Pacific Symposium on Biocomputing. Pacific Symposium on Biocomputing最新文献

Advancements in medical imaging and artificial intelligence (AI) have revolutionized the field of cardiac diagnostics, providing accurate and efficient tools for assessing cardiac function. AI diagnostics claims to improve upon the human-to-human variation that is known to be significant. However, when put in practice, for cardiac ultrasound, AI models are being run on images acquired by human sonographers whose quality and consistency may vary. With more variation than other medical imaging modalities, variation in image acquisition may lead to out-of-distribution (OOD) data and unpredictable performance of the AI tools. Recent advances in ultrasound technology has allowed the acquisition of both 3D as well as 2D data, however 3D has more limited temporal and spatial resolution and is still not routinely acquired. Because the training datasets used when developing AI algorithms are mostly developed using 2D images, it is difficult to determine the impact of human variation on the performance of AI tools in the real world. The objective of this project is to leverage 3D echos to simulate realistic human variation of image acquisition and better understand the OOD performance of a previously validated AI model. In doing so, we develop tools for interpreting 3D echo data and quantifiably recreating common variation in image acquisition between sonographers. We also developed a technique for finding good standard 2D views in 3D echo volumes. We found the performance of the AI model we evaluated to be as expected when the view is good, but variations in acquisition position degraded AI model performance. Performance on far from ideal views was poor, but still better than random, suggesting that there is some information being used that permeates the whole volume, not just a quality view. Additionally, we found that variations in foreshortening didn't result in the same errors that a human would make.

Artificial Intelligence (AI) models are substantially enhancing the capability to analyze complex and multi-dimensional datasets. Generative AI and deep learning models have demonstrated significant advancements in extracting knowledge from unstructured text, imaging as well as structured and tabular data. This recent breakthrough in AI has inspired research in medicine, leading to the development of numerous tools for creating clinical decision support systems, monitoring tools, image interpretation, and triaging capabilities. Nevertheless, comprehensive research is imperative to evaluate the potential impact and implications of AI systems in healthcare. At the 2024 Pacific Symposium on Biocomputing (PSB) session entitled "Artificial Intelligence in Clinical Medicine: Generative and Interactive Systems at the Human-Machine Interface", we spotlight research that develops and applies AI algorithms to solve real-world problems in healthcare.

The greatest known risk factor for Alzheimer's disease (AD) is age. While both normal aging and AD pathology involve structural changes in the brain, their trajectories of atrophy are not the same. Recent developments in artificial intelligence have encouraged studies to leverage neuroimaging-derived measures and deep learning approaches to predict brain age, which has shown promise as a sensitive biomarker in diagnosing and monitoring AD. However, prior efforts primarily involved structural magnetic resonance imaging and conventional diffusion MRI (dMRI) metrics without accounting for partial volume effects. To address this issue, we post-processed our dMRI scans with an advanced free-water (FW) correction technique to compute distinct FW-corrected fractional anisotropy (FAFWcorr) and FW maps that allow for the separation of tissue from fluid in a scan. We built 3 densely connected neural networks from FW-corrected dMRI, T1-weighted MRI, and combined FW+T1 features, respectively, to predict brain age. We then investigated the relationship of actual age and predicted brain ages with cognition. We found that all models accurately predicted actual age in cognitively unimpaired (CU) controls (FW: r=0.66, p=1.62x10-32; T1: r=0.61, p=1.45x10-26, FW+T1: r=0.77, p=6.48x10-50) and distinguished between CU and mild cognitive impairment participants (FW: p=0.006; T1: p=0.048; FW+T1: p=0.003), with FW+T1-derived age showing best performance. Additionally, all predicted brain age models were significantly associated with cross-sectional cognition (memory, FW: β=-1.094, p=6.32x10-7; T1: β=-1.331, p=6.52x10-7; FW+T1: β=-1.476, p=2.53x10-10; executive function, FW: β=-1.276, p=1.46x10-9; T1: β=-1.337, p=2.52x10-7; FW+T1: β=-1.850, p=3.85x10-17) and longitudinal cognition (memory, FW: β=-0.091, p=4.62x10-11; T1: β=-0.097, p=1.40x10-8; FW+T1: β=-0.101, p=1.35x10-11; executive function, FW: β=-0.125, p=1.20x10-10; T1: β=-0.163, p=4.25x10-12; FW+T1: β=-0.158, p=1.65x10-14). Our findings provide evidence that both T1-weighted MRI and dMRI measures improve brain age prediction and support predicted brain age as a sensitive biomarker of cognition and cognitive decline.

The concept of a digital twin came from the engineering, industrial, and manufacturing domains to create virtual objects or machines that could inform the design and development of real objects. This idea is appealing for precision medicine where digital twins of patients could help inform healthcare decisions. We have developed a methodology for generating and using digital twins for clinical outcome prediction. We introduce a new approach that combines synthetic data and network science to create digital twins (i.e. SynTwin) for precision medicine. First, our approach starts by estimating the distance between all subjects based on their available features. Second, the distances are used to construct a network with subjects as nodes and edges defining distance less than the percolation threshold. Third, communities or cliques of subjects are defined. Fourth, a large population of synthetic patients are generated using a synthetic data generation algorithm that models the correlation structure of the data to generate new patients. Fifth, digital twins are selected from the synthetic patient population that are within a given distance defining a subject community in the network. Finally, we compare and contrast community-based prediction of clinical endpoints using real subjects, digital twins, or both within and outside of the community. Key to this approach are the digital twins defined using patient similarity that represent hypothetical unobserved patients with patterns similar to nearby real patients as defined by network distance and community structure. We apply our SynTwin approach to predicting mortality in a population-based cancer registry (n=87,674) from the Surveillance, Epidemiology, and End Results (SEER) program from the National Cancer Institute (USA). Our results demonstrate that nearest network neighbor prediction of mortality in this study is significantly improved with digital twins (AUROC=0.864, 95% CI=0.857-0.872) over just using real data alone (AUROC=0.791, 95% CI=0.781-0.800). These results suggest a network-based digital twin strategy using synthetic patients may add value to precision medicine efforts.

A single gene can produce multiple transcripts with distinct molecular functions. Rare-variant association tests often aggregate all coding variants across individual genes, without accounting for the variants' presence or consequence in resulting transcript isoforms. To evaluate the utility of transcript-aware variant sets, rare predicted loss-of-function (pLOF) variants were aggregated for 17,035 protein-coding genes using 55,558 distinct transcript-specific variant sets. These sets were tested for their association with 728 circulating proteins and 188 quantitative phenotypes across 406,921 individuals in the UK Biobank. The transcript-specific approach resulted in larger estimated effects of pLOF variants decreasing serum cis-protein levels compared to the gene-based approach (pbinom ≤ 2x10-16). Additionally, 251 quantitative trait associations were identified as being significant using the transcript-specific approach but not the gene-based approach, including PCSK5 transcript ENST00000376752 and standing height (transcript-specific statistic, P = 1.3x10-16, effect = 0.7 SD decrease; gene-based statistic, P = 0.02, effect = 0.05 SD decrease) and LDLR transcript ENST00000252444 and apolipoprotein B (transcript-specific statistic, P = 5.7x10-20, effect = 1.0 SD increase; gene-based statistic, P = 3.0x10-4, effect = 0.2 SD increase). This approach demonstrates the importance of considering the effect of pLOFs on specific transcript isoforms when performing rare-variant association studies.

The drug development pipeline for a new compound can last 10-20 years and cost over $10 billion. Drug repurposing offers a more time- and cost-effective alternative. Computational approaches based on network graph representations, comprising a mixture of disease nodes and their interactions, have recently yielded new drug repurposing hypotheses, including suitable candidates for COVID-19. However, these interactomes remain aggregate by design and often lack disease specificity. This dilution of information may affect the relevance of drug node embeddings to a particular disease, the resulting drug-disease and drug-drug similarity scores, and therefore our ability to identify new targets or drug synergies. To address this problem, we propose constructing and learning disease-specific hypergraphs in which hyperedges encode biological pathways of various lengths. We use a modified node2vec algorithm to generate pathway embeddings. We evaluate our hypergraph's ability to find repurposing targets for an incurable but prevalent disease, Alzheimer's disease (AD), and compare our ranked-ordered recommendations to those derived from a state-of-the-art knowledge graph, the multiscale interactome. Using our method, we successfully identified 7 promising repurposing candidates for AD that were ranked as unlikely repurposing targets by the multiscale interactome but for which the existing literature provides supporting evidence. Additionally, our drug repositioning suggestions are accompanied by explanations, eliciting plausible biological pathways. In the future, we plan on scaling our proposed method to 800+ diseases, combining single-disease hypergraphs into multi-disease hypergraphs to account for subpopulations with risk factors or encode a given patient's comorbidities to formulate personalized repurposing recommendations.Supplementary materials and code: https://github.com/ayujain04/psb_supplement.

Large Language Models (LLMs) are a type of artificial intelligence that has been revolutionizing various fields, including biomedicine. They have the capability to process and analyze large amounts of data, understand natural language, and generate new content, making them highly desirable in many biomedical applications and beyond. In this workshop, we aim to introduce the attendees to an in-depth understanding of the rise of LLMs in biomedicine, and how they are being used to drive innovation and improve outcomes in the field, along with associated challenges and pitfalls.

Digital health technologies such as wearable devices have transformed health data analytics, providing continuous, high-resolution functional data on various health metrics, thereby opening new avenues for innovative research. In this work, we introduce a new approach for generating causal hypotheses for a pair of a continuous functional variable (e.g., physical activities recorded over time) and a binary scalar variable (e.g., mobility condition indicator). Our method goes beyond traditional association-focused approaches and has the potential to reveal the underlying causal mechanism. We theoretically show that the proposed scalar-function causal model is identifiable with observational data alone. Our identifiability theory justifies the use of a simple yet principled algorithm to discern the causal relationship by comparing the likelihood functions of competing causal hypotheses. The robustness and applicability of our method are demonstrated through simulation studies and a real-world application using wearable device data from the National Health and Nutrition Examination Survey.

We present a fully automated AI-based system for intensive monitoring of cognitive symptoms of neurotoxicity that frequently appear as a result of immunotherapy of hematologic malignancies. Early manifestations of these symptoms are evident in the patient's speech in the form of mild aphasia and confusion and can be detected and effectively treated prior to onset of more serious and potentially life-threatening impairment. We have developed the Automated Neural Nursing Assistant (ANNA) system designed to conduct a brief cognitive assessment several times per day over the telephone for 5-14 days following infusion of the immunotherapy medication. ANNA uses a conversational agent based on a large language model to elicit spontaneous speech in a semi-structured dialogue, followed by a series of brief language-based neurocognitive tests. In this paper we share ANNA's design and implementation, results of a pilot functional evaluation study, and discuss technical and logistic challenges facing the introduction of this type of technology in clinical practice. A large-scale clinical evaluation of ANNA will be conducted in an observational study of patients undergoing immunotherapy at the University of Minnesota Masonic Cancer Center starting in the Fall 2023.

Assembling an "integrated structural map of the human cell" at atomic resolution will require a complete set of all human protein structures available for interaction with other biomolecules - the human protein structure targetome - and a pipeline of automated tools that allow quantitative analysis of millions of protein-ligand interactions. Toward this goal, we here describe the creation of a curated database of experimentally determined human protein structures. Starting with the sequences of 20,422 human proteins, we selected the most representative structure for each protein (if available) from the protein database (PDB), ranking structures by coverage of sequence by structure, depth (the difference between the final and initial residue number of each chain), resolution, and experimental method used to determine the structure. To enable expansion into an entire human targetome, we docked small molecule ligands to our curated set of protein structures. Using design constraints derived from comparing structure assembly and ligand docking results obtained with challenging protein examples, we here propose to combine this curated database of experimental structures with AlphaFold predictions and multi-domain assembly using DEMO2 in the future. To demonstrate the utility of our curated database in identification of the human protein structure targetome, we used docking with AutoDock Vina and created tools for automated analysis of affinity and binding site locations of the thousands of protein-ligand prediction results. The resulting human targetome, which can be updated and expanded with an evolving curated database and increasing numbers of ligands, is a valuable addition to the growing toolkit of structural bioinformatics.