Integrating Chemical Language and Molecular Graph in Multimodal Fused Deep Learning for Drug Property Prediction (2312.17495v2)

Abstract: Accurately predicting molecular properties is a challenging but essential task in drug discovery. Recently, many mono-modal deep learning methods have been successfully applied to molecular property prediction. However, the inherent limitation of mono-modal learning arises from relying solely on one modality of molecular representation, which restricts a comprehensive understanding of drug molecules and hampers their resilience against data noise. To overcome the limitations, we construct multimodal deep learning models to cover different molecular representations. We convert drug molecules into three molecular representations, SMILES-encoded vectors, ECFP fingerprints, and molecular graphs. To process the modal information, Transformer-Encoder, bi-directional gated recurrent units (BiGRU), and graph convolutional network (GCN) are utilized for feature learning respectively, which can enhance the model capability to acquire complementary and naturally occurring bioinformatics information. We evaluated our triple-modal model on six molecule datasets. Different from bi-modal learning models, we adopt five fusion methods to capture the specific features and leverage the contribution of each modal information better. Compared with mono-modal models, our multimodal fused deep learning (MMFDL) models outperform single models in accuracy, reliability, and resistance capability against noise. Moreover, we demonstrate its generalization ability in the prediction of binding constants for protein-ligand complex molecules in the refined set of PDBbind. The advantage of the multimodal model lies in its ability to process diverse sources of data using proper models and suitable fusion methods, which would enhance the noise resistance of the model while obtaining data diversity.

Introduction to Multimodal Deep Learning

In the cutting-edge field of drug discovery, the accurate prediction of molecular properties is a crucial yet complex task. Traditional methods often center around mono-modal deep learning approaches, which utilize a singular form of molecular representation. These methods, while successful to a degree, are limited in their scope and can be impeded by data noise—a form of disruption that can affect the accuracy of predictions.

The Multimodal Approach

To enhance the capability of predictive models and curtail the impact of data noise, a paradigm shift towards multimodal deep learning is proposed. By integrating several forms of molecular representation—SMILES-encoded vectors, ECFP fingerprints, and molecular graphs—a more holistic view of drug molecules can be attained. The application of diverse deep learning techniques, namely Transformer-Encoder, BiGRU (bi-directional gated recurrent units), and GCN (graph convolutional network), each suited to process a specific type of molecular data, supports in capturing a wide array of naturally occurring bioinformatics characteristics.

Efficacy of Multimodal Models

Assessed on a variety of molecule datasets, these triple-modal models demonstrate superior performance as compared to mono-modal models. They exhibit greater accuracy, reliability, and a stronger defense against noise interference. The findings underscore the importance of simultaneous processing of diverse data sources, such as chemical structures and molecular graphs, using fitting models and fusion methods. There are five different fusion methods evaluated in this research, all of which lead to improvements in predictions. Yet, a fusion based on stochastic gradient descent (SGD) particularly stands out for its ability to assign optimal contributions across different modal information and consistently ensuring heightened prediction accuracy.

Applications and Future Scope

This multimodal approach holds significant promise for its application within pharmaceutical research and development. Implemented to predict a range of molecular properties, it also proves its worth on the refined set of PDBbind, demonstrating a strong generalization ability in protein-ligand complex binding constants prediction. Furthermore, multimodal deep learning models, especially those adopting an SGD-based fusion method, boast a favorable capacity to resist data noise, adding another layer of robustness to this novel predictive system. As the technology develops and scales up, the proposed methods could mark a new era in drug discovery, with far-reaching implications for medical research and treatment innovation.