Artificial Intelligence and Network Medicine: Path to Precision Medicine

Abstract

Over the past two decades, network medicine (NM) has evolved to help define disease mechanisms, identify drug targets, and guide increasingly precise therapies. In recent years, the integration of NM with artificial intelligence (AI), particularly deep learning techniques, has evolved with increasing applications. AI techniques help elucidate complex disease mechanisms and define precise therapies. The depth of useful, mechanistic information implicit in molecular interaction networks and prior deep learning successes provide a rational basis for combining NM and AI in the analyses of large multiomic datasets to enhance the speed, predictive precision, and biological insights of the computational process. In this review, we provide a summary of concepts related to the combined use of AI and NM as a path to precision medicine, illustrating the success of this joint approach to biomedical complexity and its ongoing challenges.

Figure 1.

Biological and Biomedical Challenges Addressed through Integration of Network Medicine and Artificial Intelligence. Panel A shows examples of biomedical problems addressed by combining network medicine with artificial intelligence: disease-gene prediction; disease-module and endophenotype detection; patient-specific modules (reticulotypes); drug discovery and repurposing; drug -target and protein -protein interaction prediction; and side-effect modeling. Panel B summarizes representative approaches for drug -target interaction (DTI) prediction. Molecular compounds are encoded with descriptors such as isomeric Simplified Molecular Input Line Entry System (SMILES), extended-connectivity fingerprints (ECFP), and hydrophobicity indices; protein targets are represented by primary amino acid sequence, 3D structure, and protein-family features. Network medicine methods often embed drugs and targets into a bipartite/heterogeneous network topology, applying link-prediction strategies (e.g., maximum-entropy under constraints representing the available metadata; network-proximity metrics based on shortest paths and diffusion). Deep-learning methods include knowledge-graph embeddings (KGE), which encode biomedical entities and relations into latent vector spaces to support link prediction, and graph neural networks (GNNs), which learn from graph-structured data by aggregating features from neighboring nodes, such as atoms in a molecular graph or nodes in a biological network. For example, the ERT-GFAN model integrates structural features of drugs and targets obtained using ECFP, interaction features from the RotatE KGE model, and contextual neighborhood features refined by a Transformer. These multimodal fusion features are then integrated into a graphical high-dimensional fusion feature attention network for DTI prediction. In GraphDTA, drugs are represented as molecular graphs and encoded using one of four GNN variants: GCN, GAT, GIN, or a GAT-GCN hybrid. Protein targets are represented as sequences and encoded with 1D convolutional layers. MLC-DTA is a drug -target affinity prediction model that applies equivariant GNNs to capture 3D structural features of drugs and targets, integrates network-level relationships, and employs multi-level contrastive learning. MolTrans decomposes drugs and proteins into sub-structural sequences, encodes them with an augmented transformer embedding module, constructs a pairwise interaction map refined by CNNs, and outputs DTI predictions via a decoder module. Multi-TransDTI is an end-to-end DTI prediction model that simplifies drug and protein encoding using SUDD and SUPD, and employs a multi-view strategy with a transformer to extract key local residues from protein sequences. 1D denotes one-dimensional; 3D, three-dimensional; CNN, Convolutional Neural Network; DTI, drug -target interaction; ECFP, extended-connectivity fingerprints; ERT-GFAN, ECFP + RotatE + Transformer - Graphical High-dimensional Fusion Feature Attention Network ECFP + RotatE + Transformer - Graphical High-dimensional Fusion Feature Attention Network; GAT, Graph Attention Network; GIN, Graph Isomorphism Network; GCN, Graph Convolutional Network; GNN, graph neural network; GraphDTA, Graph Drug-Target affinity; InChIKey, hashed International Chemical Identifier; KGE, knowledge-graph embedding; MLC-DTA, Multi-Level Contrastive Learning for Drug -Target Affinity prediction; MolTrans, Molecular Interaction Transformer; SMILES, Simplified Molecular Input Line Entry System; SUDD, Simple Universal Drug Embedding Dictionary; SUPD, Simple Universal Protein Embedding Dictionary; Trans, Transformer.

Figure 2.

The Synergistic Relationship between Network Medicine and Artificial Intelligence. Network Medicine and Artificial Intelligence are not simply complementary — they are coevolutionary. One maps the complexity of systems; the other learns how to decode and act upon them. AI denotes artificial intelligence.

Figure 3.

Analytical Strategies for Networks of Heterogeneous (Biological) Information. Panel A shows heterogeneous layers of interconnected biological information. Network medicine harnesses artificial intelligence to tackle the challenges of integrating multiple layers of biological data into a cohesive framework. Each layer — genomic, transcriptomic, proteomic, and metabolomic — operates on distinct spatial and temporal scales, involving unique biological entities and interactions. These layers often exhibit noise, heterogeneity, and sparsity, challenging their integration. In addition, the relationships between layers are frequently nonlinear and context dependent, requiring advanced algorithmic solutions that surpass traditional topological analyses of the interactome. In Panel B, deep learning on networks derives contextual embeddings. Deep learning models utilize multilayered biological information and attention mechanisms to produce context-specific vector representations of biological entities, such as genes and proteins, capturing their roles and interactions within specific biological samples. Panel C shows quantum enhancement for diffusion, search algorithms, and community detection. The increasing complexity of network medicine data may benefit from the adoption of quantum-inspired algorithms to navigate multilayered biological networks efficiently, enhancing tasks such as diffusion modeling, search algorithms, and community detection.