Early circulating tumor DNA dynamics and clonal selection with palbociclib and fulvestrant for breast cancer

CDK4/6 inhibition substantially improves progression-free survival (PFS) for women with advanced estrogen receptor-positive breast cancer, although there are no predictive biomarkers. Early changes in circulating tumor DNA (ctDNA) level may provide early response prediction, but the impact of tumor heterogeneity is unknown. Here we use plasma samples from patients in the randomized phase III PALOMA-3 study of CDK4/6 inhibitor palbociclib and fulvestrant for women with advanced breast cancer and show that relative change in PIK3CA ctDNA level after 15 days treatment strongly predicts PFS on palbociclib and fulvestrant (hazard ratio 3.94, log-rank p = 0.0013). ESR1 mutations selected by prior hormone therapy are shown to be frequently sub clonal, with ESR1 ctDNA dynamics offering limited prediction of clinical outcome. These results suggest that early ctDNA dynamics may provide a robust biomarker for CDK4/6 inhibitors, with early ctDNA dynamics demonstrating divergent response of tumor sub clones to treatment.

There are however caveats to the study which largely centre around discussion of clonality and subclonality based on plasma variant allele frequency -this is transparent and highlighted in the discussion and text. The inferences made surrounding ESR status and response to treatment require elaboration and the authors should highlight in the introduction that they were limited in terms of the quantity of plasma/cell free DNA they could analyse compared to other ctDNA studies (0.25ml plasma or 1.3ng cell-free DNA) which could effect limit of detection.

PIK3CA clonality
The authors track variants in the PIK3CA gene on the basis that these variants are predominantly truncal in breast cancer, furthermore they use the likely truncal nature of PIK3CA to suggest that ESR variants are largely subclonal. Based on available data (e.g. TCGA) could the authors describe and quantify the frequency of subclonal PIK3CA variants in the patient demographic analysed in this study and document this in the text? This is important to interpret the claims made in the manuscript.
Comparing CDR15 between clonal and subclonal variants considering assay limit of detection The authors draw attention to the fact that variants in ESR1 are typically subclonal whereas PIK3CA is clonal and that the summed MAF for PIK3CA variants is more than that for ESR1 variants. Given that the CDR is a ratio of day 1 copies per ml to day 15 copies per ml and given that the ESR1 variants are subclonal and consequently exhibit lower starting MAF -are they more likely to fall below limit of detection of the ddPCR assay than PIK3CA variants? If so is the comparison between subclonal and clonal CDR15 ratios robust given that subclonal variants are more likely to exhibit CDR15s of 0 by falling below LOD? This question is applicable to data presented in figure 2e (potentially figure 2d but it appears no CDR15 fell to 0) and Figure 5b? Could this limitation have effected the lack of prognostic information conferred by ESR1 CDR15 in figure 8? The authors opinion on this would be very helpful.
Since there was a linear relationship between CDR15 with both PIK3CA and ESR1 in most of the cases with dual PIK3CA and ESR1 variants (figure 4d) can the authors analyse the cases where ESR1 CDR15 fell to 0 and determine what the expected mutant ESR1 copies per ml would have been in these 6 cases if a linear relationship had existed here as well. Using this data and considering cell free DNA genomic equivalents analysed at D15 in these cases -what is the likelihood ESR1 would have been detected within the context of a linear fall in ESR1 CDR15 i.e. is this a true non-linear fall in ESR1 levels at day 15 in these 6 cases or a LOD issue at low MAFs given cfDNA genomic equivalents analysed?

Minor
It is difficult to see the median lines on the dot-plots presented in supplementary Figure 2 and 5 It is interesting that wild type PIK3CA/ESR molecules also decreased in response to treatment at day 15. Was this only in the Palbociclib treated group? Could the authors possibly comment on potential reasons why this would be observed? Could the authors describe the possible clinical role for using ctDNA within the context they outline within the manuscript. Would they consider cessation of treatment at day 15 for patients without evidence of a ctDNA response to the therapeutic? Or do they envisage use of this technology in early-phase studies as a a pharmacodynamic marker of response to novel agents?
Could the authors make it clearer that the survival analysis (using Harrell's C) presumably constitutes a training exercise and therefore requires a validation cohort to ascertain the derived CDR15 cut-off's utility in a predictive context?
Can the authors provide supplementary data regarding cell-free DNA quantity (ng) extracted and analysed for each patient at each time point?
What do the two bar charts in supplementary figure 4 represent, can the legend be clearer?
Reviewer #2 (Remarks to the Author): O'Leary et al describe an interesting analysis of mutant copies of PIK3CA (often clonal) and ESR1 (often subclonal) in cell-free DNA after 2 weeks of treatment with fulvestrant +/-palbociclib in the PALOMA-3 trial. They find that greater decrease in mutant copies of PIK3CA at this time point is prognostic for improved progression-free survival. They also find greater decrease of mutant copies of ESR1 (relative to PIK3CA) in patients treated with fulvestrant alone, but this decrease did not predict improved progression-free survival; they postulate that ESR1 subclones may be more sensitive to fulvestrant than the overall tumor, but being subclonal, this differential sensitivity may have little impact on patient outcome. Overall, this work adds significant value to the growing literature supporting using ctDNA metrics to prognosticate and predict response in advanced cancer.

Major Comments
The biggest question that emerged from this study that remains unanswered is whether CDR15 would add value to other clinical predictors of PFS, and whether it would be a better predictor than baseline copies of mutant PIK3CA (a measure not reported in this paper). I would like to see a multivariate analysis including baseline copies of PIK3CA, CDR15 of PIK3CA, disease-free interval, number of previous lines of endocrine treatment, and visceral vs bone-only disease. If CDR15 loses significance in such an analysis, this would be important to know and would temper enthusiasm over its future clinical utility. Nonetheless, understanding the underlying biology would be important. Presumably CDR15 correlates with baseline copies (e.g. it is easier to have a lower ratio if you started higher) -could it be baseline copies that matters? Alternatively, does CDR15 (and/or baseline copies) correlate with other important predictors of disease aggressiveness as outlined in clinical predictors above (e.g. one could imagine that ctDNA dynamics might be dependent on site of metastatic disease)? These are clinically relevant questions that could be explored with the existing data.

Minor Comments
Abstract -It is misleading to say used plasma samples from 455 patients for this analysis, as for the main results/conclusions only used patients who had the tested mutation as well as matched plasma samples at baseline and cycle 1 day 15. Would therefore report in the abstract N=73 (PIK3CA) and N=65 (ESR1).
-Should say explicitly what about PIK3CA ctDNA dynamics specifically predicts what outcome in the abstract -i.e. greater relative decrease in number of mutant PIK3CA copies on day 15 of treatment predicts increased progression-free survival (rather than "PIK3CA ctDNA dynamics after 15 days treatment predicts outcome"). I also find it misleading to use the hazard ratio of 4.92 here, given that that was from the discovery cohort, and would either divide the cohort into discovery and validation and report HR from validation cohort, or simply the median CDR15 as the divider (see comments on results).

Results
-Section 1: How did you determine that the second PIK3CA mutation in the 4 cases with 2 mutations was subclonal vs clonal? Was the VAF for the second mutation significantly lower than for the first in all 4 cases? -Section 3: Please report what the tertiles (and median) of CDR15 were. -Section 3: Reasonable to report what the "ideal" cut-off was per supplementary figure 7, but only reasonable to report median PFS in each group if you are using a validation (not discovery) cohort. So either divide the cohort into discovery (to find the cut-off) and validation (to find the median PFS in each group), or just report median PFS for above and below the median. Would change Fig  3B accordingly (i.e. could present both discovery and validation cohorts' K-M curves, or change to K-M curve based on median). See comments on abstract.
-Section 3: I am puzzled why you did not perform the same tertile and median K-M analysis of the patients treated with fulvestrant alone -there were patients treated with fulvestrant alone who had a CDR15 > 1, and others who had a CDR15 < 0.5, so it's not empirically obvious why there would be no differentiation there. Should report that there was no difference by tertile or median for patients treated with fulvestrant alone if that is the case. We thank the reviewers for highly constructive comments on the manuscript. We have addressed these in full. We provide a point by point rebuttal of the reviewers' comments below.

Reviewer #1 (Remarks to the Author):
Early circulating tumor DNA dynamics and clonal selection with palbociclib and fulvestrant for advanced breast cancer O'leary et al. present a methodologically robust analysis of plasma samples collected within the context of a phase III clinical trial investigating the combination of palbociclib and fulvestrant for advanced breast cancer. Of interest is the finding that circulating tumor DNA dynamics 15 days following commencement of anti-proliferative treatment has potential to predict progression free survival. The authors use the available data from the PALOMA study to determine a threshold for change in circulating tumor DNA concentration 15 days following commencement of palbociclib that best predicts relapse free survival. This work is novel and could be validated in a larger cohort as a marker of palbociclib response. If validated this approach could be used as a pharmacodynamic marker for early-phase studies of cytostatic therapies. The finding that assumed clonal variants are useful prognostic indicators in contrast to subclonal variants which do not provide prognostic information is also helpful for the field.
There are however caveats to the study which largely centre around discussion of clonality and subclonality based on plasma variant allele frequency -this is transparent and highlighted in the discussion and text. The inferences made surrounding ESR status and response to treatment require elaboration and the authors should highlight in the introduction that they were limited in terms of the quantity of plasma/cell free DNA they could analyse compared to other ctDNA studies (0.25ml plasma or 1.3ng cell-free DNA) which could affect limit of detection.
We thank the reviewer for their positive comments and recommendations for expansion. As suggested we have enriched our analysis of the TCGA data for polyclonal PIK3CA mutations and performed further experiments to further establish the potential role the assay limit of detection may have played in our results. These are detailed point by point below.

PIK3CA clonality
The authors track variants in the PIK3CA gene on the basis that these variants are predominantly truncal in breast cancer, furthermore they use the likely truncal nature of PIK3CA to suggest that ESR variants are largely subclonal. Based on available data (e.g. TCGA) could the authors describe and quantify the frequency of subclonal PIK3CA variants in the patient demographic analysed in this study and document this in the text? This is important to interpret the claims made in the manuscript.
We accessed the TCGA via cBioportal and downloaded the 2015 TCGA, isolating the ER+/HER2-subset as most comparable to the patient cohort in PALOMA-3. Examining the data for the 4 PIK3CA mutations assayed in our study we find that multiple PIK3CA mutations are rare in primary disease (1/488, 0.2%, accounting for 1/152, 0.7% of all H1047R, H1047L, E542K, E545K mutations), supportive of the hypothesis that PIK3CA mutations are generally truncal. These data are now detailed in the results section and we return to it in the discussion but caveat this observation with the fact that the landscape of PIK3CA mutations in endocrineresistant disease is less well-described. The cases of multiple PIK3CA mutations are too small in number in both the TCGA and PALOMA-3 to enable meaningful statistical analysis. We now include a detailed breakdown of the observed PALOMA-3 polyclonal PIK3CA mutations as supplementary figure 2.
Comparing CDR15 between clonal and subclonal variants considering assay limit of detection The authors draw attention to the fact that variants in ESR1 are typically subclonal whereas PIK3CA is clonal and that the summed MAF for PIK3CA variants is more than that for ESR1 variants. Given that the CDR is a ratio of day 1 copies per ml to day 15 copies per ml and given that the ESR1 variants are subclonal and consequently exhibit lower starting MAF -are they more likely to fall below limit of detection of the ddPCR assay than PIK3CA variants? If so is the comparison between subclonal and clonal CDR15 ratios robust given that subclonal variants are more likely to exhibit CDR15s of 0 by falling below LOD? This question is applicable to data presented in figure 2e (potentially figure 2d but it appears no CDR15 fell to 0) and Figure 5b? Could this limitation have effected the lack of prognostic information conferred by ESR1 CDR15 in figure 8? The authors opinion on this would be very helpful.
We have reanalysed the data to investigate whether the analysis presented in Figure 2e is affected by baseline level, i.e. to address whether ESR1 mutations are more likely have a CDR 15 s of 0, as they frequently start at a lower baseline level. There is no difference in the baseline ESR1 mutation level between those patients that had a CDR 15 of 0, and those with CDR 15 above zero. Therefore there was no suggestion that the increased suppression of ESR1 mutations observed in figure 2e is due to LOD issues, Conversely, we do observe a statistically significant difference for PIK3CA (p=0.029 Mann-Whitney, now supplementary figure 16).
We have also reanalysed the data to demonstrate that baseline mutant copies/ml was not correlated with CDR 15 , data now included as Overall these analyses demonstrate convincingly that the differences observed in figure 2e are not due to analytical issues around detecting mutations, but are due to biological differences. As there is no evidence that analytical issues affect ESR1 CDR 15 , this would not limit the potential for ESR1 CDR 15 to predict PFS as shown in supplementary figure 9. p = 0.029 p = 0.055 Since there was a linear relationship between CDR15 with both PIK3CA and ESR1 in most of the cases with dual PIK3CA and ESR1 variants (figure 4d) can the authors analyse the cases where ESR1 CDR15 fell to 0 and determine what the expected mutant ESR1 copies per ml would have been in these 6 cases if a linear relationship had existed here as well.
Using this data and considering cell free DNA genomic equivalents analysed at D15 in these cases -what is the likelihood ESR1 would have been detected within the context of a linear fall in ESR1 CDR15 i.e. is this a true non-linear fall in ESR1 levels at day 15 in these 6 cases or a LOD issue at low MAFs given cfDNA genomic equivalents analysed?
We thank the reviewer for raising this excellent point.. To investigate the possible of effect limit of detection we analysed further day 15 plasma. There are 8 cases, not 6, two pairs lie on top of each other, now corrected in the text. Of these 8 samples, 4/8 cases are above the level of detection and are a true non-linear fall, whilst 4/8 are below the level of detection. For the samples with ESR1 below the limit of detection we have analysed further plasma, to bring ESR1 within the lower the limit of detection. 6 mutations were tested, as 2 of the 4 cases were polyclonal. An ESR1 mutation was detected in only one of these cases, suggesting that falling below the limit of detection is not a major limiting factor in this analysis. These additional analyses are now included in the results, with the below tables included as supplementary figures 13 and 14. The theoretical limit of detection calculation is included in the methods. The issue of limit of detection has also been included in the discussion to emphasise caution in the subclonality observations.

Minor
It is difficult to see the median lines on the dot-plots presented in supplementary Figure  2 and 5 We have corrected this.
It is interesting that wild type PIK3CA/ESR molecules also decreased in response to treatment at day 15. Was this only in the Palbociclib treated group? Could the authors possibly comment on potential reasons why this would be observed? Could the authors describe the possible clinical role for using ctDNA within the context they outline within the manuscript. Would they consider cessation of treatment at day 15 for patients without evidence of a ctDNA response to the therapeutic? Or do they envisage use of this technology in early-phase studies as a a pharmacodynamic marker of response to novel agents?

We have now included an analysis of CDR
The optimal way to use CDR 15 in the clinic will be the subject of future trials. We plan trials of the addition of extra treatment in patients without a ctDNA response. We agree that ctDNA response could be used to screen agents in early phase trials, as a potentially highly robust surrogate endpoint. These points have been expanded in the discussion.
Could the authors make it clearer that the survival analysis (using Harrell's C) presumably  constitutes a training exercise and therefore requires a validation cohort to ascertain the derived CDR15 cut-off's utility in a predictive context?
We highlight this in the discussion.
Can the authors provide supplementary data regarding cell-free DNA quantity (ng) extracted and analysed for each patient at each time point?

This is now attached as a supplementary datasheet Excel file.
What do the two bar charts in supplementary figure 4 represent, can the legend be clearer?
We apologise for not making the legend clear, and have clarified in the legend. Each bar represents a single patient. These illustrate the contribution of each of the individual ESR1 mutations to the total ESR1 mutant copies/ml in those patients who in the baseline samples were identified as having more than one ESR1 mutation. The lower panel is an enlargement of the upper to enable a clearer look at the patients with lower mutation abundance.

Reviewer #2 (Remarks to the Author):
O'Leary et al describe an interesting analysis of mutant copies of PIK3CA (often clonal) and ESR1 (often subclonal) in cell-free DNA after 2 weeks of treatment with fulvestrant +/-palbociclib in the PALOMA-3 trial. They find that greater decrease in mutant copies of PIK3CA at this time point is prognostic for improved progression-free survival. They also find greater decrease of mutant copies of ESR1 (relative to PIK3CA) in patients treated with fulvestrant alone, but this decrease did not predict improved progressionfree survival; they postulate that ESR1 subclones may be more sensitive to fulvestrant than the overall tumor, but being subclonal, this differential sensitivity may have little impact on patient outcome. Overall, this work adds significant value to the growing literature supporting using ctDNA metrics to prognosticate and predict response in advanced cancer.
We thank the reviewer for their positive assessment and address their concerns point by point below.

Major Comments
The biggest question that emerged from this study that remains unanswered is whether CDR15 would add value to other clinical predictors of PFS, and whether it would be a better predictor than baseline copies of mutant PIK3CA (a measure not reported in this paper). I would like to see a multivariate analysis including baseline copies of PIK3CA, CDR15 of PIK3CA, disease-free interval, number of previous lines of endocrine treatment, and visceral vs bone-only disease. If CDR15 loses significance in such an analysis, this would be important to know and would temper enthusiasm over its future clinical utility. Nonetheless, understanding the underlying biology would be important. Presumably CDR15 correlates with baseline copies (e.g. it is easier to have a lower ratio if you started higher) -could it be baseline copies that matters? Alternatively, does CDR15 (and/or baseline copies) correlate with other important predictors of disease aggressiveness as outlined in clinical predictors above (e.g. one could imagine that ctDNA dynamics might be dependent on site of metastatic disease)? These are clinically relevant questions that could be explored with the existing data.
We thank the reviewer for these points that we have addressed in full matched plasma samples at baseline and cycle 1 day 15. Would therefore report in the abstract N=73 (PIK3CA) and N=65 (ESR1).
We have removed the "455" from the abstract as requested.
-Should say explicitly what about PIK3CA ctDNA dynamics specifically predicts what outcome in the abstract -i.e. greater relative decrease in number of mutant PIK3CA copies on day 15 of treatment predicts increased progression-free survival (rather than "PIK3CA ctDNA dynamics after 15 days treatment predicts outcome").
This has now been changed in accordance with the advice from the reviewer and now reads "we show that relative change in PIK3CA ctDNA level after 15 days treatment strongly predicts PFS on palbociclib and fulvestrant (hazard ratio 3.94, p = 0.0013). " I also find it misleading to use the hazard ratio of 4.92 here, given that that was from the discovery cohort, and would either divide the cohort into discovery and validation and report HR from validation cohort, or simply the median CDR15 as the divider (see comments on results).
As requested, we now report the data according to the median in the abstract, and lead with this in the results section.

Results
-Section 1: How did you determine that the second PIK3CA mutation in the 4 cases with 2 mutations was subclonal vs clonal? Was the VAF for the second mutation significantly lower than for the first in all 4 cases?
The second VAF was significantly lower in all cases. These data are now presented in full as part of supplementary figure 2.
-Section 3: Please report what the tertiles (and median) of CDR15 were.
The median CDR 15 is now reported in the text for PIK3CA and ESR1 for the palbociclib and fulvestrant patients, and in the relevant supplementary figure legends for the placebo and fulvestrant patients.
-Section 3: Reasonable to report what the "ideal" cut-off was per supplementary figure 7, but only reasonable to report median PFS in each group if you are using a validation (not discovery) cohort. So either divide the cohort into discovery (to find the cut-off) and validation (to find the median PFS in each group), or just report median PFS for above and below the median. Would change Fig 3B accordingly (i.e. could present both discovery and validation cohorts' K-M curves, or change to K-M curve based on median). See comments on abstract.
As mentioned in an earlier response, the data are now presented according to the median with further clarification highlighting this as a training set.
-Section 3: I am puzzled why you did not perform the same tertile and median K-M analysis of the patients treated with fulvestrant alone -there were patients treated with fulvestrant alone who had a CDR15 > 1, and others who had a CDR15 < 0.

Figures
- Fig 1A is cut off at the top We thank the reviewer for pointing this out and have corrected it.
-Set x-axis of Fig 1C to 1.0 (i.e. bars below the x-axis would then reflect decrease in ctDNA and bars above would reflect increase, like a waterfall plot) We thank the reviewer for this suggestion but for the ESR1 plot this makes the data somewhat difficult to interpret visually. To emphasize the fall from the value