Towards overtreatment-free immunotherapy: Using genomic scars to select treatment beneficiaries in lung cancer

In advanced non-small cell lung cancer (NSCLC), response to immunotherapy is difficult to predict from pre-treatment information. Given the toxicity of immunotherapy and its financial burden on the healthcare system, we set out to identify patients for whom treatment is effective using mutational signatures from DNA mutations in pre-treatment tissue. Analysis of single base substitutions, doublet base substitutions, indels, and copy number alteration signatures in the discovery set ( m = 101 patients) linked tobacco smoking signature (SBS4) and thiopurine chemotherapy exposure-associated signature (SBS87) to durable benefit. Combining both signatures in a machine learning model separated patients with a progression free survival hazard ratio of 0.40 +0 . 28 − 0 . 17 on the cross validated discovery set and 0.24 +0 . 31 − 0 . 14 on an independent external validation set ( m = 56). This paper demonstrates that the fingerprints of mutagenesis, codified through mutational signatures, can be used to select advanced

In non-small cell lung cancer (NSCLC), response to immunotherapy is low, with 2 radiology-assessed response typically around 20-25% [1], while the percentage 3 of patients achieving durable benefit (DB), defined as progression free survival 4 (PFS) ≥ 1 2 year, is only slightly higher. As a result, the majority of patients 5 are subjected to a futile toxic treatment. 6 Predictors can help to narrow down specific subpopulations for which treat-

15
Once the cancer cell expresses neoantigens, the cancer cell can be eliminated 16 through immune recognition and cell killing [1,2]. 17 While several studies show a clear association of TMB with response to treat-

33
Evidence is accumulating that mutational signatures in cancer are therapeu-34 tically relevant [9]. Given the partial success of TMB for predicting immunother- defense that operate by generating mutations in single-stranded DNA [9,18]; 48 and signatures similar to clock-like signature SBS1, capturing mutations that 49 steadily accrue with age, were linked to non-response to immunotherapy by 50 Chong et al. [17]. 51 Here, instead of de novo analysis we directly use previously catalogued mu- 52 tational signatures like in earlier work [4,19]. We expand previous efforts by 53 interrogating an order of magnitude more signatures, including the recently de- and females (53.5%), and the majority of the patients (85/147, 57.8%) did not 66 achieve durable benefit from immunotherapy (Table 1).

67
Mutational signatures in the discovery set.

68
Whole genome sequencing revealed a mean of 7.   Figure 1: Mutational signatures from pre-treatment tumor tissue predict immunotherapy efficacy in advanced non-small cell lung cancer. a, Single base substitutions (SBS) are determined from preimmunotherapy tumor material. After deconvolution into signature attributions, a machine learning classifier uses smoking-associated signature SBS4 and thiopurine chemotherapy-associated signature SBS87 to predict durable benefit (DB) from immunotherapy. b, Cartoon illustration of SBS signature deconvolution, where we solve for signature attribution W given mutation spectrum X and COSMIC signatures H through X ≈ W H. Nucleotide pyramids indicate SBS with flanking context; Sun, cigarette, and Erlenmeyer symbols depict example aetiologies; Shading highlights information that pertains to the corresponding patient. For illustration purposes, the size of the dots do not represent actual data. c, Signatures SBS4 [q = 0.014, Benjamini-Hochberg corrected Kolmogorov-Smirnov (B-HK-S) test] and SBS87 (q = 0.017, B-HK-S test) are linked to DB (discovery set). d, Signatures SBS4 and SBS87 correlate with mutations in genes canonically mutated in cancer (discovery set). Correlations were assessed using a B-H corrected Kendall τ (corrected p-values along the arrows) and the correlation strengths are indicated by the arrow line widths. e, Patients predicted to have DB (Signatures +, blue line) have superior progression free survival compared to those predicted to have non-DB (Signatures -, orange line) in the discovery set (left panel). The classifier's performance replicates in an independent validation set (right). Censored observations are indicated by crosses. Estimates and corresponding 95% confidence intervals are indicated by sub and superscripts.