Integration of genomic aberrations to predict clinical outcomes for patients with gastroesophageal adenocarcinoma receiving neoadjuvant chemotherapy

Abstract

Background

Esophageal cancer [esophagogastric adenocarcinoma (OGA)] shows heterogeneity at the molecular level, leading to lower efficacy rates and highlighting the need for personalized treatment strategies. We have developed a computational model that uses both mechanistic and statistical approaches to integrate a patient’s genomic aberrations, revealing signaling pathway dysregulation and variable drug response. The model output, Therapy Response Index (TRI), has been used to predict therapeutic outcomes in this study.

Design

TRI’s ability to predict patient outcomes was retrospectively evaluated in a prospectively collected cohort of patients with operable OGA from the UK Oesophageal Cancer Clinical and Molecular Stratification (OCCAMS) consortium receiving neoadjuvant chemotherapy. Stratified random sampling was used to split the data into training and validation subsets. Multivariate Cox proportional hazard and proportional odds models were used to predict survival and pathological response as a function of TRI and clinical thresholds compared with clinical factors.

Results

A total of 270 patients with OGA were selected who had 50× whole genome sequencing carried out on tissue derived from either biopsy or resection. Patients were treated with chemotherapy drugs or regimens according to UK clinical guidelines. The association of TRI with overall survival (OS) was significant above and beyond standard clinical factors (P = 0.0012). A significant association was also observed with disease-free survival (DFS; P = 0.0288). A TRI optimized for tumor regression grade also displayed a significant association (P = 0.0011).

Conclusions

TRI was predictive of OS and DFS beyond clinical factors. These positive results suggest the potential utility of personalized biosimulation-informed therapy selection and that further assessment in prospective clinical studies is warranted.

Highlights

• •Oesophageal cancer is heterogeneous at the molecular level.
• •Molecular heterogeneity can contribute to decreased therapy efficacy.
• •Computational modeling integrating genomic aberrations reveals phenotypic changes.
• •Phenotypic changes can represent signaling dysregulation and variable drug response.
• •Such model output can be used to improve patient-specific therapy efficacy.

Introduction

In 2020, esophageal cancer presented a global cancer burden of 604, 100 cases with a 20-fold rate of geographical divergence in incidence, with the higher percentages of esophageal squamous cell carcinoma occurring in Asia and Sub-Saharan Africa, and esophageal adenocarcinoma being found predominantly in Western countries.¹ Unless screening or proactive case finding is practiced, patients usually present with symptoms of esophageal obstruction, with approximately half already having distant metastases at diagnosis.² Although patients with early-stage disease have as much as 50%-75% chance of cure, patients with locally advanced disease have marginally curable cancers achieving just 25%-35% long-term survival with multimodality therapy (chemotherapy ± radiation therapy, and surgical resection ± adjuvant nivolumab), which has become the standard of care for esophageal cancer.³

For gastroesophageal adenocarcinoma, preoperative chemoradiotherapy or perioperative chemotherapy has been shown to significantly improve survival in patients with resectable, locoregionally advanced disease, as has the use of the anti-programmed cell death protein 1 (PD-1) antibody nivolumab in the adjuvant setting.4, 5, 6, 7, 8 In patients receiving preoperative chemoradiotherapy, guidelines support incorporating a doublet regimen of fluoropyrimidines, platinum analogs, or taxanes, with a preoperative radiation dose.^9,10 Currently, paclitaxel–carboplatin (PC), 5-fluorouracil–cisplatin, and FOLFOX are regarded as Category 1 recommendations by the National Comprehensive Cancer Network (NCCN) and by the European Society of Medical Oncology (ESMO),^9,10 while FLOT (fluorouracil, leucovorin, oxaliplatin, and docetaxel) is recommended in the perioperative setting.9, 10, 11 The verdict for superiority of one particular neoadjuvant chemotherapy or chemoradiotherapy strategy has not been resolved and it is unlikely that a ‘one-size-fits-all’ regimen exists.^12,13 For instance, results from CALGB 89803 demonstrated that an additional 18% and 20% of the fluorodeoxyglucose positron emission tomography nonresponders to induction chemotherapy with FOLFOX or PC could achieve pathologic complete remission via regimen switching to the alternative regimen during neoadjuvant chemoradiotherapy.¹⁴ While the results do not establish that PC or FOLFOX alone would suffice in those cases, 20% of patients appear to get improved outcomes as a result of receiving both regimens.

The use of biomarkers in gastroesophageal cancer to help drive therapy choice largely has been limited to human epidermal growth factor receptor 2 (HER2), microsatellite instability (MSI), tumor mutation burden, and programmed death-ligand 1 (PD-L1).15, 16, 17, 18, 19, 20 Emergent biomarkers such as CLD18.2, a splice-variant of CLD18, and FGFR2b, a splice-variant of fibroblast growth factor receptor (FGFR), may add to this list.²¹ In addition, molecular profiling has suggested other biomarkers of cytotoxic drug response governing DNA repair, the drug target, drug’s intracellular metabolism, and apoptosis.22, 23, 24, 25, 26, 27 Although individual biomarkers may be useful in predicting outcomes from treatment, their impact on outcome has not been sufficient to exclude or prefer any one particular form of therapy, and they have not been prospectively validated.²⁸ As a result, drug development efforts lacking effective biomarkers have resulted in the use of more agents, based on marginal benefits from randomized trials that naïvely assume that patients with adenocarcinoma of the oesophagus are genomically homogenous for the sake of study.

Computational biosimulation is an emerging approach that has shown promise in addressing the aforementioned conundrum. An in silico Computational Omics Biology Model (CBM), derived from the published literature and integrating a network of 6000+ human genes, 30 000+ molecular species, and 600 000+ molecular interactions, was developed to biosimulate the mechanistic, biochemical interactions (e.g. gene–protein, protein–protein, enzyme–substrate) that govern signaling pathways responsible for the hallmark behaviors of cancer. This model is personalized through the incorporation of genomic profiling data derived from a patient’s tumor, in essence creating an in silico model of the patient’s cancer cell. The interaction of various treatment strategies (including chemotherapy, radiation, targeted therapies, and immunotherapy) with the disease network is then computationally biosimulated to determine the treatment effect on a composite of hallmark pathways, represented by cell number or tumor volume. In this way the mechanistic model determines the relative efficacy of various treatment approaches, according to the vulnerabilities and resistance mechanisms in the genomic aberrations of each individual patient’s cancer.29, 30, 31, 32

In this study, the previously developed mechanistic model has been expanded through the development of a Therapy Response Index (TRI) in a cohort of patients with gastroesophageal adenocarcinoma receiving neoadjuvant chemotherapy. At the individual patient level, the CBM distills complex hallmark networks into summary measures, allowing the application of statistical and machine learning methods to elicit predictors of clinical outcome for a specific treatment based on the observed multivariate hallmark responses. To facilitate interpretability of clinical results, the resulting predictor is subsequently reduced to a single quantitative TRI score, scaled from 0 (unfavorable outcome) to 100 (favorable outcome) for a specific therapy. In the current study, we describe the association of TRI with disease-free survival (DFS), overall survival (OS), and tumor regression grade (TRG), and illustrate the potential clinical utility of TRI use.

Methods

Patient population

The performance of the CBM was studied in patients with gastroesophageal adenocarcinoma with operable cancers from the UK Oesophageal Cancer Clinical and Molecular Stratification (OCCAMS) International Cancer Genome Consortium study, who had either biopsy or surgical resection followed by 50× whole genome sequencing within 12 months of treatment. Individuals with no known values for nodal status or metastasis were excluded. An additional 226 individuals were subsequently excluded for data quality concerns, leaving 270 patients (Figure 1). More detailed descriptions of the cohort and OCCAMS study have previously been published.^33,34 Each patient was treated with neoadjuvant therapy according to UK clinical guidelines. Clinical and histopathology measurements were recorded for each patient, including but not limited to, OS times, DFS times, post-treatment resection pathology TRGs, tumor site, pretreatment histology, signet ring cells present, pretreatment tumor–node–metastasis (TNM), pretreatment tumor grading, and tumor response.

Figure 1.

Patient selection and filtering. A total of 496 unique patient records were initially evaluated, with 270 remaining after exclusions. Numbers of exclusion, by reason, are shown in the box to the left.

OS, overall survival.

Biosimulation

Biosimulation was carried out using a computationally derived mechanistic model which predicts changes in biomarker concentrations and hallmark cancer behaviors triggered by genomic aberrations or therapeutic intervention, see Pingle et al.³⁵ and Supplementary Methods, available at https://doi.org/10.1016/j.esmogo.2023.08.009 for more details.

TRI model development and validation

Stratified random sampling based on known clinical factors was used to split the data into training (N = 139 patients) and validation (N = 131 patients) datasets. For TRI development, multivariate Cox proportional hazards and proportional odds models were used to model OS, DFS, and TRG as a function of each individual’s CBM hallmark measures. The resulting linear predictors derived from multivariate hallmark responses were transformed into TRI values. Predicted clinical outcomes, including survival and progression probabilities, median survival, progression times, and probability of tumor regression along with corresponding 95% confidence intervals were derived.

Primary study objectives

TRI was validated in the remaining dataset of 131 patients from the OCCAMS study. Statistical analyses were carried out in accordance with the prespecified clinical study protocol. The primary objective of this study was to demonstrate that there is a statistically significant association between the therapy and patient-specific TRI value derived from the predefined TRI model and OS. OS was defined as the time from diagnosis to death. If no date of death was observed, the survival time was censored at the date of last follow-up. For this purpose, we compared the (full) Cox proportional hazards regression model for OS based on the (main) effects of patient age, patient sex, TNM tumor stage, TRG, neoadjuvant therapy, and TRI versus the (reduced) Cox proportional hazards regression model for OS excluding TRI. A P < 0.05 for the corresponding likelihood ratio statistic was considered significant. For the Kaplan–Meier analysis, a TRI score of 49 was used as the cut point for categorizing favorable versus unfavorable outcomes.

Secondary study objectives

As a secondary analysis, multivariate Cox proportional hazards and proportional odds regression models were used to examine the association between the predefined TRI and DFS (defined as the length of time during and after treatment that a patient lives without any signs or symptoms of their cancer) as well as TRG, respectively. For TRG, the Mandard classification system was used for assessment.³⁶ For analysis purposes, TNM stage was based on TNM version 7 pretreatment staging. Corresponding hazard and odds ratios and associated 95% confidence intervals are provided. In addition, predictive median survival times, median progression times, and predicted TRGs are provided for each patient and competing standard-of-care therapy where there are sufficient data to support the prediction.

Results

Patient population

Altogether, 496 patients with a diagnosis of gastroesophageal adenocarcinoma and 50× whole genome sequencing were screened for this study. Of these, 139 were excluded for not receiving preoperative treatment and 87 were excluded for missing clinical data (Figure 1). The sequencing carried out on the remaining assessable 270 patients was on tissue obtained from pretreatment biopsy (37%) or post-treatment resection (63%). The patient characteristics are described in Table 1. As standard for this population, the majority of patients were male; the median age was 65.7 years and most patients had symptomatic, locally advanced disease.

Table 1.

Baseline tumor and patient characteristics

Characteristics	Training	%	Validation	%	Total	%
Patient age at diagnosis
Missing	6	4	0	0	6	2
30-40	1	1	0	0	1	0
40-50	13	9	11	8	24	9
50-60	30	22	26	20	56	21
60-70	51	37	51	39	102	38
70-80	37	27	42	32	79	29
80-90	1	1	1	1	2	1
Total	139	100	131	100	270	100
Patient sex
Missing	7	5	0	0	7	3
Male	117	84	116	89	233	86
Female	15	11	15	11	30	11
Total	139	100	131	100	270	100
Tumor grade
Missing	9	6	13	10	22	8
Well	2	1	1	1	3	1
Well to moderate	2	1	0	0	2	1
Moderate	42	30	37	28	79	29
Moderate to poor	31	22	29	22	60	22
Poor	53	38	51	39	104	39
Total	139	100	131	100	270	100
Signet cell rings
Missing	11	8	11	8	22	8
Present	20	14	16	12	36	13
Absent	108	78	104	79	212	79
Total	139	100	131	100	270	100
Tumor regression grade
Missing	16	12	15	11	31	11
TRG1	8	6	13	10	21	8
TRG2	9	6	3	2	12	4
TRG3	17	12	15	11	32	12
TRG4	57	41	61	47	118	44
TRG5	32	23	24	18	56	21
Total	139	100	131	100	270	100
T-stage at resection (TNM7)
Missing	2	1	2	2	4	1
T0	8	6	12	9	20	7
T1	14	10	14	11	28	10
T2	17	12	14	11	31	11
T3	85	61	81	62	166	61
T4	13	10	8	6	21	8
Total	139	100	131	100	270	100
N-stage at resection (TNM7)
Missing	15	11	10	8	25	9
N0	36	26	44	34	80	30
N1	33	24	28	21	61	23
N2	25	18	28	21	53	20
N3	30	21	21	16	51	19
Total	139	100	131	100	270	100

TNM, tumor–node–metastasis; TRG, tumor regression grade.

The standard treatments received by patients represented UK clinical decision making in a diverse group of individuals with varying degrees of eligibility for cytotoxic strategies, combination treatments, and radiation (Table 2). A majority received epirubicin-based therapy (epirubicin/ciplatin/fluorouracil or epirubcin/cisplatin/capecitabine) rather than FLOT (i.e. fluorouracil–leucovorin, oxaliplatin, and docetaxel), reflecting the period during which the patients were accrued for treatment. Only 10% of patients received radiation in combination with chemotherapy, and the remainder received chemotherapy alone. Only three patients received HER2-directed treatment in the preoperative setting (two off-protocol and one as part of a clinical trial). None of the patients received PD-L1 immunotherapy, as it had not received regulatory approval in the United Kingdom for esophagogastric adenocarcinoma (OGA) at the time of the study.

Table 2.

Preoperative therapy administered

Therapy	Sample size
Chemotherapy (plus other agents)
Fluoropyrimidine agent–platinum agent–epirubicin	185
Fluoropyrimidine agent–platinum agent	42
Fluoropyrimidine agent–platinum agent–docetaxel (FLOT)	8
Fluoropyrimidine agent–platinum agent–lapatinib	3
Fluoropyrimidine agent–platinum agent–trastuzumab	3
Fluoropyrimidine agent–platinum agent–dexamethasone–epirubicin	2
Platinum agent–epirubicin	2
Fluoropyrimidine agent–platinum agent–epirubicin–lapatinib	1
Fluoropyrimidine agent–epirubicin	1
Taxane Agent–platinum agent	1
Chemotherapy and radiation
Fluoropyrimidine agent–platinum agent–radiation	9
Fluoropyrimidine agent–platinum agent–epirubicin–radiation	8
Taxane agent–platinum agent–radiation	5

A variety of DNA repair pathway aberrations were encountered through interrogation of the whole genome sequencing data, including abnormalities in homologous recombination repair, DNA checkpoints (ATM, ATR, CHEK1, CHEK2), and mismatch repair, among others (Figures 2 and 3). Output from the mechanistic model illustrated additional dysregulated pathways, including TP53 mutations, amplification of multidrug resistance pumps, aberrations of the WNT pathway, Hippo pathway, phosphatase and tensin homolog–phosphatidylinositol-3 kinase–protein kinase B (PTEN–PI3K–AKT) pathway, chromatin regulation genes, and alterations of autophagy and oxidative stress resistance mechanisms. These results are consistent with the previous molecular profiling of this cohort that has been reported elsewhere.^33,34 At the hallmark pathway level, the primary drivers of OS, DFS, and TRG were cell proliferation, survival, apoptosis, DNA damage, methylation, oxidative stress, immune evasion, and metastasis.

Figure 3.

Circosplot showing copy number variation(CNV) amplifications (brown), CNV deletions (green), and mutations (blue). Cytogenetic bands are shown in black in the outer ring. Line height represents the number of patients and is scaled the same across all aberrations.

Figure 2.

Genomic mutations among the 270 patients with esophageal adenocarcinoma, limited to the top 25 mutated genes. Number of patients are shown on the ordinate, genes are shown on the abscissa.

GOF, gain of function; LOF, loss of function; SOF, switch of function.

TRI association with OS

The primary objective of this study was to demonstrate a statistically significant association between the predefined TRI and OS in the validation set. We compared the (full) Cox proportional hazards regression model for OS based on the effects of patient age, patient sex, TNM tumor stage, TRG, neoadjuvant therapy, and the predefined TRI versus the (reduced) Cox proportional hazards regression model for OS excluding TRI (Table 3). TRG (P < 0.0001) and pretreatment N-stage (P = 0.0005) were significantly predictive of OS. However, the choice of physician-directed neoadjuvant therapy was only marginally significant (P = 0.0559) after adjusting for these factors. By contrast, the association of TRI with OS remained significant after adjustment for these clinical factors (likelihood ratio P value = 0.0012, HR = 0.548 per 25 units of TRI), meeting the primary objective of the study. We also carried out a similar post hoc analysis adjusting for post-treatment N-stage and saw comparable results (TRI likelihood ratio P = 0.0028, Table 4). Figure 4A displays the corresponding Kaplan–Meier plot for OS stratified by TRI favorable outcome and unfavorable outcome risk groups (median OS = 82.1 months, 40.7 months, respectively; TRI threshold = 49) in the validation set. The resulting log-rank P value = 0.0051, again indicating that TRI is significantly associated with OS.

Table 3.

Type 3 likelihood ratio statistics for multivariate model for OS and DFS

OS	df	Likelihood ratio χ2	P value	Hazard ratio per 25 units TRI
Patient age at diagnosis	1	1.5254	0.2168
Patient sex	1	0.5595	0.4545
Physician-prescribed treatment	5	10.7825	0.0559
Tumor regression grade	4	31.7410	<0.0001
Pretreatment T-stage	1	0.3940	0.5302
Pretreatment N-stage	1	12.2425	0.0005
TRI	1	10.4241	0.0012	0.548 (95% CI 0.378-0.798)

DFS	df	Likelihood ratio χ2	P value	Hazard ratio per 25 units TRI
Patient age at diagnosis	1	0.8022	0.3704
Patient sex	1	1.0953	0.2953
Physician-prescribed treatment	5	8.6904	0.1221
Tumor regression grade	4	19.1321	0.0007
Pretreatment T-stage	1	0.6211	0.4306
Pretreatment N-stage	1	18.0257	<0.0001
TRI	1	4.7803	0.0288	0.653 (95% CI 0.441-0.965)

TRG	df	Likelihood ratio χ2	P value	Odds ratio per 25 units TRI
Patient age at diagnosis	1	0.0670	0.7958
Patient sex	1	0.4194	0.5172
Pretreatment T-stage	1	0.0910	0.7629
Pretreatment N-stage	1	0.3612	0.5479
TRI for TRG	1	10.6885	0.0011	0.124 (95% CI 0.033-0.469)

CI, confidence interval; DFS, disease-free survival; OS, overall survival; TRG, tumor regression grade; TRI, Therapy Response Index.

Table 4.

Type 3 likelihood ratio statistics for the multivariate model for OS and DFS using postsurgical N-stage

Effect	df	Likelihood ratio χ2	P value	Hazard ratio per 25 units TRI
OS
Patient age at diagnosis	1	0.5141	0.4734
Patient sex	1	0.2246	0.6355
Physician-prescribed treatment	5	7.4018	0.1924
Tumor regression grade	4	12.8511	0.0120
Pretreatment T-stagea	3	0.7674	0.3810
Post-treatment N-stageb	3	8.4906	0.0036
TRI	1	8.9035	0.0028	0.592 (0.413-0.245)
DFS
Patient age at diagnosis	1	0.2113	0.6457
Patient sex	1	0.9556	0.3283
Physician-prescribed treatment	5	10.0396	0.0741
Tumor regression grade	4	2.4797	0.6483
Pretreatment T-stagea	1	1.1260	0.2886
Pretreatment N-stageb	1	12.4080	0.0004
TRI	1	3.2587	0.0710	0.722 (0.503-0.363)

DFS, disease-free survival; OS, overall survival; TRI, Therapy Response Index.

^aAT-Stage T1/T2 versus T3/T4/T5.

^bBN-Stage N0 vs N1/N2/N3.

Figure 4.

(A) Kaplan–Meier analysis of overall survival (OS) in the combined training and validation sets; favorable outcomes are in blue, unfavorable outcomes in green. Patients were dichotomized into risk groups using the median Therapy Response Index (TRI) of 43.6. The number of patients at risk by 6-month increment is shown on the lower table. (B) Kaplan–Meier analysis of progression-free survival in the combined training and validation sets; favorable outcomes are in blue, unfavorable outcomes in green. Patients were dichotomized into risk groups using a TRI of 49. The number of patients at risk by 6-month increment is shown on the lower table.

DFS, disease-free survival.

TRI association with DFS and TRG

In a secondary analysis, a statistically significant association between the predefined TRI and DFS was demonstrated in the validation set. We compared the (full) Cox proportional hazards regression model for DFS based on the (main) effects of patient age, patient sex, TNM tumor stage, TRG, neoadjuvant therapy, and the predefined TRI versus the (reduced) Cox proportional hazards regression model for DFS excluding TRI (Table 3). Again, TRG (P = 0.0007) and pretreatment N-stage (P < 0.0001) were significantly associated with OS. Physician-directed neoadjuvant therapy was not significant (P = 0.1221) after adjusting for TRI and other clinical factors. By contrast, TRI had a likelihood ratio P value = 0.0288 (adjusted HR = 0.653 per 25 units of TRI), indicating that the predefined TRI is significantly associated with DFS even after adjustment for neoadjuvant therapy and clinical factors; the corresponding Kaplan–Meier plot comparing favorable versus unfavorable outcome risk groups is shown in Figure 4B (median PFS 104 months versus 31 months, respectively, log-rank P = 0.0296, TRI threshold = 49).

In additional secondary analysis, we aimed to demonstrate a statistically significant association between the predefined TRI and TRG. For this purpose, TRG1|TRG2 was compared with TRG3|TRG4|TRG5. We compared the (full) logistic regression model for TRG based on the (main) effects of patient age, patient sex, TNM tumor stage, TRG, and the predefined TRI versus the (reduced) logistics regression model for DFS excluding TRI (Table 3). Because of the limited number of patients with complete data, physician-prescribed treatment was not included. TRI had a likelihood ratio P value = 0.0011 (adjusted OR = 0.124 per 25 units of TRI), indicating that the predefined TRI is significantly associated with TRG even after adjustment for clinical factors; the corresponding receiver operating characteristic plot is provided in Figure 5 and yields an area under the curve of 0.747 and a sensitivity and specificity of 70% and 80%, respectively (Table 5)

Figure 5.

Receiver operating characteristic curve for Therapy Response Index(TRI)prediction of tumor regression grade. The area under the curve (AUC) was 0.747. The blue line represents the tangent line to the point that maximizes Youden’s J index, which provides the maximum sensitivity (70%) and specificity (81%) for the TRI.

Table 5.

Performance characteristics of the TRG TRI model

Endpoint	Sensitivity, %	Specificity, %	Area under the curve	TRI cut point
TRG	70.0	81.3	0.747	71.9

TRG, tumor regression grade; TRI, Therapy Response Index.

Examples of clinical utility using TRI

To illustrate the potential clinical utility of TRI, a predictiveness curve for OS as a function of the TRI score for physician-directed neoadjuvant therapy in the validation set is shown in Figure 6A, with a range of TRI scores being observed (TRI 25th percentile = 33.4, median = 42.8, 75th percentile = 52.9). In some instances, neoadjuvant therapy was associated with a low to medium TRI score and likewise a poorer predicted OS (left of the graph), leading to the possibility that alternative therapies might have a higher TRI and provide an improved outcome (Figure 6B and C). An example of why a patient with a low TRI does not respond to neoadjuvant therapy is shown at the molecular, mechanistic level in Figure 7A. In this instance, the mechanistic model identified patient-specific aberrations that led to the lack of efficacy, including alterations in CES1, encoding carboxylesterase 1 which is necessary for the conversion of capecitabine to 5-fluorouracil, and SLC31A1 (also known at CTR1), which encodes a transporter essential for the entry of platinum agents into the cell.37, 38, 39, 40, 41, 42 In addition, amplifications in multiple genes, including numerous transporters such as ABCC1, ABCC2, ABCC4, ABCC5, and ABCB1, which lead to capecitabine, cisplatin, and epirubicin resistance (Supplementary References, available at https://doi.org/ 10.1016/j.esmogo.2023.08.009) were observed. Alternatively, there are instances where the physician-directed neoadjuvant therapy was associated with a higher TRI (right of the graph) and the clinician can be assured that their therapeutic choice will result in an optimal outcome versus other alternative therapies (Figure 6D). An example of why a patient with a high TRI responds to physician-directed neoadjuvant therapy is shown at the molecular, mechanistic level in Figure 7B. Here, numerous aberrations in the patient’s genome supported the use of ECX, including an amplification in TOP2A and deletion in PRKDC which confers sensitivity to epirubicin,^43,44 amplifications in RRM1 and RRM2 conferring sensitivity toward capecitabine,⁴⁵ and numerous additional aberrations conferring sensitivity to both capecitabine and cisplatin (Figure 7B, Supplemental References, available at https://doi.org/10.1016/j.esmogo.2023.08.009).

Figure 6.

(A) Predictiveness curve for overall survival (OS) as a function of the Therapy Response Index (TRI) score. Ordinate is predicted OS at 36 months, abscissa is TRI. Vertical marks above the abscissa represent distribution of the patients in the validation set. (B) TRI probability curve for an hypothetical individual, showing poor efficacy for physician-directed neoadjuvant therapy and better predicted efficacy for therapies A and B. (C) TRI probability curve for an hypothetical individual, showing intermediate efficacy for physician-directed neoadjuvant therapy and poorer predicted efficacy for therapy A and better predicted efficacy for therapy B. (D) TRI probability curve for an hypothetical individual, showing good efficacy for physician-directed neoadjuvant therapy and poorer predicted efficacy for therapies A and B.

PPT, physician-prescribed treatment.

Figure 7.

Molecular mechanistic examples of patients receiving neoadjuvant therapy (epirubicin, cisplatin, capecitabine—ECX) with (A) low Therapy Response Index (TRI) and (B) high TRI. (A) The patient was treated with ECX [overall survival (OS) = 12.15 months and disease-free survival (DFS) = 9.4 months]. The patient progressed as inferred from the DFS status. The genomic profile of the patient presents aberrations in the CES1 and SLC31A1 genes (brown boxes), which lead to resistance in capecitabine and cisplatin, respectively. The patient has amplifications in multiple genes, including numerous transporters such as ABCC1, ABCC2, ABCC4, ABCC5, and ABCB1, which lead to capecitabine, cisplatin, and epirubicin resistance. (B) The patient was treated with ECX (OS = 38.17 months and DFS = 36.96 months). The patient was disease free during the course of the treatment as inferred from the DFS status. The genomic profile of the patient presents amplifications in the TOP2A gene (blue box) and deletion in PRKDC (brown box), which confer sensitivity to epirubicin. The patient has amplifications in RRM1 and RRM2 genes (gray box) which confer sensitivity toward capecitabine. In addition, aberrations in CUX1, ATR, KMT2D, MLH1, UNG, and SMUG1 confer sensitivity toward capecitabine (brown boxes). Aberrations in ERCC1/2, USP1, REV3L, ATR, CDK12, XRCC1, POLH, HMGA1, and ARID1A confer sensitivity toward cisplatin (brown boxes). For both (A) and (B): wild-type phenotype = light gray; knockdown = brown; overexpression = grayish blue. Inhibitory interactions are shown in red, activating interactions are shown in black.

Discussion

Despite advances in gastroesophageal cancer therapy, only a minority of patients achieve long-term survival from early-stage cancers that are potentially curable at diagnosis. Although clinical practice has evolved to include therapies developed via precision-based approaches, including the identification of genomic biomarkers shown to predict patient outcomes across multiple treatments,⁴⁶ molecular diagnosis has been limited in the ability to predict therapy response and by single biomarker–drug associations that over-simplify disease and treatment. Personalized ‘N-of-1’ approaches that incorporate a patient’s complete tumor genomic landscape are lacking, and the size and complexity of comprehensive genomic panels present a daunting and complex biology that is difficult for oncologists to translate into clinical decision-making.

Computational biosimulation provides a potential solution to these challenges. The mechanistic model described in this work, based on a patient’s genomic aberration-driven alterations in hallmark signaling pathways, provides insight into how unique features of the individual’s tumor biology impact specific therapeutic strategies. The development of the TRI score, which weighs the alterations of an individual patient’s hallmark signaling pathways with regard to disease prognosis and treatment efficacy, may permit the prediction of which treatment approach is expected to produce the best DFS and OS. This study finds that physician-selected therapies may be optimal in some cases, while for others, an alternative treatment based on comprehensive genomic analysis and biosimulation might have produced a better outcome (Figure 6).

In the current study, use of TRI showed that the benefits of specific treatment strategies, including fluoropyrimidines, taxanes, anthracyclines, and platinum compounds, varied among patients (Figures 4 and 6). Patients with unfavorable TRI scores had worse outcomes (Figure 4) and could be candidates for alternate therapies. Although the therapies in this study were not sufficiently powered to make reliable alternate therapy predictions, it is easy to picture the clinical utility of such predictions (Figure 6B–D), which can be supported at the molecular level (Figure 7A and B; Supplementary References, available at https://doi.org/10.1016/j.esmogo.2023.08.009). In the patient described in Figure 7A, the low TRI score suggests that an alternate therapy might have proven more beneficial than ECX. Interrogation of the patient’s genomic profile showed amplification of ERBB2, suggesting that trastuzumab might have been a better alternate therapy (data not shown). The ability of the mechanistic model to identify the molecular underpinnings for a given therapeutic response translates to the identification of the correct therapy for a given patient. The generation of TRI scores across the approved formulary for an indication such as gastroesophageal cancer will allow clinicians to identify which therapy is most likely to benefit their patient and, importantly, which will not.

This study has a number of limitations. First, the use of certain therapies, including radiation therapy, HER2-directed therapy, and immunotherapy among others, was limited in the OCCAMS cohort. As a result, while biosimulation can identify response and resistance mechanisms for these interventions, TRI predictions based on statistical modeling are not yet validated in patients with gastroesophageal adenocarcinoma who received those therapies. Second, the use of adjuvant therapy and postprogression treatment may have affected OS. Third, as this was a prospective–retrospective study, patients were not randomized to different treatments, nor were they uniform with regard to treatment eligibility. Hence, it is not possible to discern the extent to which the observed survival outcomes may have been confounded by unmeasured factors (e.g. fitness, nutrition) or were determined by prognosis rather than prediction of treatment responsiveness. While computational biosimulation is mechanistically driven by drug sensitivity and resistance mechanisms to specific treatment strategies, only a genomically randomized dataset can definitively segregate predictive from prognostic factors. Conceivably, the adjustment of the mechanistic model scores to fit the survival outcomes in the training set diminishes the prognostic impact on outcome. It is noteworthy that clinicians would use biosimulation after diagnosis, diminishing the importance of this distinction. Nevertheless, we acknowledge the value of a randomized patient data to confirm the predictive utility of biosimulation in situations where a decision for treatment has been made. Fourth, genomic analysis was carried out on samples collected before initiation of therapy via biopsy as well as during resection, after therapy completion. It is possible that therapeutic bias might exist in the post-therapy samples. In addition, sequencing carried out on a single biopsy sample may not adequately account for intratumoral molecular heterogeneity. For example, the disease that has already metastasized to lymph nodes may possess drivers of invasion, metastases, and drug resistance that are absent in the primary site. Notably, the failure to achieve a remission at sites of lymph node metastases negates the achievement of complete tumor regression in the primary site.⁴⁷ Tissue sampling from multiple sites would present a more accurate picture of intratumoral heterogeneity and thereby enhance the initial molecular diagnosis and the therapeutic decisions that are based on it, although it is not currently practical to implement.

Finally, with regard to implementing biosimulation in the clinic, the tissues in this study underwent whole genome sequencing that was able to embrace the complexity of the underlying disease and therapy response, including detailed chromosomal structural abnormalities. However, the read depth of the sequencing at ∼50× was lower than that achieved via exome sequencing, which could miss low-frequency mutations. Similarly, it is unlikely that molecular diagnosis based on small gene panels will provide a sufficiently comprehensive picture of signaling pathways to address the spectrum of cytotoxic therapy and radiation, while the profile it does provide tends to capture low-frequency mutations due to their higher read depth.

Despite the aforesaid limitations, this study provides an early glimpse into the benefits provided by a personalized ‘N-of-1’ approach to therapy selection versus population- or precision-based approaches and suggests that this approach may be beneficial in other indications. By utilizing the complete genomic picture provided by a patient’s tumor, biosimulation can help provide an upfront understanding of which therapy will be effective and, equally importantly, which therapy will not be. Further validation of this approach in additional prospective–retrospective cohorts as well as larger, prospective studies will help demonstrate and cement the clinical utility of this approach.