Zero-Shot Prediction

The model projects both drugs and diseases into a shared, high-dimensional latent space using auxiliary information such as gene expression signatures, chemical structures, or clinical text. A novel disease is positioned in this space based on its biological description, allowing the model to infer potential treatments by identifying drugs embedded in close proximity without ever observing that specific drug-disease pair during training.

Instead of memorizing specific drug-disease associations, the model learns to recognize the mechanistic attributes that make a drug effective. For example, it learns that 'kinase inhibition' is a property associated with treating certain cancers. When a completely new disease is characterized as having 'dysregulated kinase activity,' the model composes these learned attributes to predict a kinase inhibitor as a repurposing candidate, even if that specific disease was absent from the training data.

In a biomedical knowledge graph, drugs and diseases are nodes connected by edges representing known interactions, targets, and pathways. Zero-shot prediction is framed as a link prediction task: the model learns a scoring function that evaluates the plausibility of a connection between any two nodes based on the graph's relational patterns. When a new disease node is added with its known connections (e.g., associated genes), the model can infer missing links to existing drug nodes without retraining.

The model is trained to understand the relationship between a drug's gene expression perturbation signature and a disease's differential expression signature. The core logic is that a drug which reverses a disease's expression pattern is a potential therapeutic. In a zero-shot context, the model can compute the reversal score for a disease signature it has never encountered by comparing it against a library of learned drug perturbation profiles, generalizing the signature-reversal principle to novel pathologies.

Rigorous evaluation of zero-shot models requires specialized data splitting strategies. Unlike random splitting, which causes data leakage, zero-shot validation holds out entire drug or disease entities from the training set. For instance, all associations for a specific disease are placed in the test set. The model must then predict treatments for this held-out disease, providing a realistic measure of its ability to generalize to truly novel clinical scenarios.

Zero-shot prediction directly addresses the cold-start problem in drug discovery, where new diseases or orphan indications lack any known pharmacological treatments. By relying on auxiliary biological data rather than historical association matrices, the model can generate actionable hypotheses for these cold-start entities from day one, bypassing the data scarcity that paralyzes traditional collaborative filtering or matrix factorization approaches.

A matrix factorization variant that incorporates side information—such as drug chemical structures and disease gene signatures—to predict associations for entities not present in the original training matrix. Unlike transductive methods, it learns a mapping from feature space to latent space, enabling zero-shot inference for completely new drugs or diseases by projecting their features directly.

A machine learning technique that projects entities and relations from a biomedical knowledge graph into a low-dimensional vector space. By learning translational distance models (e.g., TransE) or semantic matching models (e.g., ComplEx), the system can predict missing links. For zero-shot drug repurposing, embeddings for a novel disease node can be inferred from its textual description or ontological connections without retraining the entire graph.

A self-supervised representation learning method that pulls together augmented views of the same biological entity while pushing apart representations of different entities. In zero-shot settings, a model pre-trained with contrastive loss on drug-target pairs learns a joint embedding space where molecular structure and protein sequence are aligned. A novel drug can then be matched to a disease target by proximity in this shared space without explicit training on that pair.

An inductive transfer approach that trains a single model simultaneously on multiple related prediction tasks—such as binding affinity, toxicity, and solubility. By sharing hidden representations across tasks, the model learns a generalized molecular grammar. This shared representation enables zero-shot generalization: a drug optimized for one target class can be evaluated against an unseen disease target because the model has learned transferable structure-activity principles.

A critical validation error where information from the test set inadvertently influences training, leading to over-optimistic performance estimates. In zero-shot drug repurposing, rigorous temporal splitting (training on older drugs, testing on newer approvals) or strict scaffold separation (ensuring no chemical similarity between train and test molecules) is mandatory. Without these controls, a model may appear to perform zero-shot prediction when it is actually memorizing structural analogs.

A computational approach that compares a disease's gene expression pattern against a reference database of drug-induced gene expression changes, such as the Connectivity Map (CMap) . For zero-shot prediction, a model trained on the L1000 assay data learns to embed both disease signatures and drug perturbation profiles into a common space. A novel disease with only a transcriptional signature can then be matched to existing drugs that reverse that pattern without prior clinical association data.