KRAS G12C mutations are important oncogenic mutations that confer sensitivity to direct G12C inhibitors. We retrospectively identified patients with KRAS+ NSCLC from 2015 to 2019 and assessed the imaging features of the primary tumor and the distribution of metastases of G12C NSCLC compared to those of non-G12C KRAS NSCLC and NSCLC driven by oncogenic fusion events (RET, ALK, ROS1) and EGFR mutations at the time of initial diagnosis. Two hundred fifteen patients with KRAS+ NSCLC (G12C: 83; non-G12C: 132) were included. On single variate analysis, the G12C group was more likely than the non-G12C KRAS group to have cavitation (13% vs. 5%, p = 0.04) and lung metastasis (38% vs. 21%; p = 0.043). Compared to the fusion rearrangement group, the G12C group had a lower frequency of pleural metastasis (21% vs. 41%, p = 0.01) and lymphangitic carcinomatosis (4% vs. 39%, p = 0.0001) and a higher frequency of brain metastasis (42% vs. 22%, p = 0.005). Compared to the EGFR+ group, the G12C group had a lower frequency of lung metastasis (38% vs. 67%, p = 0.0008) and a higher frequency of distant nodal metastasis (10% vs. 2%, p = 0.02). KRAS G12C NSCLC may have distinct primary tumor imaging features and patterns of metastasis when compared to those of NSCLC driven by other genetic alterations.

Purpose Next-generation sequencing technologies are actively applied in clinical oncology. Bioinformatics pipeline analysis is an integral part of this process; however, humans cannot yet realize the full potential of the highly complex pipeline output. As a result, the decision to include a variant in the final report during routine clinical sign-out remains challenging. Methods We used an artificial intelligence approach to capture the collective clinical sign-out experience of six board-certified molecular pathologists to build and validate a decision support tool for variant reporting. We extracted all reviewed and reported variants from our clinical database and tested several machine learning models. We used 10-fold cross-validation for our variant call prediction model, which derives a contiguous prediction score from 0 to 1 (no to yes) for clinical reporting. Results For each of the 19,594 initial training variants, our pipeline generates approximately 500 features, which results in a matrix of > 9 million data points. From a comparison of naive Bayes, decision trees, random forests, and logistic regression models, we selected models that allow human interpretability of the prediction score. The logistic regression model demonstrated 1% false negativity and 2% false positivity. The final models’ Youden indices were 0.87 and 0.77 for screening and confirmatory cutoffs, respectively. Retraining on a new assay and performance assessment in 16,123 independent variants validated our approach (Youden index, 0.93). We also derived individual pathologist-centric models (virtual consensus conference function), and a visual drill-down functionality allows assessment of how underlying features contributed to a particular score or decision branch for clinical implementation. Conclusion Our decision support tool for variant reporting is a practically relevant artificial intelligence approach to harness the next-generation sequencing bioinformatics pipeline output when the complexity of data interpretation exceeds human capabilities.

3079 Background: ALK tyrosine kinase inhibitors (TKIs) are effective in treating advanced anaplastic lymphoma kinase (ALK) fusion-positive non-small-cell lung cancers (NSCLC), and specific ALK variants are associated with the development of resistance to specific TKIs. Humans struggle to harness the full potential of the highly complex next-generation sequencing bioinformatics pipeline output. As a consequence, the decision to report a variant remains difficult, and we considered the discrete nature of the data and the binary decision (report vs. not-report) as an ideal setting to apply an artificial intelligence (AI) approach for variant reporting. Methods: We assessed diagnostic performance of an AI model in calling ALK-resistance mutations in n = 50 consecutive ALK fusion positive patients who relapsed on TKI-therapy and underwent repeat biopsy at MGH. The random forest model was derived from independent datasets (training and validation) capturing the reporting decision on > 36,000 variants with ~500 features per variant resulting in a matrix of > 18 million data points. The model output is a contiguous prediction score from 0 (not report) to 1 (report) and a visual drill-down functionality allows exploration of the underlying features that contributed to the decision. Results: Examination of n = 76 tests from n = 50 patients with a total of n = 130 reported variants (and = 115 not reported variants) included a total of n = 31 ALK point mutations: p.1156(n = 2), p.1171(n = 8), p.1174(n = 2), p.1180(n = 2), p.1196(n = 1), p.1198(n = 1), p.1202(n = 8), p.1203(n = 1), p.1204(n = 1), p.1206(n = 1), p.1269(n = 4). Setting a screening threshold of the model at > 10% for reporting showed only one false-negative (p.Ile1171Asn) variant and 96.7% sensitivity. The average model score for ALK variants was 0.664 (range: 0.08–0.98; median 0.8) and did not show significant differences from other reported variants (0.636; 0–1; 0.7; t-test 0.66). The model would have called n = 18 of the non-reported control variants (average 0.07; range < 0.001–0.64; P < 0.0001) and was 84% specific. Review of the drill-down function identified prior call frequency, allelic ratio, and predicted transcript consequences as common model features. Importantly, the model is currently agnostic to the medical literature and does not take clinical parameters (e.g. TKI type) into account, which may further improve performance. Conclusions: Applying artificial intelligence to large discrete datasets is one approach to help identify clinically relevant variants in the setting of ALK resistance in ALK-fusion positive NSCLC.