. By segmenting the interactions into these layers, we can analyze the interaction patterns across various spatial scales more effectively. R={ri}i=1Nr represent the set of Nr atoms in the protein and L={lj}j=1Nl represent the set of Nl atoms in the ligand. The spatial distance layer feature fs is then defined as the number of atomic pairs within each layer:(2)fs=∑i=1Nr∑j=1NlI(dmin≤dij