Drug Combination Prediction

A specialized neural network architecture designed explicitly for drug combination prediction. DeepSynergy uses multi-branch input processing to handle heterogeneous drug and cell line features.

• Drug branches: Separate subnetworks process molecular fingerprints, physicochemical descriptors, or graph embeddings for each compound
• Cell line branch: Encodes genomic features such as gene expression, mutation status, and copy number variations
• Fusion layer: Concatenates learned representations and passes them through fully connected layers to predict synergy scores
• Output: A continuous synergy score or classification (synergistic vs. antagonistic) for the drug-drug-cell line triplet

A quantitative metric that measures the degree of interaction between two drugs. It determines if a combination is synergistic (effect greater than the sum of parts), additive, or antagonistic.

• Loewe Additivity: Assumes a drug cannot interact with itself; the gold standard for dose-effect analysis.
• Bliss Independence: A probabilistic model based on the assumption of independent action.
• HSA (Highest Single Agent): A simple reference model where the combination effect cannot exceed the maximum single-agent effect.

The design or functional capacity of a single drug molecule to interact with multiple distinct biological targets simultaneously. This is a core driver of both therapeutic efficacy and off-target toxicity.

• Exploits the fact that many diseases are multifactorial, requiring modulation of several pathways.
• Computational polypharmacology uses proteome-wide docking and network analysis to map a drug's full interaction profile.
• Kinase inhibitors are a classic example, often hitting dozens of kinases beyond the primary target.

The physical binding event between a chemical compound and a specific biomolecule, typically a protein, which modulates a biological process. Predicting these interactions is fundamental to understanding combination mechanisms.

• Binding Affinity (Kd, IC50): Quantifies the strength of the interaction.
• Molecular Docking: A structure-based computational method that predicts the preferred orientation of a drug to its target.
• Deep learning models like Graph Neural Networks now predict binding affinities directly from 3D structural data.

The computational forecasting of adverse drug reactions by modeling the interaction between a drug's chemical structure and off-target biological pathways. This is critical for de-risking combination therapies.

• Uses transcriptomic signature matching to link drugs to gene expression changes associated with known toxicities.
• Knowledge graph embeddings can predict unobserved side effects by traversing drug-protein-disease networks.
• A major challenge is predicting emergent side effects that only appear when two drugs are combined, not individually.

The specific biochemical interaction through which a drug substance produces its pharmacological effect. Understanding MoA is essential for rational combination design.

• Target Deconvolution: The experimental or computational process of identifying the molecular target of a compound.
• CRISPR-Cas9 knockout screens are a powerful experimental method for MoA elucidation.
• Computational approaches use causal inference on gene expression data to distinguish direct targets from downstream effects.

A computational approach that compares a disease's gene expression pattern against a reference database of drug-induced gene expression changes. The goal is to find compounds that reverse the disease state.

• The foundational resource is the Connectivity Map (CMap).
• For combination prediction, the goal is to find two drugs whose combined signature more completely reverses the disease signature than either alone.
• Modern approaches use contrastive learning to embed gene expression profiles into a space where therapeutic relationships are preserved.