ABSTRACT. In this presentation that partly subsumes and summarises in the form of adapted excerpts some recent articles of which I am author or co-author [2], [3], [11], I suggest that cancer is fundamentally a disease of the control of cell differentiation in multicellular organisms, uncontrolled cell proliferation being a mere consequence of blockade, or unbalance, of cell differentiations. Cancer cell populations, that can reverse the sense of differentiations, are extremely plastic and able to adapt without mutations their phenotypes in order to transiently resist drug insults [10], which is likely due to the reactivation of ancient, normally silenced, genes [5], [7], [12]. Stepping from mathematical models of non genetic plasticity in cancer cell populations [4], [6] and questions they raise, I propose an evolutionary biology approach to shed light on this problem a) from a theoretical viewpoint by a description of multicellular organisms in terms of multi-level structures, which integrate function and matter from lower to upper levels, and b) from a practical point of view by proposing future tracks for cancer therapeutics, as cancer is primarily a failure of multicellularity in animals and humans. This approach resorts to the emergent field of knowledge known as philosophy of cancer [1], [8], [9].

ABSTRACT. Here, we present new version 2.1 of tugHall (tumor gene-Hallmark) cancer-cell evolution simulator which is accelerated by clone-based approach. The tool is based on the model connected to the well-known cancer hallmarks with the specific mutational states of tumor-related genes. The hallmark variables depend linearly on the mutational states of tumor-related genes with specific weights. The cell behavior phenotypes are stochastically determined and the phenotypic probabilities are probabilistically interfered by the hallmarks. Approximate Bayesian Computation is applied to find the personalized specific parameters of the model. The variant allele frequencies are used as target data for analysis. In tugHall 2.1, the Darwinian evolutionary competition amongst different clones is computed due to clone’s death/birth processes. The open-source code is available in the repository www.github.com/tugHall 

ABSTRACT. Somatic mutations accumulated with age in healthy tissues are one of the major risk factors of cancer. Recently accumulation of specific mutations in DNMT3A, TET2, ASXL1, JAK2, and TP53 genes in hematopoietic tissue, termed clonal hematopoiesis (CH), were observed in healthy volunteers. Here, we design a stochastic mathematical model of hematopoietic stem and progenitor cells to study the evolutionary dynamics of CH. We analyzed a Moran model and found that CH is promoted by age, stem cell division rate, mutation rate, the size of the cell population, and relative fitness of mutated cells. Moreover, using realistic parameter sets we succeeded in reproducing the clinical observation that age-dependent increase of CH. Finally, we will discuss the effect of age-dependent radiation exposure on CH.

ABSTRACT. Background

Immunotherapy is shaping care for patients with advanced cancers. Unfortunately, the vast majority of immunotherapy trials remain negative, hampering clinical benefit for patients and increasing research and development costs. Currently, immunotherapy trials are designed based on ‘black box’ statistical methods that lack a coupling between disease mechanism and clinical trial outcome. We constructed an in silico mechanism-based multiscale model to simulate randomized-controlled immunotherapy trials. This approach enables the explicit incorporation of clinical and biomolecular knowledge to gain power and improve the design of immunotherapy trials. 

Methods

Our model is based on ordinary differential equations describing essential tumor-immune dynamics in the tumor microenvironment. Specifically, our model describes 1) growth of a tumor, 2) activation and proliferation of T cells, 3) migration of these cells into the tumor microenvironment, and 4) T cell-dependent killing of tumor cells. Computational implementation of these equations was performed using the ‘odeint’ function(s) derived from the C++ boost library. Integration of treatment modalities based on variation in model parameters (e.g., the killing capacity of CD8⁺ T cells) enables simulation of individualized disease courses in patients. Available pre-clinical or (early phase) clinical data can be used to fit model parameters. Upscaling this approach by simulating a patient population facilitates clinical trial modeling and helps investigate treatment-induced mechanistic effects on trial outcomes. 

Results

Our computational approach – mechanistic microsimulations – can link mechanistic insight from a patient’s tumor microenvironment to the outcome of a clinical trial. As proof of principle, we demonstrate our models’ ability to generate realistic survival data by fitting the model to the publicly available NCCTG lung cancer survival dataset. Based on (pre)clinical knowledge concerning therapeutic mechanisms, the model parameters can be altered to simulate treatment effects in patients. Specifically, we show that treatments that increase the killing rate of CD8⁺ T cells (e.g., anti-PD-1 checkpoint inhibitors) can induce a substantial survival benefit in a subset of patients. Even more interestingly, the survival kinetics arising as emergent behavior from our model are typical for many immunotherapy trials – namely, it shows a delayed curve separation. This delayed curve separation implies that a survival benefit only becomes apparent at the later stages of the trial. While conventional statistical methods cannot incorporate these unique immunotherapy-induced survival kinetics into the planning of randomized-controlled immunotherapy trials, they are crucial to consider. 

The importance of these typical kinetics becomes evident in two applications of microsimulations: sample size calculations and endpoint selection for randomized-controlled immunotherapy trials. Sample size calculations determine the number of required participants in an immunotherapy trial to show a particular treatment effect – mostly, a survival benefit. We demonstrate that conventional formula-based sample size calculations overestimate the power of immunotherapy trials: while considering immunotherapy-specific survival kinetics, microsimulations estimate that to increase the 2-year overall survival with 15-20%, around 30 additional patients are required per treatment arm in a 1 : 1 randomized-controlled trial to reach a power of 80% compared to the conventional calculation. 

In a related experiment with identical effect size, similar observations come to light with regard to endpoint selection. According to classical methods (i.e., in the absence of a delayed curve separation), a randomized immunotherapy trial (n = 800) with 2-year overall survival as primary endpoint possesses an unrealistically high power near 100%. Again, microsimulations are able to consider the survival kinetics, leading to a more realistic power of approximately 35% with 2-year overall survival as endpoint. A power >80% would be obtained after 36 months; only after 48 months, the power approaches 100%. 

Microsimulation-based predictions increase the trial design accuracy, preventing futile exposure of patients to potential adverse events and increasing the cost-effectiveness of immunotherapy trials in clinical practice. 

Conclusion

Mechanistic microsimulations enable clinicians to design more robust and cost-effective immunotherapy trials by explicitly incorporating biomolecular and clinical knowledge into the design process. 

ABSTRACT. Human Leukocyte Antigen (HLA) receptors play a key role in cellular immunity, binding intracellular peptides and displaying them for recognition by T-cell lymphocytes. T-cell activation is partially driven by structural features of these peptide-HLA complexes, making structural modeling and analysis of these complexes a desired goal for cancer immunotherapy projects. Unfortunately, these analyses are limited by their computational cost, and the small number of experimentally-determined structures of peptide-HLA complexes. Here we describe HLAArena, a computational environment designed to overcome the challenges associated with structural modeling, visualization, and analysis of peptide-HLA complexes. To illustrate the capabilities of HLA-Arena, we conducted a large-scale virtual screening of peptides for multiple HLA alleles. Our dataset included experimentally-determined binders, as well as decoys, and our results demonstrate the enrichment of true binders among top scoring peptides. HLA-Arena can be integrated within larger computational pipelines, and could be used to conduct structural analyses for personalized cancer immunotherapy, neoantigen discovery or vaccine development.

ABSTRACT. Recently developed technologies allow us to view the transcriptome at high resolution while preserving the spatial location of samples. These advances are particularly relevant to cancer research, since clonal theory predicts that nearby cells are likely to belong to the same expanding subclone. Using this evolutionary hypothesis, we develop a statistical procedure which uses a test of local spatial association along with a graph-based approach to infer subclones from spatially resolved RNA-seq data. Our method is robust, scalable, and can be applied to data from any of the existing spatial transcriptomics technologies. On data from spatially resolved RNA-seq of breast cancer tissue, our method infers seven distinct subclones and identifies potential driver genes.

ABSTRACT. Synonymous mutations inside the coding region, which do not alter the amino acid chain, are usually considered to have no effect on the protein. However, in recent years it was shown that they may regulate expression levels via various mechanisms, suggesting that they may also play an important role in tumorigenesis. In the current study, we suggest a pipeline for detecting cancerous silent mutations that affect the cancer fitness via regulation of transcription. We demonstrate our approach by reporting for the first-time cases where cancerous synonymous mutations inside the coding regions of the gene BCL2 are under selection in germinal center B-cell (GCB) lymphoma patients. We provide various lines of evidence that suggest that these mutations contribute to the cancer cell survival via improving the expression levels of anti-apoptotic BCL2 protein.

ABSTRACT. Introduction: Rising cancer care costs impose financial burdens on health systems. Applying artificial intelligence (AI) to diagnostic algorithms may reduce testing costs and avoid wasteful therapy-related expenditures. To evaluate the financial and clinical impact of incorporating artificial intelligence-based determination of mismatch repair/microsatellite instability (MMR/MSI) status into the first-line metastatic colorectal carcinoma (mCRC) setting.

Methods: We developed a financial model to compare financial and clinical impact of eight mCRC tumor tissue MMR/MSI strategies: next-generation sequencing (NGS) alone, high-sensitivity polymerase chain reaction (PCR) or immunohistochemistry (IHC) panel (“panel,” for short) alone, high-specificity panel alone, high-specificity AI alone, high-sensitivity AI followed by NGS, high-specificity AI followed by NGS, high-sensitivity AI followed by high-sensitivity panel, and high-sensitivity AI followed by high-specificity panel. We modeled the United States health-care system, using a hypothetical, nationally representative, population-based sample of individuals receiving first-line treatment for de novo mCRC (N = 32,549). Model inputs were derived from secondary research (peer-reviewed literature and Medicare data). We estimated the population-level diagnostic costs, treatment costs, and clinical implications associated with each testing strategy.

Results: The testing strategy that resulted in the greatest project cost savings (including testing and first-line drug cost) compared to NGS alone in newly-diagnosed mCRC was using high-sensitivity AI followed by confirmatory high-specificity PCR or IHC panel for patients testing negative by AI ($400 million in savings, 12.9%). The high-specificity AI-only strategy resulted in the most favorable clinical impact, with 97% diagnostic accuracy in guiding genotype-directed treatment and average time to treatment initiation of less than one day.

Conclusion: AI has the potential to reduce both time to treatment initiation and costs in the mCRC setting without meaningfully sacrificing diagnostic accuracy. We expect the AI value proposition to improve in coming years, with increasing diagnostic accuracy and decreasing costs of processing power. To extract maximal value from the technology, health systems should evaluate integrating diagnostic histopathologic artificial intelligence into institutional protocols, perhaps in place of other genotyping methodologies.

ABSTRACT. Recently, there has been an explosion of slide-free microscopy techniques for pathology, designed to simplify the traditional histology workflow for faster specimen processing and imaging. These include techniques such as microscopy with ultraviolet surface excitation (MUSE), stimulated Raman histology (SRH), and confocal and multiphoton microscopy, structured illumination, optical coherence microscopy and others.

Despite the recent interest in such techniques, the current gold standard in histology is still brightfield imaging of hematoxylin-and-eosin (H&E) stained tissue slides; these dyes render cell nuclei blue and cytoplasm pink, respectively. However, other techniques often differ from the traditional brightfield hues and can sometimes contain more than two major sets of shading contrast sources. In order to bridge the gap between slide-free techniques and traditional histological examination, we aim to digitally modify slide-free microscopy images to match H&E images. We especially focus on MUSE imaging. In previous work, a spectral unmixing color mapping model was used, but this required user input of expected colors and is limited to conversion of the major nuclear and cytoplasm stain pairs, failing to handle cases in which a larger gamut of colors is generated. Therefore, we aim to utilize deep learning methodologies to learn the appropriate transformation for generating visually convincing virtual H&E (vH&E) images that works well on a variety of tissue and cell types. We focus on applying our techniques for imaging of tumors, where slide-free techniques can provide a significant benefit in accelerating clinical decision-making, for example, for providing intraoperative surgical guidance.

To develop a deep learning model for this task, we used unpaired image-to-image translation techniques. While other modality conversion techniques also use generative adversarial networks (GANs), such methods require imaging of identical samples (images) using both modalities (paired), a constraint that typically cannot be satisfied with slide-free techniques like MUSE. Unpaired image-to-image translation models such as cycle-consistent GANs (CycleGANs) on the other hand, are able to convert images between domains using unpaired exam-ples. Here we compared several image translation models: CycleGAN, GANILLA, and DualGAN. Other architectural and training modifications were also tested in order to study which aspects of the model are relevant for successful predictions. Finally, we used a non-AI-based, deterministic color mapper see what can be accomplished without machine-learning input. The models were trained on either 512 x 512 patches or 256 x 256 patches of the specimen image. To apply the resulting conversion models to larger images, the latter were separated into patches and a post-conversion montage construction algorithm was developed. To quantitatively evaluate conversion, we trained an external model to discriminate between real and vH&E images. Success would be signalled if the critic could not distinguish real from virtual images. Experiments (training and testing) were run with datasets derived from anonymized specimens of urothelial carcinoma and glioblastoma tissues.

We observed that training with unaltered MUSE images (example shown in Fig. 1a) generated poor results, failing to appropriately transfer the content of the MUSE image to a vH&E image (example shown in Fig. 1b). The CycleGAN interpreted bright fluorescent nuclear stains as background white spaces (red box, Fig. 1a,b), and, conversely, converted darker regions to contain fictitious (hallucinated) nuclear structures (yellow box, Fig. 1a,b). On the other hand, if MUSE images were initially color- and intensity-inverted, and then used as input (Fig 1c), appropriate color conversions were obtained (Fig. 1d). Based on both qualitative and quantitative evaluation, the CycleGAN model outperformed the other models. We observed that simply applying the model to contiguous patches from large image regions, followed by naïve stitching resulted in prominent tiling artifacts. To address this, we utilized overlapping patches and by averaging between overlapping regions, we obtained large converted images without evident tiling.

We have demonstrated successful MUSE-to-H&E modality conversion using deep learning. In the future, we plan to obtain pathologist ratings as a more reliable metric for model performance. Additionally, we hope to adopt neural com-pression techniques to allow inference on resource-constrained hardware such as smaller GPUs or even CPUs alone. We hope that our framework for converting slide-free microscopy images, generated potentially by a variety of methods, to H&E-like images will help catalyze the adoption of novel slide-free techniques and ultimately improve the speed and efficiency of pathology workflow.

ABSTRACT. A cancer initiates, grows and metastasizes over time and space. It often involves dynamic, genotypical and phenotypical evolution and interaction of millions of cells, belonging to hundreds of cell types. Successful cancer prevention and treatment require quantitative approaches that can identify key factors that are causal to cancer evolution and can be therapeutically intervened. Achieving such a goal has been challenging, due partly to limitations in data collection, analysis and interpretation. In this talk, I will highlight ongoing efforts that involve various aspects of experimental design, application of high-throughput multiomics technologies such as single-cell DNA, RNA and ATAC sequencing, and statistical computational approaches to tackle such an important challenge.

ABSTRACT. Pancreatic ductal adenocarcinoma (PDAC) patients, who often present with stage III or IV disease, face a dismal prognosis as the 5-year survival rate remains below 10%. Recent studies have revealed that CD4+ T, CD8+ T, and/or B cells in specific spatial arrangements relative to intratumoral regions correlate with clinical outcome for patients, but the complex functional states of those immune cell types remain to be incorporated into prognostic biomarker studies. Here, we developed an interpretable machine learning model to analyze the functional relationship between leukocyte-leukocyte or leukocyte-tumor cell spatial proximity, correlated with clinical outcome of 46 therapy-naïve PDAC patients following surgical resection. Using a multiplex immunohistochemistry imaging data set focused on profiling leukocyte functional status, our model identified features that distinguished patients in the fourth quartile from those in the first quartile of survival. The top ranked important features identified by our model, all of which were positive prognostic stratifiers, included CD4+ T helper cell frequency among CD45+ immune cells, frequency of Granzyme B-positivity among CD4+ and CD8+ T cells, as well as the frequency of PD-1 positivity among CD8+ T cells. The spatial proximity of CD4+ T to B cells, and between CD8+ T cells and epithelial cells, were also identified as important prognostic features. While spatial proximity features provided valuable prognostic information, the best model required both spatial and phenotypic information about tumor infiltrating leukocytes. Our analysis links the immune microenvironment of PDAC tumors to outcome of patients, thus identifying features associated with more progressive disease.

ABSTRACT. Noninvasive follicular thyroid neoplasms with papillary-like nuclear features (NIFTP) are follicular-patterned thyroid neoplasms defined by nuclear atypia and indolent behavior. They harbor RAS mutations, rather than BRAF mutations as is observed in papillary thyroid carcinomas with extensive follicular growth (PTC-EFGs). Reliably identifying NIFTPs aids in safe therapy de-escalation, but has proven to be challenging due to interobserver variability and morphologic heterogeneity. While identifying this neoplasm by genetic features is attractive, not all NIFTPs harbor RAS mutations. The genomic scoring system BRS (BRAF-RAS score) was developed to quantify the extent to which a tumor’s expression profile resembles a BRAF or RAS-mutant neoplasm. We proposed that deep learning prediction of BRS could differentiate NIFTP from other follicular-patterned neoplasms. A deep learning model was trained on slide images from a dataset of 115 thyroid neoplasms to predict tumor subtype (NIFTP, PTC-EFG, or classic PTC), and was used to generate predictions for 497 thyroid neoplasms within The Cancer Genome Atlas (TCGA). Most follicular-patterned neoplasms in the TCGA test set were predicted to be NIFTP by this model, rather than PTC-EFG (72% vs 17%). Within follicular-patterned neoplasms, tumors with positive BRS (RAS-like) were 8.5 times as likely to carry a NIFTP prediction than tumors with negative BRS (89.7% vs 10.5%, P<0.0001). To test the hypothesis that BRS may serve as a surrogate for biological processes that determine tumor subtype, a separate model was then trained on TCGA slides to predict BRS as a linear outcome. This model performed well in cross-validation on the training set (R2=0.67, dichotomized AUC=0.94). A final predictive model was trained across the entire TCGA cohort and used to generate BRS predictions on our internal cohort. NIFTPs were near universally predicted to have RAS-like BRS; as a sole discriminator of NIFTP status, predicted BRS performed with an AUC of 0.99 globally and 0.97 when restricted to follicular-patterned neoplasms. BRAF-mutant PTC-EFG had BRAF-like predicted BRS (mean -0.49), non-mutant PTC-EFG had more intermediate predicted BRS (mean -0.17), and NIFTP had RAS-like BRS (mean 0.35; P<0.0001). In summary, histologic features associated with the BRAF-RAS gene expression spectrum are detectable by deep learning and can aid in distinguishing indolent NIFTP from PTCs. These results advocate for further validation of the use of BRS testing as a diagnostic aid for follicular-patterned neoplasms.

ABSTRACT. The objective of this work is to confront artificial neural network models with time-to-event data, using specific ways to handle censored observations such as pseudo-observations and tailored loss functions. Different neural network models were compared. Cox-CC (Kvamme et al., 2019) uses a loss function based on a case-control approximation. DeepHit (Lee et al., 2018) is a model that estimates the probability mass function and combines log-likelihood with a ranking loss. DNNSurv (Zhao and Feng, 2019) circumvent the problem of censoring by using pseudo-observations. We also proposed other ways of computing pseudo-observations. We investigated the prediction ability of these models using data simulated from an AFT model as proposed by Friedman (2001), with different censoring rates. We simulated 100 datasets of 4,000 samples and 20 variables each, with pairwise interactions and non-linear effects of random subsets of these variables. Models were compared using the concordance index and integrated Brier score. We applied the methods to the METABRIC breast cancer data set, including 1,960 patients, 6 clinical covariates and the expression of 863 genes. In the simulation study, we obtained the highest C-indices and lower integrated Brier score with CoxTime for low censoring and pseudo-discrete with high censoring. On the METABRIC data, the models obtained very comparable 5-year and 10-year discrimination performances with slightly higher values for the models based on optimised pseudo-observations.