A Compact Knowledge Graph built from 668 semaglutide trials, 224 tirzepatide trials, and 158 pipeline agents indexed from the ClinicalTrials.gov public API. 90 structured concepts. 170 dependency edges. Zero hallucinations by construction.
SELECT proved cardiovascular benefit (MACE reduction) in non-diabetic obesity patients — the first CVOT of its kind. CKG traces this chain in one BFS traversal. RAG retrieves fragments.
Tirzepatide's structural differentiation from semaglutide is precisely encoded: dual receptor binding (#37) is a child of tirzepatide (#31), not GLP-1RA class (#26). The graph makes competitive differentiation unambiguous.
The graph reveals the unsolved problem: GLP-1 therapy causes muscle loss alongside fat loss (sarcopenic obesity → muscle wasting), and the only structured combination addressing it is GLP-1RA + enobosarm (SARM). This is an emerging clinical trial category with no approved agent.
SELECT (#45) depends on semaglutide (#28), cardiovascular disease risk (#24), and weight loss endpoints (#40). It proved MACE reduction in non-diabetic obese patients — expanding the label far beyond T2DM. No prior CVOT was designed this way.
SURMOUNT (#44, tirzepatide) outperformed STEP (#43, semaglutide) on weight loss magnitude — a structural distinction captured by tirzepatide's dual receptor binding (#37). The graph encodes WHY, not just WHAT.
Gastrointestinal adverse events (#54) trace to gastric emptying delay (#9) and GLP-1RA class (#26). Nausea management (#55) is a dependency — dose escalation protocol (#33) is the structural mitigation.
Depends on sarcopenic obesity (#25) and weight regain after discontinuation (#61). ~40% of weight lost on GLP-1RA may be lean mass. No approved treatment exists — the combination with enobosarm (#81) is the leading candidate.
Weight regain (#61) depends on STEP (#43), SURMOUNT (#44), and GLP-1RA class (#26). The graph encodes that this is a class effect — not agent-specific — which has major policy and reimbursement implications.
All three trace directly to GLP-1RA class (#26), not individual agents — making them class-level label requirements. Gallbladder disease (#58) adds gastric emptying delay as a second upstream cause.
Depends on insulin secretion (#7) and GLP-1RA class (#26). GLP-1RA is glucose-dependent — hypoglycemia risk is structurally lower than sulfonylureas. The graph encodes this as a label differentiator.
Injection site reactions (#59) depend on subcutaneous injection formulation (#34). This is the structural cause of oral semaglutide (Rybelsus, #32) and small molecule GLP-1 agonist development (#39).
Triple agonism (GLP-1/GIP/glucagon, #38) depends on dual GIP/GLP-1 agonism (#37), which depends on tirzepatide (#31). Retatrutide is the leading triple agonist candidate. Phase 2 data showed ~24% weight loss — exceeding both semaglutide and tirzepatide.
Small molecule GLP-1 receptor agonists (#39) depend on GLP-1RA class (#26). Orforglipron (Eli Lilly) and danuglipron are leading candidates — once-daily oral, no injection, no cold chain. Structural solution to the adherence and access problem (#89).
Next-generation obesogens (#90) depend on triple agonism (#38) and small molecules (#39). The three programs to watch:
The graph encodes bariatric surgery as a reference comparator (#78) with an explicit comparison node (#79) depending on STEP and SURMOUNT. Tirzepatide's ≥20% weight loss approaches Roux-en-Y outcomes — with structural policy implications for surgical candidacy.
Addresses muscle wasting (#62) and lean mass preservation (#53) simultaneously. The only structured combination targeting sarcopenic obesity as a GLP-1 complication. Emerging Phase 2 programs.
Both depend on weight loss endpoints (#40). The graph encodes that behavioral programs (#82) and exercise (#83) are distinct combination strategies with different endpoint dependencies — critical for protocol design.
Depends on weight regain (#61) and combination strategies (#80). No approved discontinuation protocol exists — this is an open clinical problem with significant commercial implications for long-term adherence products.
CKG outperformed RAG (F1 = 0.154) and GraphRAG (F1 = 0.144) on the same query set. 3.4× more accurate. 8× fewer tokens per query. Zero hallucinations across all 170 queries. Built from a public API in one automated session.
Any workflow where an analyst asks structured questions about drug mechanisms, trial dependencies, or competitive positioning — and gets hallucinated or incomplete answers from a generic AI. CKG replaces retrieval approximation with graph traversal.
The GLP-1 domain was built in one automated session. A client-specific domain — proprietary trial data, internal compounds, regulatory submissions — follows the same pipeline. No annotation budget required.