Inductive Matrix Completion

The foundational collaborative filtering technique that decomposes a sparse drug-disease association matrix into low-rank latent factor matrices. Unlike IMC, classic matrix factorization operates solely on the observed interaction matrix without integrating auxiliary side information. It learns embeddings for drugs and diseases by minimizing the reconstruction error on known associations, then predicts novel links by computing the dot product between the corresponding latent vectors. This approach is limited to transductive settings and cannot generalize to new drugs or diseases absent from the training matrix.

A graph-based representation where nodes represent drugs and edges connect compounds sharing chemical, biological, or side-effect profiles. These networks encode the side information that IMC exploits as feature matrices. Key similarity measures include:

• Chemical similarity: Tanimoto coefficient computed on molecular fingerprints
• Therapeutic similarity: Shared ATC classification codes
• Side-effect similarity: Overlap in reported adverse event profiles
• Genomic similarity: Correlation of drug-induced gene expression signatures These networks are often constructed from databases like DrugBank or STITCH.

A machine learning paradigm where a model predicts drug-disease associations for completely unseen entities without any specific training examples for that particular pair. IMC inherently supports zero-shot prediction for new drugs or diseases because it learns a mapping from the feature space to the latent space. When a novel drug arrives, its chemical fingerprint is projected directly into the latent factor matrix via the learned feature transformation, enabling immediate inference without retraining. This contrasts sharply with standard matrix factorization, which requires retraining the entire model to incorporate new rows or columns.

An inductive transfer approach that trains a single model simultaneously on multiple related prediction tasks to improve generalization through shared representations. In drug repurposing, a multi-task framework might jointly predict:

• Binding affinity against multiple kinase targets
• Toxicity endpoints such as hERG channel inhibition
• Therapeutic indication across disease categories IMC can be extended to multi-task settings by sharing the drug feature transformation matrix across tasks while learning task-specific disease projections, leveraging common chemical patterns to improve performance on data-poor targets.

A machine learning technique that projects the entities and relations of a biomedical knowledge graph into a low-dimensional vector space to predict missing links. While IMC uses fixed feature matrices derived from chemical or genomic data, knowledge graph embedding methods like TransE, RotatE, or ComplEx learn embeddings directly from the graph topology. These learned embeddings can serve as the side information matrices for IMC, creating a hybrid approach where the feature representations themselves are optimized from relational data rather than hand-crafted descriptors.

A graph machine learning task that estimates the likelihood of a missing connection between two nodes in a heterogeneous biomedical network. IMC frames drug repurposing as a link prediction problem on a bipartite drug-disease graph, where the goal is to score unobserved edges. The inductive capability allows scoring links involving cold-start nodes that have no existing connections. Evaluation metrics for link prediction include:

• AUROC: Area under the receiver operating characteristic curve
• AUPRC: Area under the precision-recall curve, more informative for sparse matrices
• Hits@K: The fraction of true associations ranked in the top K predictions