Using clinically rich transcriptomic data, we have identified key regulatory networks and novel therapeutic opportunities in OAC
Oesophageal adenocarcinoma (OAC) is a devastating disease, with rising incidence, poor survival rates, and no effective targeted treatments despite decades of research. To develop precision therapies for OAC, we need data that captures the full complexity of the disease. Yet too often, the datasets meant to drive insights fall critically short, lacking proper clinical annotation, missing key disease stages, or failing to include the most relevant controls and data types.
In collaboration with the Innovation for Translation Research Group in the School of Cancer Sciences, University of Southampton, we have generated one of the largest known clinically annotated OAC patient datasets, which will allow us to redefine how this deadly cancer is treated in the clinic.
OAC is on the rise, particularly in North America and Europe, where thousands of new cases are diagnosed each year. The increasing prevalence of obesity and chronic acid reflux, both key drivers of OAC and its intermediate metaplastic condition — Barrett’s Oesophagus — is fueling this trend. Yet despite growing awareness, most cases are still caught too late [1,2,3]. Only about a quarter of patients are diagnosed at an early stage, when treatment has the best chance of success. For the rest, the outlook is far more severe—by the time OAC reaches stage IV, fewer than 5% of patients survive beyond five years, making its prognosis among the poorest of all cancers [4,5]. Barrett’s Oesophagus increases the risk of developing OAC by 10-50x, so understanding this intermediate disease is crucial to understanding OAC itself [5,6].
Adding to this issue is the lack of precision therapies for OAC. Standard of care (SOC) treatment includes chemo- or chemoradiotherapy before surgery (NACT), but the response rate of late-stage patients' to NACT is less than 20%, with generally poor survival outcomes [1,3,5]. Recent clinical trials of precision therapies—such as agents targeting HER2 and EGFR, as well as immune checkpoint inhibitors like pembrolizumab—have yielded mixed or negative results [3,5].
The development of precision therapies in OAC has been severely limited, often stalling before clinical translation due to inadequate selection of robust targets and predictive biomarkers. This is driven by the application of the wrong datasets for target discovery. Major initiatives like The Cancer Genome Atlas (TCGA) and Genotype-Tissue Expression (GTEx) have provided a wealth of publicly available data across cancer subtypes, but with a number of significant shortcomings.
GTEx profiles tissues from non-diseased deceased individuals, but these are not matched with paired tumour samples from the same patient. Comparing GTEx to tumour samples therefore requires the application of complex batch correction algorithms, introducing bias and obscuring real biological signals.
Although TCGA provides matched tumour and normal samples, the so-called 'normal' tissues are often taken from regions immediately adjacent to the tumour, based on the assumption that histologically normal appearance equates to normal biological behaviour. However, the regions surrounding a tumour don’t exist in isolation, but are shaped by the tumour’s presence through genetic, epigenetic, and metabolic alterations, and may not provide the necessary insights to effectively dissect tumour biology.
In addition, both datasets lack full clinical annotation and are not representative of the real clinical patient population. There is therefore an urgent need for new, clinically annotated datasets that tackle the major issues present in the existing available data.
With our partners in the Underwood and Walters labs at the University of Southampton, we have built a large proprietary RNA-seq dataset spanning three distinct tissue types— normal oesophagus, Barrett’s, and OAC — offering a comprehensive view of disease progression.
Layered with extensive clinical metadata, including survival, treatment response, node stage, and other key pathological/molecular features, this dataset provides a powerful foundation for uncovering biomarkers and identifying new therapeutic opportunities.
Coupling the dataset with Evariste’s AI-driven Frobenius Target platform will redefine how novel targets are selected, and will drive breakthroughs in the treatment options available for OAC patients.
To demonstrate the unique features of our dataset, we have performed a set of example analyses that illustrate the importance of using the right data to derive meaningful insights.
Using deconvolution analysis to estimate the proportion of cellular subtypes from bulk transcriptomic sequencing, along with Southampton’s largest publicly available single-cell OAC dataset, we can observe that normal samples derived from TCGA appear to be “contaminated” with a significant proportion of tumour cells (Fig. 1, TCGA ANT vs TCGA Tumour).
Our dataset bridges this gap by including normal tissue sampled from the lower oesophageal lining, with patient-matched Barrett’s and OAC samples. Deconvolution analysis reveals that these normal samples clearly differentiate from both Barrett’s (as a precancerous lesion), and OAC, offering a unique window into understanding cancer progression (Fig.1, EV Normal vs Barrett’s vs Tumour).
.png)
A key insight from this early analysis is that many of the key oncogenic changes found in tumour cells are already present in Barrett’s tissue (Fig. 2). Additionally, sampled normal tissue adjacent to the tumour is likely to fail to capture many of the most critical changes in gene expression that underlie the transformation into cancer.
OAC is characterised by extensive genomic instability, with a markedly high mutational burden of eight mutations per megabase. However, its genetic alterations are highly heterogeneous, rarely recurrent, and can vary significantly within different regions of the same tumour. This extreme heterogeneity makes it difficult to define reliable genetic signatures or identify consistently actionable mutations.
While DNA sequencing has been the cornerstone of molecular research, its limitations in capturing the variability within OAC highlight the need for alternative strategies. Transcriptomic analysis provides a powerful solution, offering a high-throughput approach to identifying key networks and potential therapeutic targets.
After adjusting for unwanted sources of variation, our AI tools build intricate gene networks to reveal hidden layers of co-regulation and functional interplay. For example, for a set of hard-to-drug synthetic lethal pairs, we identified key druggable vulnerabilities within the same functional network, where knockout triggers profound effects in OAC models (Fig. 3).
To refine our understanding of disease progression and optimise therapeutic strategies, genomic alterations must be examined in the context of clinical outcomes. As an example, we highlight a novel target that acts as a key regulator of G2 arrest and cell fate under genotoxic stress. This target shows a stepwise increase in expression from normal tissue to Barrett’s oesophagus and on to tumour, suggesting a role in malignant transformation (Fig. 4A). Notably, elevated expression correlates with poor survival, positioning it as a potential prognostic biomarker (Fig. 4B). Furthermore, its association with tumour regression grade (TRG) suggests that higher expression correlates with a reduced response to neoadjuvant chemo or chemo-radiotherapy (Fig. 4C), making it a potentially promising therapeutic option for patients who would respond poorly to SOC treatments.
Taken together, these results show the critical importance of generating datasets with the correct normal controls, full clinical annotation, and the right data type. They also demonstrate that analysing these complex multimodal datasets is challenging, and that AI techniques are well placed to identify networks within the data that can inform clinical practice and facilitate novel target identification. As our understanding of OAC biology progresses, therapeutic strategies will be refined, integrating targeted drug discovery and biomarker-driven approaches to enhance treatment precision and efficacy.
Stay updated with Evariste's blogposts
Stay updated with Evariste's blogposts
Stay updated with Evariste's blogposts
Stay updated with Evariste's blogposts
Stay updated with Evariste's blogposts
Stay updated with Evariste's blogposts
.png)
.png)
