Eugene Fluder

Artificial intelligence models using digital histopathology slides stained with hematoxylin and eosin offer promising, tissue-preserving diagnostic tools for patients with cancer. Despite their advantages, their clinical utility in real-world settings remains unproven. Assessing EGFR mutations in lung adenocarcinoma demands rapid, accurate and cost-effective tests that preserve tissue for genomic sequencing. PCR-based assays provide rapid results but with reduced accuracy compared with next-generation sequencing and require additional tissue. Computational biomarkers leveraging modern foundation models can address these limitations. Here we assembled a large international clinical dataset of digital lung adenocarcinoma slides (N = 8,461) to develop a computational EGFR biomarker. Our model fine-tunes an open-source foundation model, improving task-specific performance with out-of-center generalization and clinical-grade accuracy on primary and metastatic specimens (mean area under the curve: internal 0.847, external 0.870). To evaluate real-world clinical translation, we conducted a prospective silent trial of the biomarker on primary samples, achieving an area under the curve of 0.890. The artificial-intelligence-assisted workflow reduced the number of rapid molecular tests needed by up to 43% while maintaining the current clinical standard performance. Our retrospective and prospective analyses demonstrate the real-world clinical utility of a computational pathology biomarker.

Visualizations of biomolecular networks assist in systems-level data exploration in many cellular processes. Data generated from high-throughput experiments increasingly inform these networks, yet current tools do not adequately scale with concomitant increase in their size and complexity. We present an open source software platform, interactome-CAVE (iCAVE), for visualizing large and complex biomolecular interaction networks in 3D. Users can explore networks (i) in 3D using a desktop, (ii) in stereoscopic 3D using 3D-vision glasses and a desktop, or (iii) in immersive 3D within a CAVE environment. iCAVE introduces 3D extensions of known 2D network layout, clustering, and edge-bundling algorithms, as well as new 3D network layout algorithms. Furthermore, users can simultaneously query several built-in databases within iCAVE for network generation or visualize their own networks (e.g., disease, drug, protein, metabolite). iCAVE has modular structure that allows rapid development by addition of algorithms, datasets, or features without affecting other parts of the code. Overall, iCAVE is the first freely available open source tool that enables 3D (optionally stereoscopic or immersive) visualizations of complex, dense, or multi-layered biomolecular networks. While primarily designed for researchers utilizing biomolecular networks, iCAVE can assist researchers in any field.

Recent breakthroughs in self-supervised learning have enabled the use of large unlabeled datasets to train visual foundation models that can generalize to a variety of downstream tasks. While this training paradigm is well suited for the medical domain where annotations are scarce, large-scale pre-training in the medical domain, and in particular pathology, has not been extensively studied. Previous work in self-supervised learning in pathology has leveraged smaller datasets for both pre-training and evaluating downstream performance. The aim of this project is to train the largest academic foundation model and benchmark the most prominent self-supervised learning algorithms by pre-training and evaluating downstream performance on large clinical pathology datasets. We collected the largest pathology dataset to date, consisting of over 3 billion images from over 423 thousand microscopy slides. We compared pre-training of visual transformer models using the masked autoencoder (MAE) and DINO algorithms. We evaluated performance on six clinically relevant tasks from three anatomic sites and two institutions: breast cancer detection, inflammatory bowel disease detection, breast cancer estrogen receptor prediction, lung adenocarcinoma EGFR mutation prediction, and lung cancer immunotherapy response prediction. Our results demonstrate that pre-training on pathology data is beneficial for downstream performance compared to pre-training on natural images. Additionally, the DINO algorithm achieved better generalization performance across all tasks tested. The presented results signify a phase change in computational pathology research, paving the way into a new era of more performant models based on large-scale, parallel pre-training at the billion-image scale.

As personalized medicine becomes more integrated into healthcare, the rate at which human genomes are being sequenced is rising quickly together with a concomitant acceleration in compute and storage requirements. To achieve the most effective solution for genomic workloads without re-architecting the industry-standard software, we performed a rigorous analysis of usage statistics, benchmarks and available technologies to design a system for maximum throughput. We share our experiences designing a system optimized for the "Genome Analysis ToolKit (GATK) Best Practices" whole genome DNA and RNA pipeline based on an evaluation of compute, workload and I/O characteristics. The characteristics of genomic-based workloads are vastly different from those of traditional HPC workloads, requiring different configurations of the scheduler and the I/O subsystem to achieve reliability, performance and scalability. By understanding how our researchers and clinicians work, we were able to employ techniques not only to speed up their workflow yielding improved and repeatable performance, but also to make more efficient use of storage and compute resources.

The productivity of computational biologists is limited by the speed of their workflows and subsequent overall job throughput. Because most biomedical researchers are focused on better understanding scientific phenomena rather than developing and optimizing code, a computing and data system implemented in an adventitious and/or non-optimized manner can impede the progress of scientific discovery. In our experience, most computational, life-science applications do not generally leverage the full capabilities of high-performance computing, so tuning a system for these applications is especially critical. To optimize a system effectively, systems staff must understand the effects of the applications on the system. Effective stewardship of the system includes an analysis of the impact of the applications on the compute cores, file system, resource manager and queuing policies. The resulting improved system design, and enactment of a sustainability plan, help to enable a long-term resource for productive computational and data science. We present a case study of a typical biomedical computational workload at a leading academic medical center supporting over $100 million per year in computational biology research. Over the past eight years, our high-performance computing system has enabled over 900 biomedical publications in four major areas: genetics and population analysis, gene expression, machine learning, and structural and chemical biology. We have upgraded the system several times in response to trends, actual usage, and user feedback. Major components crucial to this evolution include scheduling structure and policies, memory size, compute type and speed, parallel file system capabilities, and deployment of cloud technologies. We evolved a 70 teraflop machine to a 1.4 petaflop machine in seven years and grew our user base nearly 10-fold. For long-term stability and sustainability, we established a chargeback fee structure. Our overarching guiding principle for each progression has been to increase scientific throughput and enable enhanced scientific fidelity with minimal impact to existing user workflows or code. This highly-constrained system optimization has presented unique challenges, leading us to adopt new approaches to provide constructive pathways forward. We share our practical strategies resulting from our ongoing growth and assessments.