Thread by @RielyMD

THREAD: (1/10) Our work in JCO PO looked at NGS data from patients with met lung adeno. Some goals: (1) explore notion of molecular staging (within stage IV) using somatic mut & (2) see what drives poorer outcomes under current standard of care ascopubs.org/doi/full/10.12… #MSKIMPACT

(2/10) We knew some single-gene mutations were good (e.g. EGFR, ALK) or bad predictive/prognostic determinants. OncoCast helped broadly explore effect of genes & gene combinations using an ensemble statistical learning strategy

(3/10)Some genes like STK11 and KEAP1 came out as neg survival factors on most runs of model. EGFR came out as positive. Interesting to look at the size of effect on survival (KRAS came out a lot, but wasn’t as strong a negative effect as STK11 or KEAP1)

(4/10) We categorized the cohort into 4 risks groups. There was reasonable separation in survival outcomes. (perhaps a molecular staging within stage IV lung adenocarcinoma). [For external validation, look for us #ASCO2019 @ASCO ]

(5/10) OncoCast risk score (based on all somatic mutations on MSK-IMPACT panel in a tumor) was better than individual mutations or clinical factors in predicting overall survival. Tools to calculate risk score for individuals and populations of patients are freely available.

(6/10) What are some of the mutations/characteristics associated with risk? Patients with the lowest risk tumors had EGFR, ALK. KRAS and TP53 are broadly present across the risk spectrum. STK11/KEAP1 alterations are common in the highest risk group.

(7/10) We used a similar technique to derive a risk model specific to patients who received targeted therapies (OncoCast-TR). In this model, we looked at patients with EGFR, ALK, or ROS1. Using this, we found two distinct groups.

(8/10) In this group of people (with tumors with EGFR, ALK, ROS1) who receive targeted therapies, the additional factors that led to reduction in overall survival were things like TP53, STK11, and ARID1A.

(9/10) This was a cool use of advanced statistical techniques to derive additional value from data we get with routine clinical broad-panel sequencing. OncoCast would be useful for describing risk profiles of patients in trials (e.g. in Table 1) as well as in “real-world” data

(10/10) Software available at github.com/shenmskcc/Onco…
Check out the paper. Thanks. ascopubs.org/doi/full/10.12…