Multiscale DAGI Validation on Cancer Drug Synergy: Information-Theoretic Decomposition of the NCI–ALMANAC

The National Cancer Institute (NCI) ALMANAC is one of the largest publicly available resources on pairwise cancer drug combinations, comprising more than 300, 000 combination experiments across the NCI--60 cell-line panel and reporting both monotherapy and combination responses for 104 FDA-approved anticancer agents. This dataset has been widely used to develop predictive models of drug synergy, but the multiscale informational structure of the combination effects---how much of the response is attributable to individual drugs, pairwise interactions, or higher-order context---remains poorly quantified. Here we present a first biological validation of the Directed Acyclic Graph Informational (DAGI) framework on the NCI--ALMANAC resource. Using a Möbius inversion-based multiscale information decomposition previously applied to quantum Darwinism experiments and holographic stabilizer, we analyze how information about tumor cell-kill is distributed across three variables: drug~1, drug~2, and cellular context (cell line). We define two target variables: (i) the Bliss-excess growth inhibition, comparing observed viability to an independence model; and (ii) the ALMANAC ComboScore. For each drug pair with sufficient sampling, we estimate the mutual information between these targets and the triplet \ ( (D₁, D₂, C) \), then decompose this information into irreducible contributions from single variables, pairs, and the full triple. Across 500 drug pairs (each with at least 100 combination experiments), we find that synergy is predominantly a third-order, context-dependent phenomenon. Between 94% and 97. 2% of pairs require the full triple \ (\D₁, D₂, C\\) to reach their maximal synergy order; pairwise drug--drug terms alone are insufficient. On average, triple-order terms are 3--4 times larger than drug--drug terms, with a mean irreducible context-dependent synergy of \ (f₃䃑₃䃒₂ 0. 056\) bits compared to \ (f₃䃑₃䃒 0. 012\) bits. For Bliss-excess responses, first-, second-, and third-order contributions account for approximately 52. 4%, 35. 4%, and 12. 2% of total information, respectively; for ComboScore these fractions shift to 36. 0%, 44. 0%, and 20. 0%, indicating that ComboScore is more sensitive to truly emergent combination effects. These results provide an independent, biological validation of a core DAGI prediction: that many phenomena of interest are governed by irreducible higher-order synergies that cannot be recovered from pairwise interactions alone. In the cancer therapy context, they also formalize a long-held intuition in precision oncology: drug synergy is not an intrinsic property of drug pairs, but an emergent, context-dependent feature of the \ ( (drug₁, drug₂, tumor) \) triad. Our analysis establishes a reusable DAGI-based pipeline for multiscale information decomposition on large biological datasets, opening the door to systematic studies of informational thresholds in living systems.