APOBEC mutagenesis, a major driver of cancer evolution, is known for targeting TpC sites in DNA. Recently, we showed that APOBEC3A (A3A) targets DNA hairpin loops. Here, we show that DNA secondary structure is in fact an orthogonal influence on A3A substrate optimality and, surprisingly, can override the TpC sequence preference. VpC (non-TpC) sites in optimal hairpins can outperform TpC sites as mutational hotspots. This expanded understanding of APOBEC mutagenesis illuminates the genomic Twin Paradox, a puzzling pattern of closely spaced mutation hotspots in cancer genomes, in which one is a canonical TpC site but the other is a VpC site, and double mutants are seen only in trans, suggesting a two-hit driver event. Our results clarify this paradox, revealing that both hotspots in these twins are optimal A3A substrates. Our findings reshape the notion of a mutation signature, highlighting the additive roles played by DNA sequence and DNA structure.

Purpose: Current standard initial therapy for advanced, ROS proto-oncogene 1, receptor tyrosine kinase fusion (ROS1)-positive (ROS1+) non–small cell lung cancer (NSCLC) is crizotinib or entrectinib. Lorlatinib, a next-generation anaplastic lymphoma kinase/ROS1 inhibitor, recently demonstrated efficacy in ROS1+ NSCLC, including in crizotinib-pretreated patients. However, mechanisms of lorlatinib resistance in ROS1+ disease remain poorly understood. Here, we assessed mechanisms of resistance to crizotinib and lorlatinib. Experimental Design: Biopsies from patients with ROS1+ NSCLC progressing on crizotinib or lorlatinib were profiled by genetic sequencing. Results: From 55 patients, 47 post-crizotinib and 32 post-lorlatinib biopsies were assessed. Among 42 post-crizotinib and 28 post-lorlatinib biopsies analyzed at distinct timepoints, ROS1 mutations were identified in 38% and 46%, respectively. ROS1 G2032R was the most commonly occurring mutation in approximately one third of cases. Additional ROS1 mutations included D2033N (2.4%) and S1986F (2.4%) post-crizotinib and L2086F (3.6%), G2032R/L2086F (3.6%), G2032R/S1986F/L2086F (3.6%), and S1986F/L2000V (3.6%) post-lorlatinib. Structural modeling predicted ROS1L2086F causes steric interference to lorlatinib, crizotinib, and entrectinib, while it may accommodate cabozantinib. In Ba/F3 models, ROS1L2086F, ROS1G2032R/L2086F, and ROS1S1986F/G2032R/L2086F were refractory to lorlatinib but sensitive to cabozantinib. A patient with disease progression on crizotinib and lorlatinib and ROS1 L2086F received cabozantinib for nearly 11 months with disease control. Among lorlatinib-resistant biopsies, we also identified MET amplification (4%), KRAS G12C (4%), KRAS amplification (4%), NRAS mutation (4%), and MAP2K1 mutation (4%). Conclusions: ROS1 mutations mediate resistance to crizotinib and lorlatinib in more than one third of cases, underscoring the importance of developing next-generation ROS1 inhibitors with potency against these mutations, including G2032R and L2086F. Continued efforts are needed to elucidate ROS1-independent resistance mechanisms.