Data Collection

ComboAtlas was constructed through manual curation of 10,469 publications spanning 1985 to 2026. Target entities were primarily standardized using HGNC identifiers, while disease entities were normalized using the MONDO disease ontology. For each disease-target combination association, ComboAtlas annotates the supporting evidence level (clinical, in vivo, or in vitro) and the intervention direction of each target within the combination. To further enrich the annotation of the curated data, ComboAtlas integrates drug-target interaction information from ChEMBL, while IUPHAR and PANTHER are used to provide complementary target-level annotations. Together, these data establish an interconnected framework linking targets, target combinations, diseases, and drugs, enabling users to navigate therapeutic combination evidence from multiple perspectives.

Interpretation

Once these target combinations were collected, a natural question was whether they are organized randomly, or whether recurring patterns can be found in how therapeutic targets are paired. Networks provide a useful way to explore relationships among biological entities, so ComboAtlas examines target combinations from both protein interaction and pathway perspectives, using the PPI network from Cheng et al. and the Reactome pathway hierarchy.

Rather than looking at PPI relationships with a single metric, we examine them at three progressively broader scales. We start with shortest-path distance to see how closely two targets are connected, then move to Resource Allocation (RA) to ask whether they share informative neighbors, and finally use RWR-based similarity to explore whether their influence spreads into similar regions of the broader network.

Beyond physical interaction networks, we also ask whether paired targets are functionally related at the pathway level. Using the Reactome pathway hierarchy, ComboAtlas applies Lin-BMA similarity to evaluate how closely the pathway annotations of two targets are positioned within the functional hierarchy. Importantly, these network and pathway relationships are not interpreted in isolation. For each target pair, we construct a degree-matched random background and compare the observed relationship with what would be expected by chance, allowing us to examine whether individual combinations and therapeutic target combinations as a whole show non-random organizational patterns.

Prediction

After exploring the organizational patterns of known target combinations, the next question naturally follows: can the evidence we have collected help us prioritize new combinations for experimental investigation? To address this, we developed a disease-specific target-combination prediction model using the curated ComboAtlas data. The model combines the network-based characteristics identified above with a deep-learning framework to learn from known therapeutic combinations, and has been evaluated across multiple prediction and ranking settings (see our publication for details).

ComboAtlas provides an online prediction tool in which users select a disease of interest and an anchor target to obtain a ranked list of potential partner targets. This setup is intended to reflect a practical research scenario, where a biologically or therapeutically relevant target has already been identified and the next step is to search for suitable combination partners. Users can then explore high-ranking candidates and select promising target pairs for further investigation and experimental validation. To provide additional context for anchor selection, ComboAtlas also integrates disease-target association scores from Open Targets, helping users assess how strongly the chosen anchor target is associated with the selected disease before interpreting the predicted partners.