Programme

DNA methylation is an epigenetic mark and large-scale alterations in DNA methylation patterns are observed in development, ageing and diseases such as cancer. Previous mathematical models have aided in understanding methylation patterns, but the computational expense associated with these studies have limited their use to the study of small-scale patterns. During this talk, I will discuss using cluster mean-field models to quickly, and accurately, predict statistical properties associated with large-scale DNA methylation patterns. Much of my talk is based on the publication https://royalsocietypublishing.org/doi/10.1098/rsif.2021.0707

My past academic research focused on using mathematical and computational models to investigate the growth and collective migration of cancer cell populations. The main focus of that research was to look at heterogeneous populations – where subtypes of cancer cells had different abilities – and how that can drive metastasis and spread of the cancer. This work was mathematically rigorous and explored methods for modelling these phenomena, however, was not a particularly data-driven approach.

My current position, as a quantitative systems pharmacology (QSP) scientist at Certara, has allowed me to continue using similar mathematical methods in real world applications. At Certara, we use QSP modeling to predict how new treatments may work in humans. Within these models, we use preclinical and clinical datasets to calibrate and validate the modelling techniques. These models can then be used to support the design and approval of further clinical trials.

My talk will briefly cover some of the differences and similarities between these theoretical and data-driven research approaches. Both approaches are essential—working together, they advance cancer research and more generally, biomedical research.

My career has been shaped by a fascination with understanding complex biological systems through mathematical and computational approaches. A central focus of my research is understanding how cancer cells interact with their surroundings. The tumour microenvironment is a highly dynamic and spatially organised ecosystem in which interactions between cancer cells and the extracellular matrix (ECM) can drive tumour invasion and metastatic progression. Capturing these processes requires approaches that bridge biological complexity with quantitative and computational thinking.

In this talk, I will present my work on developing 3D individual-based models of the tumour microenvironment that explicitly represent cancer cells, ECM structures, and other tissue components. These models allow us to investigate how cell movement, ECM remodelling, and local interactions collectively shape tumour behaviour. I will also discuss the translation of this work into PhysiMeSS, an extension of the widely used PhysiCell platform developed through collaboration with researchers across the international community. This has enabled a framework originally developed within my own research programme to become a shared community resource, demonstrating the power of collaborative and community-driven computational science.

Alongside the science, I will reflect on my journey into computational biology and the many unexpected twists and turns that have shaped my career. Navigating an interdisciplinary research path while balancing the complexities of life beyond academia has influenced not only the questions I ask, but also how I approach collaboration, leadership, and mentoring. I will share some of the lessons learned along the way, highlighting the importance of supportive networks, strong mentors, and diverse perspectives in tackling some of the most challenging problems in modern biomedical research.

In recent years many institutions have adopted research culture strategies and action plans, influenced at least in part by the inclusion of new criteria in the Research Evaluation Framework. Equality, Diversity and Inclusion principles are often core to these plans. But although the “research culture” terminology is relatively recent, the EDI challenges in research environments and efforts to address them are not new. In this talk I will give some personal perspectives on this landscape and in particular, how the picture can become even more complicated when multiple disciplines, with different default cultures, collide.

Background

Tumour-infiltrating lymphocytes (TILs) are clinically informative in breast cancer, but conventional assessment is limited by observer variability and spatial heterogeneity. ECTIL (label-efficient computational tumour-infiltrating lymphocyte assessment) is a recently validated computational pathology model that estimates stromal TIL abundance from routine H&E whole-slide images (1). Using three complementary breast cancer cohorts, CPTAC-BRCA, TCGA-BRCA and TransNEO, we examined the biological, prognostic and treatment-response relevance of ECTIL. We investigated what biological information ECTIL captures, whether its correlation with immune-related signalling is consistent across cohorts, and whether ECTIL provides prognostic information beyond established transcriptomic immune signatures.

Methods

CPTAC-BRCA (2) was used as the primary molecular characterization cohort. In 110 patients with matched RNA and 273 eligible H&E slides, the ECTIL-compatible workflow included foreground segmentation, normalization to 0.5 µm/pixel, extraction of non-overlapping 512×512-pixel tiles, RetCCL feature extraction and frozen ECTIL-TCGA inference. For patients with multiple eligible slides, patient-level ECTIL was calculated as the tile-count-weighted mean.

Before transcriptome-wide analysis, ECTIL was tested against six prespecified immune RNA measures representing distinct biological axes: cytolytic activity (CYT), IFNG-6, CD8/T-cell core, T-cell-inflamed positive-17, B-cell/humoral-5 and ESTIMATE ImmuneScore. Spearman correlations were estimated with 10,000-bootstrap confidence intervals and Benjamini-Hochberg (BH) correction across the six measures. Transcriptome-wide analysis was then used to characterize gene-level relationships with ECTIL, with covariate-adjusted HC3 regression as a complementary analysis; gene-level significance was defined at BH FDR<0.05. Nested Elastic Net stability selection was subsequently used for compact RNA-score development; no clinical outcome was used during this process.

TCGA-BRCA (3, 4) provided independent RNA prognostic validation and a separate training aware image analysis. Frozen RNA scores were calculated in 1,058 primary-tumour cases. ECTIL was successfully generated for 1,122 processable H&E slides from 1,051 patients; seven slides lacking essential microns-per-pixel metadata were excluded for technical reasons. After the primary TCGA package was frozen, released ECTIL train, validation and test files were used to reconstruct the original labelled TCGA cohort of 356 slides from 342 patients and identify 710 image-RNA patients outside that cohort. Post-freeze sensitivity analyses repeated ECTIL-RNA correlations in these patients and examined receptor-phenotype consistency. Progression-free interval (PFI) was the primary clinical endpoint; DSS and OS were secondary and DFI exploratory. Cox models used a frozen clinical adjustment backbone comprising age, ordinal pathologic stage, and receptor-defined clinical phenotype. To assess complementary prognostic information, PFI models compared the frozen clinical backbone alone, clinical+ECTIL, clinical+RNA and joint clinical+ECTIL+RNA models. Incremental prognostic value of ECTIL beyond each of the six established RNA signatures, and RNA beyond ECTIL, was assessed using nested likelihood ratio tests with BH correction across the six-signature family (BH-adjusted p values reported as q values). ECTIL×RNA interactions were evaluated separately.

TransNEO (5) provided independent neoadjuvant-response validation. ECTIL was generated for 203 H&E slides from 203 patients; 147 patients had matched ECTIL, RNA and response data. Associations with pCR and continuous residual cancer burden (RCB) were tested and joint ECTIL-RNA models assessed complementary response information.

Results

Across cohorts, patient-level ECTIL ranged from 0.070–0.557 in CPTAC-BRCA, 0.034–0.617 in TCGA-BRCA and 0.051–0.604 TransNEO image cohort. In CPTAC-BRCA, ECTIL was positively associated with all six prespecified immune RNA measures. Five remained significant after multiplicity correction: IFNG-6 (ρ=0.343, q=0.002), T-cell-inflamed positive-17 (ρ=0.277, q=0.010), CYT (ρ=0.261, q=0.012), CD8/T-cell core (ρ=0.251, q=0.012) and ESTIMATE ImmuneScore (ρ=0.223, q=0.023); B-cell/humoral-5 was positive but weaker (ρ=0.149, q=0.121). Transcriptome-wide analysis identified 7,061 genes at FDR<0.05, of which 1,622 were significant in both the primary and age-, stage- and TNBC-adjusted analyses. Stability development yielded a compact nine-gene ECTIL RNA score (BATF2, IFNG, LAG3, TNFRSF13C, KIR2DL4, PLA2G2D, HMGB2, HSPD1 and CD320), with repeated held-out internal CPTAC correlation ρ=0.618; in independent TCGA RNA validation, the score was associated with longer adjusted PFI (HR per SD=0.67, 95% CI 0.53–0.84; p<0.001).

In the 1,051 patient TCGA image-RNA cohort, ECTIL was positively correlated with all six established immune RNA measures and the nine-gene score (all BH q<0.05). In the 710 patients outside the original labelled ECTIL TCGA cohort, ECTIL remained correlated with all six established RNA measures (ρ=0.235–0.317; all BH q<0.05) and the nine-gene score (ρ=0.427, q=4.83×10-32). The ECTIL–nine-gene score relationship remained positive across major receptor phenotypes, including TNBC, with no evidence of heterogeneity in the full cohort. In fully adjusted joint PFI models using a common 895-patient, 96-event analysis set, higher ECTIL remained associated with lower progression risk when modelled alongside each established RNA signature, with ECTIL HRs per SD ranging from 0.61 to 0.65. ECTIL added significant prognostic information beyond all six established RNA signatures after multiplicity correction (BH q=0.0017–0.0029). In contrast, none of the six established RNA signatures added significant prognostic information beyond ECTIL after correction, and no ECTIL×RNA interaction was significant across the six-signature family. This incremental-value pattern persisted in the post-freeze TCGA sensitivity restricted to patients outside the original labelled ECTIL cohort (n=599, 60 PFI events): ECTIL added significant prognostic information beyond each established RNA signature after BH correction, whereas reciprocal RNA-beyond-ECTIL and interaction tests were not significant after correction.

In the TransNEO response validation (n=147; 38 pCR), higher ECTIL was associated with increased odds of pCR (adjusted OR per SD=1.73, 95% CI 1.18–2.54; p=0.005) and lower continuous RCB (adjusted β per SD=−0.33, 95% CI −0.52 to −0.14; p<0.001). None of the six established RNA signatures added significant pCR or RCB information beyond ECTIL after multiplicity correction; ECTIL retained incremental information beyond B-cell/humoral-5 (pCR q=0.036; RCB q=0.034).

Conclusions

ECTIL showed reproducible correlation with multiple established immune transcriptomic signatures, retained prognostic information beyond individual RNA immune measures, and independently associated with pathological treatment response. These findings support ECTIL as a biologically interpretable computational pathology biomarker complementary to bulk immune transcriptomics. The nine-gene score may provide a compact secondary RNA representation of the ECTIL-associated transcriptomic signal.

References

1. Schirris Y, Voorthuis R, Opdam M, Liefaard M, Sonke GS, Dackus G, et al. Label-efficient computational tumour infiltrating lymphocyte assessment in breast cancer (ECTIL): multicentre validation in 2340 patients with breast cancer. The Lancet Digital Health. 2025;7(11).
2. Krug K, Jaehnig EJ, Satpathy S, Blumenberg L, Karpova A, Anurag M, et al. Proteogenomic Landscape of Breast Cancer Tumorigenesis and Targeted Therapy. Cell. 2020;183(5):1436–56.e31.
3. Comprehensive molecular portraits of human breast tumours. Nature. 2012;490(7418):61–70.
4. Liu J, Lichtenberg T, Hoadley KA, Poisson LM, Lazar AJ, Cherniack AD, et al. An Integrated TCGA Pan-Cancer Clinical Data Resource to Drive High-Quality Survival Outcome Analytics. Cell. 2018;173(2):400–16.e11.
5. Sammut SJ, Crispin-Ortuzar M, Chin SF, Provenzano E, Bardwell HA, Ma W, et al. Multi-omic machine learning predictor of breast cancer therapy response. Nature. 2022;601(7894):623–9.

Cancer cachexia is a systemic metabolic rewiring resulting due to advanced cancer in severe muscle and fat loss. It markedly reduces tolerance to treatment and contributes to two million deaths annually. Despite its clinical impact, cancer cachexia remains poorly characterised [1]. Positron Emission Tomography (PET) provides a whole-body functional readout of metabolism and is routinely acquired in cancer care [2]. We present an automated PET quantification pipeline to explore the underlying metabolic mechanisms of cachexia through characterisation of organ-level changes, inter-organ relationships, and cancer-type specific cachexia signatures.

Two large multicentre [¹⁸F]fluorodeoxyglucose PET datasets were collated: Kaiser Permanente (lung n=2,289, pancreatic n=572, colorectal n=305) and TRACERx (n=643 NSCLC). Kaiser Permanente patients were stratified by weight-loss trajectory relative to scan date. TRACERx patients were categorised based on body composition changes. Whole-body PET/CT scans were segmented using TotalSegmentator [3], with organ-level SUV adaptive thresholds and growth filters applied for primary and metastatic tumour segmentation to exclude physiologically implausible uptake values. Organ-level metabolic activity was normalised to aortic uptake, yielding Standardised Uptake Value Ratios (SUVR), which mitigate scanner and protocol variability across sites. Organ-level SUVR features were extracted across 70 anatomical structures. Univariate discrimination was assessed using Cliff’s delta effect sizes and the Mann-Whitney U test. Multi-organ classification models (XGBoost, Random Forest) were built using minimum Redundancy Maximum Relevance (mRMR) feature selection. Balanced accuracy and Area Under the Receiver Operating Characteristic Curve (AUROC) were used to measure separation, and feature importances were used to derive metabolic fingerprints. Linear Discriminant Analysis (LDA) characterised metabolic differences across cancer types and demographic subgroups, alongside stratified repetitions of the primary analysis. Mutual information (MI) quantified altered inter-organ metabolic coupling networks, revealing systemic coordination beyond individual organ effects.

Individual organ variations were identified across the pan-cancer cohort in the Kaiser Permanente dataset. Most prominent was reduced uptake across several heart chambers in all three cancer types, also evident as a mild effect prior to weight loss. In colorectal and lung cancer, increased uptake in spleen and bone marrow indicated an inflammatory response. Across most organs, pancreatic cancer showed reduced PET uptake compared to the non-cachectic cohort.

LDA investigations into demographic variation demonstrated that while the severity of FDG uptake changes is influenced by age and sex, the overall trends are broadly consistent. Smoking status and comorbidities, however, had a substantial effect on PET presentation. MI demonstrated changes in muscle and cardiac inter-organ connectivity. The dynamic nature of cachexia was also evident, with several organs such as torso fat and kidney, showing significant uptake changes only at higher levels of weight loss.

Overall separation between cachectic and non-cachectic patients was achieved in Random Forest with 69.7 ± 2.2% balanced accuracy and AUROC of 0.772 ± 0.018 across the pan-cancer cohort.

TRACERx NSCLC analysis is at an earlier stage but has demonstrated trends consistent with the lung cancer Kaiser Permanente cohort, including reduced myocardium uptake at baseline and a similar increase in thyroid gland and skeletal muscle uptake.

This work demonstrates that PET, approached as a high-dimensional whole-body dataset rather than a lesion-level clinical readout, can reveal whole-body metabolic rewiring in cancer cachexia. Integrating this imaging layer with molecular and clinical data is a natural next step.

References

1. Ferrer et al. Cell. 2023
2. Ferrara et al. J Cachexia Sarcopenia Muscle 2024
3. Wasserthal et al. Radiology : AI 2023

Microscopy datasets are often spatially sparse, wherein relevant structures occupy only a small fraction of the total field of view (FOV), leaving large regions of background devoid of signal. This inherent inefficiency creates file sizes that are larger than needed, which increases the time needed for computational image data analysis and processing, and means unnecessarily large data volumes. In this work, a classical open-source method for content-aware spatial compaction of microscopy images (CASC) is reported that explicitly removes spatial redundancy by reorganising foreground objects into a new, smaller image. CASC combines adaptive intensity normalisation, statistical thresholding, morphological refinement, and connected-component analysis to isolate foreground structures. These structures are then extracted with contextual padding and repacked into a compact domain using a heuristic shelf-based spatial packing strategy. CASC intentionally destroys the spatial topology of the image but preserves pixel intensities exactly, retaining object-level information. The method achieves reductions in image area and background content while maintaining high object-level preservation of biological structures, with a reduction in file size of more than 390-fold shown in real biological image datasets.