A knowledge graph screen of innovative versus incremental lung cancer trials for need-aligned designs

Background. ClinicalTrials.gov lists >500,000 studies, raising the question of how often “new” trials are truly novel versus incremental design variants. Lung cancer remains a leading cause of cancer-related death worldwide, underscoring the need for innovative over incremental trials. Prior work has used clinical trial knowledge graphs (KGs) to support design recommendations. Here, we apply a small cell lung cancer (SCLC) patient-anchored KG to classify lung cancer targeted therapy trials as “novel additions” versus “aggregate similarity” and to identify gaps where new, need-aligned trial designs are warranted. Methods. We curated 286 adult lung cancer targeted therapy trials from ClinicalTrials.gov that were retrieved as matches to a SCLC case. We built a trial-to-trial KG by linking each trial to key entities (e.g., tumor type, genomic alterations, targets/pathways, drug classes) and learned graph-based embeddings to obtain trial-level representations. Cosine similarity was used to quantify trial similarity between embeddings; community detection was used to define trial clusters. Clusters with ≥5 trials and median within-cluster similarity ≥0.80 were labeled “aggregate similar.” A novelty score (1 − maximum similarity to any other trial) and betweenness centrality were combined to label “novel additions”. A predefined set of SCLC molecular report-matched trials was cross-referenced to characterize their distribution across identified clusters. Results. Altogether, we defined 4 trial clusters (median size 67), three of which met our criteria for aggregate similarity, comprising 74% (n=212) of large community trials of highly similar designs. Across all 286 trials, median pairwise (0.74) and nearest-neighbor similarity (0.98) were consistent with dense replication of existing biomarker- and line-of-therapy-defined templates. Only 38 trials (13%) were classified as novel additions based on combined novelty (median=0.070) and betweenness centrality (median 0.004) thresholds, and showed significantly (p < 0.01) higher median scores in both as compared to non-novel trials. Novel additions were enriched in a single aggregate similarity cluster (24/74 trials) that contained most case molecular report-matched trials (65/68). The remaining trial clusters had 1-10% novel additions, indicating that case-level matching occurs within dense targeted therapy communities, whereas a KG analysis can surface structurally distinctive, novel addition trials. Conclusions. KG analysis of lung cancer trials provides a principled way to operationalize “novel addition” versus “aggregate similarity” at the portfolio level. This patient-anchored framework can support sponsors, investigators, and regulators in prioritizing innovative trials, reducing redundancy in an already crowded lung cancer trial landscape, and ultimately aligning trial development with unmet patient needs.