Abstract

Inactivation of the retinoblastoma (Rb) tumor suppressor has long been considered a molecular hallmark of small cell lung carcinoma (SCLC), an aggressive form of bronchogenic, smoking-induced lung cancer with an extremely poor prognosis. Currently, SCLC can be classified into the following molecular subtypes: ASCL1, NEUROD1, POU2F3, and Inflamed. We analyzed a real-world cohort of ~1,400 SCLC tumors sequenced with both Tempus xT (DNA sequencing) and xR (RNA sequencing) and assigned them to molecular subtypes based on non-matrix factorization. Interestingly, we determined that RB1 alterations (e.g. single nucleotide variants, insertions/deletions) were not detected in approximately 30% of the Tempus SCLC cohort, regardless of tumor molecular subtype. However, absence of RB1 genomic alterations does not guarantee Rb protein function in SCLC, as alternate mechanisms such as RB1 promoter hypermethylation may result in gene silencing and loss-of-function (LOF), along with intronic splicing mutations that may not be well captured in probe-based methods. Therefore, we developed a machine learning model (Nested Random Forest) to predict Rb LOF in the Tempus SCLC cohort. We trained our model using 21,656 features including genomic, transcriptomic and clinical variables from 409 Tempus SCLC patients (cross-validation AUC 0.924 ± 0.020), whose tumors were either (1) RB1-altered and had low RB1 expression (defined as less than 25th percentile of expression of RB1-altered samples) or (2) RB1 wild-type and had high RB1 expression (defined as greater than 75th percentile of expression of RB1 wild-type samples). We then applied the model to remaining N=1,224 SCLC tumors in the Tempus dataset. Surprisingly, our model predicted that ~30% (N=241) of RB1-altered tumors (N=837) had an Rb wild-type or functional phenotype. This finding contradicts the notion that all RB1 genomic aberrations lead to loss of Rb pathway activity. Conversely, our model predicted that a proportion of RB1 wild-type tumors had Rb LOF, due to mechanisms such as CDKN2A (p16INK4a) deletion. Furthermore, we evaluated the frequency of predicted Rb function in the four SCLC subtypes. Our results showed that approximately one half of POU2F3 and Inflamed tumors had predicted Rb function, while approximately only 20% of ASCL1 and NEUROD1 tumors had predicted Rb function. In conclusion, our machine learning model predicted that a considerable number of real-world SCLC patients may contain a functional Rb pathway despite having genomic alterations in RB1. The impact of our finding on clinical outcomes including response to first-line chemotherapy and immunotherapy in real-world SCLC patients is currently being evaluated.

VIEW THE PUBLICATION