Notch signaling maintains a progenitor-like subclass of hepatocellular carcinoma

Reviewer #1 (Public review):

Summary:

The significance of Notch in liver cancer has been inconsistently described to date. The authors conduct a PDX screen using JAG1 ab and identify 2 sensitive tumor models. Further characterization with bulk RNA seq, scRNA seq, and ATAC seq of these tumors was performed.

Strengths:

The reliance on an extensive panel of PDXs makes this study more definitive than prior studies.

Gene expression analyses seem robust.

Identification of a JAG1-dependent signature associated with hepatocyte differentiation is interesting.

Weaknesses:

The introduction is rather lengthy and not entirely accurate. HCC is a single cancer type/histology. There may be variants of histology (allusion to "mixed-lineage" is inaccurate as combined HCC-CCa are not conventionally considered HCC and are not treated as HCC in clinical practice as they are even excluded from HCC trials), but any cancer type can have differences in differentiation. Just state there are multiple molecular subtypes of this disease.

There is minimal data on the PDXs, despite this being highlighted throughout the text. Clinical and possibly some molecular characterization of these cancers should be provided. It is also odd that the authors include only 35 HCC and then a varied sort of cancer histologies, which is peculiar given their prior statements regarding the heterogeneity of HCC.

"super-responder" is not a meaningful term, I would eliminate this use as it has no clinical or scientific convention that I am aware of.

The "expansion" of the PDX screen is poorly described. Why weren't these PDXs included in the first screen? This is quite odd as the responses in the initial screen were underwhelming. What was the denominator number of all PDXs that were assessed for JAG1 and NOTCH2 expression? This is important as it clarifies how relevant JAG1 inhibition would be to an unselected HCC population.

Was there some kind of determination of the optimal dose or dose dependency for the JAG1 ab? The original description of the JAG1 ab was in mouse lungs, not malignant or liver cells. In addition, supplementary Figure 2D is missing. There needs to be data provided on the specificity of the human-specific JAG1 ab and the anti-NOTCH2 ab. I'm not familiar with these ab, and if they are not publicly accessible reagents, more transparency on this is needed. In addition, given the reliance of the entire paper on these antibodies, I would recommend orthogonal approaches (either chemical or genetic) to confirm the sensitivity and insensitivity of select PDXs to Notch inhibition.

scRNA-seq data seems to add little to the paper and there is no follow-up of the findings. Are the low-expressing JAG1 cells eventually enriched in treated tumors contributing to disease recurrence?

The discussion should be tempered. The finding of only 2 PDXs that are sensitive out of 45+ tumors treated or selected for indicates that JAG1/NOTCH2 inhibition is likely only effective in rare HCC.

Reviewer #2 (Public review):

Summary:

The authors used a large panel of hepatocellular carcinoma patient-derived xenograft models to test the hypothesis that the developmental dependence of the liver on Jagged1-Notch2 signaling is retained in at least a subset of hepatocellular carcinomas. This led to the identification of two models that were extraordinarily sensitive to well-characterized, specific inhibitory antibodies against Jagged1 or Notch2. Based on additional analyses in these in vivo models, the authors provide compelling evidence that the response is due to the inhibition of human Notch2 and human Jagged1 on tumor cells and that this inhibition leads to a change in gene expression from a progenitor-like state to a hepatocyte-like state accompanied by cell cycle arrest. This change in cell state is associated with up-regulation of HNF4a and CEBPB and increased accessibility of predicted HNF4a and CEBPB genomic binding sites, accompanied by loss of accessibility to sequences predicted to bind TFs linked to multipotent liver progenitors. The authors put forth a plausible model in which inhibition of Notch2 downregulates transcriptional repressors of the Hairy/Enhancer of Split family, leading to increased expression of CEBPB and changes in gene expression that drive hepatocyte differentiation.

Strengths:

The strengths of the paper include the breadth of the preclinical screen in PDX models (which may be of an unprecedented size as preclinical trials go), the high quality of the well-characterized antibodies used as therapeutics and as biological perturbagens, the quality of the data and data analysis, and the authors balanced discussion of the strengths and weaknesses of their findings.

Weaknesses:

The principal weakness is the inability to clearly demonstrate the "translatability" of the PDX findings to primary human hepatocellular carcinoma.

Additional Comments:

Hepatocellular carcinoma is increasing in frequency and is difficult to treat; cure is only possible through early diagnosis and surgery, often in the form of liver transplantation. It is also a common cancer, and so even if only 5% of tumors (a value based on the frequency of super-responders in this preclinical trial) fall into the Jagged1-Notch2 group defined by Seidel et al., the development of an effective therapy for this subgroup would be a very important advance. The chief limitation of their work is that it stops short of identifying primary human hepatocellular carcinomas that correspond to the super-responder PDX models. It can be hoped that their intriguing observations will spur work aimed at filling this gap

There are several other loose ends. An unusual feature of this model is that both Jagged 1 and Notch2 are expressed in the same cells, and even in the same individual cells. In developmental systems, the expression of ligands and receptors in the same cell generally produces receptor inhibition rather than activation, a phenomenon described as cis inhibition. Their super-responder tumor models appear to break this rule, and how and why this is so remains to be understood. A follow-up question is what explains the observed heterogeneity in tumor cells, both at the level of Notch2 activation and scRNAseq clustering, and whether these different cell states are static or interchangeable.

Another unanswered issue pertains to the nature of the tumor response to Notch signaling blockade, which appears to be mainly cell cycle arrest. There are a number of human tumors with cell autonomous Notch signaling due to gain of function Notch receptor mutations that also respond to Notch blockade with cell cycle arrest, such as T cell acute lymphoblastic leukemia (T-ALL). In general, clinical trials of pan-Notch inhibitors such as gamma-secretase inhibitors have been disappointing in such tumors, perhaps reflecting a limitation of treatments with significant toxicity that do not kill tumor cells directly. It could be argued that this limitation will be mitigated by the apparently excellent safety profile of Notch2 blocking antibody, which perhaps could be administered for a sustained period, akin to the use of tyrosine kinase inhibitors in chronic myeloid leukemia---but this remains to be determined.

A minor comment is reserved for the statement in the discussion that "In chronic myelomonocytic leukemia, which results from an inactivating mutation in the y-secretase complex component nicastrin, Notch signaling has a tumor suppressive function, that is mediated through direct repression of CEBPA and PU.1 by HES1 (Klinakis et al., 2011)". Thousands of cases of CMML and related myeloid tumors have been subjected to whole exome and even whole genome sequencing without the identification of Notch signaling pathway mutations. Thus, an important tumor suppressive role for Notch-mediated through HES1 in myeloid tumors is not proven.

Reviewer #3 (Public review):

Summary:

Notch is active in HCC, but generally not mutated. The authors use a JAG1-selective blocking antibody in a large panel of liver cancer patient-derived xenograft models. They find JAG-dependent HCCs, and these are aggressive and proliferative. Notch inhibition induces cycle arrest and promotes hepatocyte differentiation, through upregulation of CEBPA expression and activation of existing HNF4A, mimicking normal developmental programs.

The authors use aJ1.b70, a potent and selective therapeutic antibody that inhibits JAG1 against PDX models. They tested over 40 PDX models and found a handful of super-responders to single-agent inhibition. In LIV78 and Li1035 cancer cells, NOTCH2 was expressed and required, in contrast to NOTCH1. RNA-seq showed that the responsive HCCs resembled the S2 transcriptional class of HCCs, which were enriched for Notch-dependent models. They conclude that these dependent tumors have transcriptomes that resemble a hybrid progenitor cell expressing FGF9 and GAS7. Inhibition was able to induce hepatocyte differentiation away from a NOTCH-driven progenitor program. scRNA-seq analysis showed a large population of NOTCH-JAG expressing cells but also showed that there are cells that did not. Not surprisingly, NOTCH2 inhibition leads to increased CEBPA and HNF4A transcriptional activity, which are standard TFs in hepatocytes.

Strengths:

The paper provides useful information about the frequency of HCCs and CCA that respond to NOTCH inhibition and could allow us to anticipate the super-responder rate if these antibodies were actually used in the clinic. The inhibitor tools are highly specific, and provide useful information about NOTCH activities in liver cancers. The large number of PDXs and the careful transcriptomic analyses were positives about the study.

Weaknesses:

The paper is mostly descriptive.

Author response:

Public Reviews:

Reviewer #1 (Public review):

Summary:

The significance of Notch in liver cancer has been inconsistently described to date. The authors conduct a PDX screen using JAG1 ab and identify 2 sensitive tumor models. Further characterization with bulk RNA seq, scRNA seq, and ATAC seq of these tumors was performed.

Strengths:

The reliance on an extensive panel of PDXs makes this study more definitive than prior studies.

Gene expression analyses seem robust.

Identification of a JAG1-dependent signature associated with hepatocyte differentiation is interesting.

Weaknesses:

The introduction is rather lengthy and not entirely accurate. HCC is a single cancer type/histology. There may be variants of histology (allusion to "mixed-lineage" is inaccurate as combined HCC-CCa are not conventionally considered HCC and are not treated as HCC in clinical practice as they are even excluded from HCC trials), but any cancer type can have differences in differentiation. Just state there are multiple molecular subtypes of this disease.

We will shorten the Introduction, in part by eliminating the discussion of histological variation in HCC and focusing on the molecular classifications.

There is minimal data on the PDXs, despite this being highlighted throughout the text. Clinical and possibly some molecular characterization of these cancers should be provided. It is also odd that the authors include only 35 HCC and then a varied sort of cancer histologies, which is peculiar given their prior statements regarding the heterogeneity of HCC.

We agree that clinical and molecular characterizations of the PDX models would be helpful and will follow up with the relevant contract research organization to determine what characterization is available.

Regarding the liver cancer PDX panel, we suggest that a major strength of the manuscript is the large number of HCC models that were tested (the reviewer also notes the importance of the “extensive” panel); thus, we are a bit confused by the reference to “only 35 HCC”. To clarify the choice of models in the PDX screen, it may help to put the screen in historical perspective as the project unfolded. In retrospect, our preliminary efficacy studies using only two HCC models were fortunate to identify the highly sensitive model, LIV78. To go beyond the simple diagnostic hypothesis that focused on Jag1, Notch2 and Hes1 expression, we took an unbiased approach to discover features linked to Notch dependence. This approach meant running an efficacy screen in all liver cancer models that were up and running at our chosen research organization, without biased selection criteria. That set of models is what is represented in the “pre-clinical screen” in Fig. 1B

"super-responder" is not a meaningful term, I would eliminate this use as it has no clinical or scientific convention that I am aware of.

We were aware of the interchangeable terms of “exceptional-“ or “super-responder” and prefer to leave this language in the text. Some references are as follows:

● Prasad et al., Characteristics of exceptional or super responders to cancer drugs. Mayo Clinic Proceedings, 2015.

● NCI Press Release 2020: https://www.cancer.gov/news-events/press-releases/2020/cancer-exceptional-responders-study-genetic-alterations-may-contribute

● NIH Info: https://www.nih.gov/news-events/nih-research-matters/understanding-exceptional-responders-cancer-treatment

● “What is a Super Responder? Bradley Jones, Cancer Today, June 26, 2020.

● “What is a Super Responder?” AACR. https://www.aacr.org/patients-caregivers/progress-against-cancer/what-is-a-super-responder/

The "expansion" of the PDX screen is poorly described. Why weren't these PDXs included in the first screen? This is quite odd as the responses in the initial screen were underwhelming. What was the denominator number of all PDXs that were assessed for JAG1 and NOTCH2 expression? This is important as it clarifies how relevant JAG1 inhibition would be to an unselected HCC population.

We will revise the writing here to clarify as requested. For now, we can hopefully clarify by building on the historical context described above. As the reviewer notes and as we describe in the text, the in vivo screen revealed only a modest JAG1 dependence. The screen also highlighted that LIV78 was exceptional, and we wanted to understand why. Hypothesizing that the expression of progenitor markers in LIV78 were important for understanding its JAG1 dependence, we identified four additional models at other contract research organizations. It is this set of four that comprises the “expansion” cohort.

Was there some kind of determination of the optimal dose or dose dependency for the JAG1 ab? The original description of the JAG1 ab was in mouse lungs, not malignant or liver cells. In addition, supplementary Figure 2D is missing. There needs to be data provided on the specificity of the human-specific JAG1 ab and the anti-NOTCH2 ab. I'm not familiar with these ab, and if they are not publicly accessible reagents, more transparency on this is needed. In addition, given the reliance of the entire paper on these antibodies, I would recommend orthogonal approaches (either chemical or genetic) to confirm the sensitivity and insensitivity of select PDXs to Notch inhibition.

First, we note that the anti-human/mouse Jagged1 and Notch2 blocking antibodies used in our study have been extensively characterized as potent and selective and have been widely used outside of our group by the Notch research community (for the human/mouse cross-reactive antibodies, see Wu et al., Nature, 2010 for anti-NOTCH2 and Lafkas et al., Nature 2015 for anti-JAG1). As noted, the antibodies have been used in studies of normal mouse lungs (Lafkas et al.). Please note that the characterization also includes mouse models of primary liver cancer that formed the foundation for the current work (please refer to Huntzicker et al, 2015).

While we show dose responses in Figures 1A and 1D, we have not optimized dosing, for example by determining the minimal drug exposures needed for pharmacodynamic changes (pathway inhibition) and efficacy. For the purposes of this study, we erred on the side of dosing at high concentrations to minimize the risk of false negative responses.

Regarding the specificity of the human-specific anti-JAG1 antibody, which is revealed here for the first time, we apologize that we incorrectly provided a text reference to Supplementary Figure 2D instead of Supplementary Figure 1D. We will revise accordingly. Fig. 1D shows results from a reporter assay demonstrating that the antibody blocks signaling induced by human but not mouse JAG1.

We appreciate the value of orthogonal methods in establishing the credibility of a novel finding. We note that genetic approaches are technically highly challenging in PDX models. Chemically, we could have tested y-secretase inhibitors (GSIs). Our position is that such inhibitors are poor substitutes for the selective antibodies that we employed, at least for addressing the questions that are relevant in this study. Although commonly used to perturb Notch signaling, GSIs target numerous proteins and signaling cascades independent of Notch. Moreover, their use in vivo leads to intestinal and other toxicities, limiting exposure.

scRNA-seq data seems to add little to the paper and there is no follow-up of the findings. Are the low-expressing JAG1 cells eventually enriched in treated tumors contributing to disease recurrence?

We respectfully disagree with this sentiment. The single-cell RNA sequencing dataset revealed the enrichment of hepatocyte-like tumor cells following Notch inhibition. Importantly, this dataset also allowed us to identify transcription factor activities regulating different cell states, which we could not have done otherwise. This understanding in turn was fundamental to develop our hypothesis that Notch inhibition, through derepressing CEBPA expression, allows chromatin engagement of HNF4A and CEPBA and thereby promotes a hepatocyte differentiation program that is not compatible with tumor maintenance.

The discussion should be tempered. The finding of only 2 PDXs that are sensitive out of 45+ tumors treated or selected for indicates that JAG1/NOTCH2 inhibition is likely only effective in rare HCC.

We agree that strong responses to Notch inhibition in the PDX models are rare (~5%) and state as much in both the Results and Discussion sections. We maintain that it is important to put this PDX response frequency into a larger context. First, establishing PDX models---human tumor samples that grow on the flanks of immunocompromised mice---represents a strong selective pressure. In other words, we don’t know precisely how the frequency of responses in this selected set of PDX models may compare to the frequency that would be observed in human patient populations. Second, the magnitude of the response points to important and hitherto unappreciated biology, with blocking JAG1 or NOTCH2 reproducibly inducing regressions in the most sensitive models. Our hope is that the field can build from this study to generate diagnostic tools that identify sensitive patient tumors, define the true frequency of this patient group within the larger HCC population (even though likely rare), and direct the relevant Notch-based therapeutics to these patients. Within this context, and while noting the rarity of PDX responses, we hope that we have not overstated the case.

Reviewer #2 (Public review):

Summary:

The authors used a large panel of hepatocellular carcinoma patient-derived xenograft models to test the hypothesis that the developmental dependence of the liver on Jagged1-Notch2 signaling is retained in at least a subset of hepatocellular carcinomas. This led to the identification of two models that were extraordinarily sensitive to well-characterized, specific inhibitory antibodies against Jagged1 or Notch2. Based on additional analyses in these in vivo models, the authors provide compelling evidence that the response is due to the inhibition of human Notch2 and human Jagged1 on tumor cells and that this inhibition leads to a change in gene expression from a progenitor-like state to a hepatocyte-like state accompanied by cell cycle arrest. This change in cell state is associated with up-regulation of HNF4a and CEBPB and increased accessibility of predicted HNF4a and CEBPB genomic binding sites, accompanied by loss of accessibility to sequences predicted to bind TFs linked to multipotent liver progenitors. The authors put forth a plausible model in which inhibition of Notch2 downregulates transcriptional repressors of the Hairy/Enhancer of Split family, leading to increased expression of CEBPB and changes in gene expression that drive hepatocyte differentiation.

Strengths:

The strengths of the paper include the breadth of the preclinical screen in PDX models (which may be of an unprecedented size as preclinical trials go), the high quality of the well-characterized antibodies used as therapeutics and as biological perturbagens, the quality of the data and data analysis, and the authors balanced discussion of the strengths and weaknesses of their findings.

Weaknesses:

The principal weakness is the inability to clearly demonstrate the "translatability" of the PDX findings to primary human hepatocellular carcinoma.

We agree that translatability has not been fully addressed. As noted in our response to Reviewer 1, our hope is that the field can build from this study to generate diagnostic tools that identify sensitive patient tumors, define the true frequency of this patient group within the larger HCC population, and direct the relevant Notch-based therapeutics to these patients. We remain encouraged by the strength of the response in the sensitive models.

Additional Comments:

Hepatocellular carcinoma is increasing in frequency and is difficult to treat; cure is only possible through early diagnosis and surgery, often in the form of liver transplantation. It is also a common cancer, and so even if only 5% of tumors (a value based on the frequency of super-responders in this preclinical trial) fall into the Jagged1-Notch2 group defined by Seidel et al., the development of an effective therapy for this subgroup would be a very important advance. The chief limitation of their work is that it stops short of identifying primary human hepatocellular carcinomas that correspond to the super-responder PDX models. It can be hoped that their intriguing observations will spur work aimed at filling this gap.

There are several other loose ends. An unusual feature of this model is that both Jagged 1 and Notch2 are expressed in the same cells, and even in the same individual cells. In developmental systems, the expression of ligands and receptors in the same cell generally produces receptor inhibition rather than activation, a phenomenon described as cis inhibition. Their super-responder tumor models appear to break this rule, and how and why this is so remains to be understood. A follow-up question is what explains the observed heterogeneity in tumor cells, both at the level of Notch2 activation and scRNAseq clustering, and whether these different cell states are static or interchangeable.

We enthusiastically agree that these are fascinating questions, worthy of further study. As noted, the majority of tumor cells express both ligand and receptor and seem to be “on” for Notch signaling. We have not been able to determine whether the signal is induced in a cell autonomous or non-autonomous manner (or both). As the reviewer notes, the HCC features we observe are inconsistent with the dogma that has arisen from studies on Notch signaling in developmental contexts.

We do not yet have the experimental data to fully address the second question of what causes the heterogeneity of Notch2 activation and scRNAseq clustering. We speculate that the cell states may be dynamic, which would be consistent with the changes in cell populations observed after antibody treatment.

Another unanswered issue pertains to the nature of the tumor response to Notch signaling blockade, which appears to be mainly cell cycle arrest. There are a number of human tumors with cell autonomous Notch signaling due to gain of function Notch receptor mutations that also respond to Notch blockade with cell cycle arrest, such as T cell acute lymphoblastic leukemia (T-ALL). In general, clinical trials of pan-Notch inhibitors such as gamma-secretase inhibitors have been disappointing in such tumors, perhaps reflecting a limitation of treatments with significant toxicity that do not kill tumor cells directly. It could be argued that this limitation will be mitigated by the apparently excellent safety profile of Notch2 blocking antibody, which perhaps could be administered for a sustained period, akin to the use of tyrosine kinase inhibitors in chronic myeloid leukemia---but this remains to be determined.

We agree that a full understanding of the tumor response warrants further investigation. Like the reviewer, we speculate that the improved safety profile of selective antibodies relative to pan-Notch inhibitors may enable greater and sustained therapeutic coverage of Notch inhibition than has been feasible in T-ALL trials. Given that in the sensitive PDX models we observe rapid tumor regressions, not just stasis, it would seem to follow that the mechanism underpinning the tumor response involves more than just cell cycle blockade. Whether tumor shrinkage reflects additional cell death mechanisms or simply tumor cell turnover after cell cycle arrest remains to be determined.

A minor comment is reserved for the statement in the discussion that "In chronic myelomonocytic leukemia, which results from an inactivating mutation in the y-secretase complex component nicastrin, Notch signaling has a tumor suppressive function, that is mediated through direct repression of CEBPA and PU.1 by HES1 (Klinakis et al., 2011)". Thousands of cases of CMML and related myeloid tumors have been subjected to whole exome and even whole genome sequencing without the identification of Notch signaling pathway mutations. Thus, an important tumor suppressive role for Notch-mediated through HES1 in myeloid tumors is not proven.

We agree that our sentence about Notch and CMML does not fit well with the prevalent paradigm established by genome wide sequencing and other methods. We will edit this paragraph accordingly, focusing on Hes1 negative regulation of CEBPA in myeloid fate control and how that shapes our thinking on molecular mechanisms in the Notch-dependent HCCs.

Reviewer #3 (Public review):

Summary:

Notch is active in HCC, but generally not mutated. The authors use a JAG1-selective blocking antibody in a large panel of liver cancer patient-derived xenograft models. They find JAG-dependent HCCs, and these are aggressive and proliferative. Notch inhibition induces cycle arrest and promotes hepatocyte differentiation, through upregulation of CEBPA expression and activation of existing HNF4A, mimicking normal developmental programs.

The authors use aJ1.b70, a potent and selective therapeutic antibody that inhibits JAG1 against PDX models. They tested over 40 PDX models and found a handful of super-responders to single-agent inhibition. In LIV78 and Li1035 cancer cells, NOTCH2 was expressed and required, in contrast to NOTCH1. RNA-seq showed that the responsive HCCs resembled the S2 transcriptional class of HCCs, which were enriched for Notch-dependent models. They conclude that these dependent tumors have transcriptomes that resemble a hybrid progenitor cell expressing FGF9 and GAS7. Inhibition was able to induce hepatocyte differentiation away from a NOTCH-driven progenitor program. scRNA-seq analysis showed a large population of NOTCH-JAG expressing cells but also showed that there are cells that did not. Not surprisingly, NOTCH2 inhibition leads to increased CEBPA and HNF4A transcriptional activity, which are standard TFs in hepatocytes.

Strengths:

The paper provides useful information about the frequency of HCCs and CCA that respond to NOTCH inhibition and could allow us to anticipate the super-responder rate if these antibodies were actually used in the clinic. The inhibitor tools are highly specific, and provide useful information about NOTCH activities in liver cancers. The large number of PDXs and the careful transcriptomic analyses were positives about the study.

Weaknesses:

The paper is mostly descriptive.