Article Text

Download PDFPDF

33 Dynamic monitoring of response to immune checkpoint blockade through deep-learning empowered ultra-sensitive liquid biopsy in melanoma
  1. Adam Widman1,
  2. Cole Khamnei2,
  3. Jake Bass2,
  4. Will Liao2,
  5. Minita Shah2,
  6. Nicolas Robine2,
  7. Jedd Wolchok1,
  8. Margaret Callahan1 and
  9. Dan Landau3
  1. 1Memorial Sloan Kettering Cancer Center, New York, NY, USA
  2. 2New York Genome Center, New York, NY, USA
  3. 3Weill Cornell Medicine, New York, NY, USA


Background Clearance of circulating tumor DNA (ctDNA) following checkpoint blockade (CB) can precede radiographic response,1 2 though current state of the art ctDNA detection via targeted panels faces limited sensitivity in low burden disease (figure 1). We previously showed that whole genome sequencing (WGS) of plasma can overcome low input of ctDNA to dynamically track low volume malignancy using matched tumor tissue.3 We therefore sought to evaluate ctDNA for tracking early response to checkpoint blockade (CB) in melanoma, and developed a novel classifier that allows us to track disease without matched tumor tissue for expanded applicability in immunotherapy.

Methods To identify ctDNA sparsely diluted in noncancerous plasma cell free DNA (cfDNA), we developed Phoenix, a deep-learning classifier that uses genomic and epigenomic features to distinguish single nucleotide variants (SNVs) in melanoma from sequencing noise. We evaluated Phoenix on a retrospective cohort of serially sampled plasma from patients with advanced cutaneous melanoma on CB (nivolumab alone or with ipilimumab). Plasma was collected at 0, 3, 6 and 12 weeks after first dose of immunotherapy. ctDNA dynamics were compared to radiographic imaging results at 12 weeks.

Results We trained Phoenix on tumor-confirmed SNVs in plasma from a single patient with high tumor mutational burden (TMB) melanoma and cfDNA from age-matched patients without known cancer. Overall ctDNA signal-to-noise enrichment ranged from 100 - 260x in validation patients (n=2) with bulky disease. Phoenix learned key features of melanoma ctDNA including the UV mutational signature and short fragment size (figure 2), and sensitively tracked persistent low burden disease seen on imaging (figure 3). To validate these findings, we expanded our cohort (n= 15) of serially tracked tumors. In our preliminary analysis of 12 patients, Phoenix detected pretreatment ctDNA in 92% of patients at a specificity of 97% (figure 4), compared with only 17% with the benchmark in the field (iChorCNA, a plasma-based WGS liquid biopsy tool; table 1). Phoenix detected a decrease in ctDNA 3 weeks after initiation of CB in 80% of patients (figure 5) with an objective response on imaging. No change in ctDNA was seen in patients who did not respond to treatment.

Conclusions Phoenix successfully identified pretreatment melanoma ctDNA without matched tumor tissue and identified response to CB as early as 3 weeks after treatment. Our ongoing studies aim to optimize this technology for early identification of CB response in clinical practice.

Abstract 33 Figure 1

WGS of plasma increases sensitivity in low-burden diseaseLikelihood of ctDNA SNV detection is a function of tumor fraction, depth, and breadth (number of candidate sites). Because the limited number of genomic equivalents exhausts depth in targeted sequencing, detection sensitivity is limited by the relatively small number of sites in a clinical panel. In contrast, WGS at modest depth (35x) can detect low tumor fraction by integrating signal from thousands of SNVs across the genome.

Abstract 33 Figure 2

Phoenix learns key covariates for melanoma ctDNAPhoenix was trained on tumor-confirmed SNVs in plasma from patients with high burden melanoma and cfDNA from age-matched patients without known cancer. We aggregated Phoenix positive (ctDNA, blue) and negative (cfDNA, red) predictions on SNVs from a held out validation melanoma plasma sample. Phoenix ctDNA predictions correctly reflect important melanoma SNV attributes including UV-signature (C>T trinucleotide context, a), low DNase accessibility (b), late replication timing (c), and short fragment length (d).

Abstract 33 Figure 3

Phoenix sensitively tracks response to nivolumabPlasma samples were collected to monitor treatment response to nivolumab. Treatment monitoring by computed tomography (CT) shows response to therapy but residual disease after 3 months of therapy (a). Phoenix quantifies tumor response, matching radiographic changes, in higher temporal resolution than what is feasible with imaging (b). IchorCNA sensitivity captures initial treatment response dynamics but does not detect residual disease after 3 months of treatment (c). Log z score is calculated from a single plasma sample for each timepoint compared to a panel of control samples (n = 37).

Abstract 33 Table 1

Characteristics of patients at baseline and ctDNA dynamics

Abstract 33 Figure 4

Phoenix detects pre- and intratreatment melanoma ctDNAWe evaluated Phoenix post-filter sample-level detection rate. Phoenix detects ctDNA in 92% of pretreatment melanoma plasma samples (green, n=12) at a specificity of 97% relative to held-out noncancerous controls (blue, n=38). Phoenix detected ctDNA in 84% of postreatment plasma samples (n=38, yellow), indicating full ctDNA clearance in 7/38 samples.

Abstract 33 Figure 5

ctDNA response to checkpoint blockade after 3 weeksSerial plasma samples were taken from patients on checkpoint blockade (nivolumab alone or with ipilimumab). ctDNA burden was measured as detection rate among post-filter candidate SNVs and compared to a 97% specificity boundary among a panel of healthy controls. Phoenix detects a response to checkpoint blockade, measured as a decrease in ctDNA detection rate, as early as 3 weeks as shown in 3 patients (MSK-38, MSK-40, MSK 42).

Acknowledgements Thanks to support from the Conquer Cancer Foundation

Ethics Approval Use of human data in this study was approved by Memorial Sloan Kettering’s IRB, Assurance Number FWA0000499


  1. Zhang Q, Luo J, et al. Prognostic and predictive impact of circulating tumor DNA in patients with advanced cancers treated with immune checkpoint blockade. Cancer Discov 2020 pp: CD-20-0047. doi:10.1158/2159-8290.CD-20-0047

  2. Bratman SV, Yang SYC., Iafolla MAJ, et al. Personalized circulating tumor DNA analysis as a predictive biomarker in solid tumor patients treated with pembrolizumab. Nat Cancer (2020).

  3. Zviran A, Schulman RC, Shah M, et al. Genome-wide cell-free DNA mutational integration enables ultra-sensitive cancer monitoring. Nat Med 2020;26(7):1114–1124. doi:10.1038/s41591-020-0915-3

  4. Adalsteinsson VA, Ha G, Freeman SS, et al. Scalable whole-exome sequencing of cell-free DNA reveals high concordance with metastatic tumors. Nat Commun2017;8(1):1324. Published 2017 Nov 6. doi:10.1038/s41467-017-00965-y

This is an open access article distributed in accordance with the Creative Commons Attribution 4.0 Unported (CC BY 4.0) license, which permits others to copy, redistribute, remix, transform and build upon this work for any purpose, provided the original work is properly cited, a link to the licence is given, and indication of whether changes were made. See:

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.