Machine Learning-Powered High-Content Image Analysis of H&E Slides Predicts Unfavourable Outcomes in HPV+ Oropharyngeal Squamous Cell Carcinoma Patients

Patients with Human Papillomavirus positive oropharyngeal Squamous Cell Carcinoma (HPV+opSCC) have better prognosis than HPV- counterparts, raising the possibility of treatment de-escalation in the former. About 20% of HPV+opSCC patients demonstrate unfavourable outcomes, contraindicating de-escalation regimes in this subgroup; however, diagnostic procedures to identify such patients are currently unavailable. To address this issue, we developed an automated workflow for quantitative high content image analysis (HCA) of H&E stained sections of HPV+opSCCs. We quantified known prognostic features such as number and spatial distribution of tumour-infiltrating lymphocytes (TILs) in the stromal and intra-tumour regions. In addition, we measured and attempted to identify other less established features such as stromal plasma cells, tumour nuclei features and morphological heterogeneity within tumour cells. Finally, we trained and validated a model to retrospectively prognosticate outcomes in a cohort of 58 HPV+opSCC patients. Univariate and multivariate statistical analyses revealed that plasma cells, stromal and intra-tumour TILs were more numerous in favourable outcome (FO) patients. Tumour cell nuclei were rounder, less eccentric in morphology and packed closer to one another in patients with FO. Tumour nuclei in FO had more nucleoli and higher texture and granularity features than patients with unfavourable outcomes (UO). UO patients had greater tumour heterogeneity in morphological, spatial and textural measurements. To attempt separating the groups according to these variables we performed statistical discriminant analyses (either LDA or QDA). QDA had an accuracy of 81.7% and 88.1% in predicting UO and FO on our cohort. We validated our analysis by k-fold cross-validation, revealing an estimated overall accuracy of 76.2% and a Kappa statistic of 0.523, indicating a good model considering the complexity of the problem at hand. Single-cell quantitative image analysis of HPV+opSCC allows us to identify prognostic factors and quantify their heterogeneity within the tumour. We have shown that some of these measures are predictive in their own right, and that their variance within a tumour can itself be prognostic and improve the accuracy of statistical discriminant models. Our open-source HCA workflow on routine H&E slides and statistical modelling can aid prognostication of HPV+opSCCs outcome with promising accuracy. Our work supports the use of ML-powered HCA followed by statistical modelling in digital pathology to exploit clinically relevant features in routine diagnostic pathology without additional biomarkers. For more details see: https://www.medrxiv.org/content/10.1101/2022.06.24.22276368v1 Multi-scale and Machine Learning Algorithms for Modeling Large Blood Clots Yuefan Deng Stony Brook University Multiscale modeling in biomedical engineering is gaining momentum because of progress in supercomputing, applied mathematics, and quantitative biomedical engineering. For example, scientists in various disciplines have been advancing, slowly but steadily, the simulation of blood including its flowing and the physiological properties of such components as red blood cells, white blood cells, and platelets. Aggregated platelets stimulate blood clotting that causes heart attacks and strokes, resulting in more than 20 million deaths annually (for a comparison, the lethal Covid-19 causes 5.7 million deaths as of Jan. 2022). To reduce such deaths, we must discover new drugs. To discover new drugs, we must understand the mechanism of platelet activation and aggregation. To model platelets’ dynamics involves setting up the basic space and time discretization in huge ranges of 5-6 orders of magnitudes, resulting from the relevant fundamental interactions at atomic, to molecular, to cell, to fluid scales. To achieve the desired accuracy at the minimal computational costs, we must select the correct physiological parameters in the force fields such as the Morse potential and Hooke’s law as well as the spatial and temporal discretization, by machine learning. We demonstrate our results of a multiscale 250-platelet (125 million particles) aggregation simulation and their corroborations with in vitro experiments. For more details see: https://arxiv.org/abs/2205.14121 *Contents* 0:00 - Introduction 5:46 - Jonas Hue 22:32 - Yuefan Deng 44:35 - Q&A Session Check out our other videos on modeling, infection and immunology: https://youtube.com/playlist?list=PLiEtieOeWbMKh9VcQoinSwODcSZKMTGat Consider joining our IMAG/MSM WG on Multiscale Modeling and Viral Pandemics: https://www.imagwiki.nibib.nih.gov/content/msm-viral-pandemics-meetings Consider joining the Global Alliance for Immune Prediction and Intervention: http://glimprint.org/