3DICE is proposed, a novel framework leveraging co-attention-based fusion and massively pre-trained 3D structural encoders for both drugs and proteins that provides intrinsic interpretability, highlighting and enabling qualitative analysis of most influential atoms or residues.
Abstract
Abstract Motivation Drug–target interaction (DTI) prediction is a crucial step in modern drug discovery. Accurate and efficient predictions can substantially reduce costs and development time. Applications of deep learning methods for this purpose have been extensively studied in recent years, yielding instrumental contributions to this field. However, existing methods face issues pertaining to efficient learning of drug and target feature representations, which is detrimental to generalizability and performance in cold-start scenarios. Most approaches extract representations from SMILES strings for drugs and FASTA sequences for target proteins, which encode limited 3D structural information. Additionally, many models lack explainability, being black boxes that provide little physical insight into the underlying mechanisms behind such interactions. Results We propose 3DICE, a novel framework leveraging co-attention-based fusion and massively pre-trained 3D structural encoders for both drugs and proteins. Uni-Mol and ESM-IF1 are employed to generate high-fidelity, 3D structure-aware embeddings which enable richer geometric and chemical understanding. Cross-modal fusion modules further augment representations to model intermolecular binding relationships. Importantly, this mechanism also provides intrinsic interpretability, highlighting and enabling qualitative analysis of most influential atoms or residues. Experiments conducted on two canonical benchmark datasets display the competitiveness of our model in real-world scenarios. 3DICE outperformed state-of-the-art models across multiple metrics on the DrugBank and KIBA datasets. Additional experiments provide a more rigorous analysis of interpretability than is typically reported in prior DTI studies, and we find that attention consistently highlights decision-critical regions which is not intrinsically class-specific. Availability Our model and dataset are freely available at: https://github.com/austinatose/3DICE.
Predicting the toxicity of pharmaceutical compounds remains a major challenge in drug discovery. Early and accurate toxicity assessment is essential for eliminating harmful candidates before costly preclinical and clinical testing, thereby improving patient safety, reducing development costs, and accelerating the drug development process. Despite advances in computational toxicology, existing methods often struggle to capture complex molecular characteristics and maintain robust performance under limited-data conditions. To address these challenges, we propose 3Br-MGD, a novel three-branch framework that integrates deep learning and meta-learning for molecular toxicity prediction. The architecture combines complementary molecular representations: FingerprintMLP encodes Morgan fingerprint descriptors, Graph Convolutional Networks (GCNs) capture structural information from molecular graphs, and one-dimensional Deep Convolutional Neural Networks (1D-CNNs) extract sequential features from SMILES strings. These embeddings are integrated within a Prototypical Network-based few-shot learning framework, enabling rapid adaptation to new prediction tasks with limited labeled samples and improving generalization in low-resource settings. Experimental results on benchmark toxicity datasets demonstrate that 3Br-MGD consistently outperforms conventional baselines in predictive accuracy, robustness, and generalization. Furthermore, the integration of heterogeneous molecular encoders reduces dependence on large training datasets while enhancing interpretability through the exploitation of complementary chemical information from multiple molecular views.
This paper proposes TextDTI, a multimodal framework that simultaneously exploits sequential and structural representations and enhances feature alignment through adversarial learning and contrastive loss, resulting in robust and high-performance DTI prediction.
Jiaqi Deng, Senyu Tang, Jijun Tang et al.· Journal of Chemical Informat...· 0 citations
Prediction of Drug Target Affinity (DTA) is essential for accelerating computational drug discovery and reducing experimental costs. However, traditional experimental approaches for DTA estimation are resource-intensive and are further challenged by the structural flexibility of both drugs and target proteins. In this work, we propose the PCBERT-GAT-DFFNN-DTA model, a three-stage deep cross-modal representation fusion framework for accurate DTA prediction. In the first stage, variable-length protein sequences are transformed into contextual representations using ProtBERT to obtain fixed-size protein embeddings. Drug molecules are represented using two modalities: sequence-based embeddings generated from ChemBERT and structure-based embeddings learned from molecular graphs using a Graph Attention Network (GAT). In the second stage, each modality is processed through dedicated subnetworks to refine features and reduce dimensionality while preserving modality-specific information. In the final stage, the refined representations are fused and passed to a Deep Feed-Forward Neural Network (DFFNN) to predict drug target binding affinity. The proposed model consistently outperformed most baseline methods under the S1-S3 evaluation settings across the benchmark datasets. Under the more challenging S4 blind setting, the model achieved strong performance on the KIBA dataset and competitive results on the Davis and Metz datasets. Compared with LLMDTA, the proposed approach achieves significant improvements in R2 scores across all datasets, demonstrating its effectiveness in learning complex drug protein interactions for reliable DTA prediction.
Essmily Simon, Sanjay S. Bankapur· Analytical Biochemistry· 0 citations
Accurately predicting binding affinities between drugs and targets is crucial for drug discovery but remains challenging due to the complexity of modeling interactions between small drug and large targets. This research presents Dual modality feature fused-drug target affinity (DMFF-DTA), a model for drug-target affinity anticipation using dual-modality neural networks that considers both the sequence and graph structure of medicines and proteins. To facilitate more exact and efficient drug-target interaction modeling, the model incorporates a binding site-focused graph generation method for extracting binding information. Experimental results show that DMFF-DTA is far more effective than current state-of-the-art approaches. By outperforming state-of-the-art approaches by more than 8%, the model demonstrates remarkable generalizability to hitherto unexplored medicines and targets. The model's biological relevance is confirmed by the model interpretability analysis. This paper presents a reliable and understandable method for improving computational drug discovery by integrating multi-view protein and drug properties.
Ghazala Sultan, J. Vincent, Ratna Sahaya et al.· International Conference Com...· 0 citations
Unexpected drug–drug interaction events (DDIEs) pose substantial clinical risks, yet many remain unannotated due to data scarcity and the rapid emergence of novel drug combinations. Conventional deep learning approaches struggle to generalize to these unseen interaction types and often lack interpretability under severe class imbalance. To address these challenges, we propose ZSCAN-DDIE, an interpretable zero-shot learning framework for DDIE prediction. The model integrates a biomedical pre-trained language model with an attention-based graph convolutional network (AGCN) to encode DDIE textual semantics and drug molecular structures, respectively. A cross-attention network (CAN) is introduced to align molecular substructures with pharmacological semantic components, enabling fine-grained cross-modal reasoning and improving interpretability at the substructure level. To mitigate modality bias and long-tailed distribution effects, we design a bimodal dynamic alignment (BDA) loss that combines hyperspherical embedding regularization with a stage-adaptive loss-switching mechanism. Experimental results under both conventional and generalized zero-shot settings demonstrate that ZSCAN-DDIE consistently outperforms state-of-the-art baselines across multiple evaluation metrics. The proposed framework not only enhances prediction accuracy for unseen DDIE categories but also provides biologically meaningful insights into molecular interaction mechanisms, offering a robust and clinically relevant solution for pharmacovigilance and drug safety assessment
The source code and data are available at https://github.com/GSX-0429/ZSCAN-DDIE.
Shaoxi Gao, Zhanpeng Gan, Fangfang Han et al.· Bioinformatics Advances· 0 citations
Evaluations on eight MoleculeNet datasets show that MSMPP significantly outperforms state-of-the-art models, demonstrating its effectiveness in integrating multi-view intra-molecular features, inter-molecular features and cross-task information.
Jiongfeng Chen, Yulian Ding, Yan Yan et al.· IEEE journal of biomedical a...· 0 citations
