Search

Artificial Intelligence (AI) in healthcare has generated remarkable innovation and progress in the last decade. Significant advancements can be attributed to the utilization of AI to transform physiology data to advance healthcare. In this review, we will explore how past work has shaped the field and defined future challenges and directions. In particular, we focus on three areas of development. First, we give an overview of AI, with special attention to the most relevant AI models. We then detail how physiology data has been harnessed by AI to advance the main areas of healthcare such as automating existing healthcare tasks, increasing access to care, and augmenting healthcare capabilities. Finally, we discuss emerging concerns surrounding the use of individual physiology data and detail an increasingly important consideration for the field, namely the challenges of deploying AI models to achieve meaningful clinical impact.

The lack of annotated publicly available medical images is a major barrier for innovations. At the same time, many de-identified images and much knowledge are shared by clinicians on public forums such as medical Twitter. Here we harness these crowd platforms to curate OpenPath, a large dataset of 208,414 pathology images paired with natural language descriptions. This is the largest public dataset for pathology images annotated with natural text. We demonstrate the value of this resource by developing PLIP, a multimodal AI with both image and text understanding, which is trained on OpenPath. PLIP achieves state-of-the-art zero-shot and transfer learning performances for classifying new pathology images across diverse tasks. Moreover, PLIP enables users to retrieve similar cases by either image or natural language search, greatly facilitating knowledge sharing. Our approach demonstrates that publicly shared medical information is a tremendous resource that can be harnessed to advance biomedical AI.

Patients with cancer may be given treatments that are not officially approved (off-label) or recommended by guidelines (off-guideline) for multiple reasons including a lack of effective approved treatments. Here we present a systematic characterization of the patterns of off-label and off-guideline usage in 153,122 U.S. patients with 14 common cancer types using a large electronic health record (EHR)-derived de-identified database. We find that 18.3% and 3.9% of patients have received at least one line of off-label and off-guideline cancer drugs, respectively. Out of the 14 malignancies investigated, advanced bladder cancer has the highest proportion with 8.1% of patients receiving off-guideline treatments, most of which are recommended for non-small cell lung cancer. Patients with worse performance status, in later lines, or treated at academic hospitals are significantly more likely to receive off-label and off-guideline drugs. Underrepresented minority patients are less likely to receive off-guideline treatments in several cancer types. To quantify how predictable off-guideline usage is, we developed machine learning models to predict which drug a patient is likely to receive based on their clinical characteristics and previous treatments. Finally, we demonstrate that our systematic analysis of large real-world cohorts can identify interesting candidates for potential label expansion by identifying off-label treatments that demonstrate effectiveness in the real world setting. For example, we find that hormonal agents approved for breast cancer are used off-label in patients with ovarian cancer. Moreover, these hormonal agents show promising effectiveness in ovarian cancer with adjusted hazard ratio 0.53 (0.44, 0.65) compared to standard-of-care. This work demonstrates the power of large-scale computational analysis of real-world data for investigating non-standard cancer treatment usages.

The advance of science rests on a robust peer review process. However whether or not a paper is accepted can depend on random external factors--e.g. the timing of the submission, the matching of editors and reviewers--that are beyond the quality of the work. This article systematically investigates the impact of these random factors independent of the paper’s quality on peer review outcomes in a major biomedical journal,eLife. We analyzed all of the submissions toeLifebetween 2016 to 2018, with 23,190 total submissions. We examined the effects of random factors at each decision point in the review process, from the gatekeeping senior editors who may desk-reject papers to review editors and reviewers who recommend the final outcome. Our results suggest that the peer-review process ineLifeis robust overall and that random external factors have relatively little quantifiable bias.

A citation map connecting patents to biomedical publications provides insights that can be used to better evaluate productivity, diversity and translational impact.

Ablation is a common therapeutic procedure for atrial fibrillation (AF), that is guided to specific targets in the heart often after laborious segmentation of computed tomography. Machine learning (ML) can automate such tasks but requires large training datasets. Inspired by natural intelligence, which builds conceptual models to learn without large datasets, we mathematically encoded domain knowledge of atrial geometry to accelerate ML segmentation. In test cohorts (N=160) at 2 institutions, Dice scores were 96.7% (IQR: 95.3% to 97.7%) and 93.5% (IQR: 91.9% to 94.7%) with similar anatomic agreement to experts (r=0.99; p

Study Objectives Periodic limb movement in sleep is a common sleep phenotype characterized by repetitive leg movements that occur during or before sleep. We conducted a genome-wide association study (GWAS) of periodic limb movements in sleep (PLMS) using a joint analysis (i.e., discovery, replication, and joint meta-analysis) of four cohorts (MrOS, the Wisconsin Sleep Cohort Study, HypnoLaus, and MESA), comprised of 6843 total subjects. Methods The MrOS study and Wisconsin Sleep Cohort Study (N = 1745 cases) were used for discovery. Replication in the HypnoLaus and MESA cohorts (1002 cases) preceded joint meta-analysis. We also performed LD score regression, estimated heritability, and computed genetic correlations between potentially associated traits such as restless leg syndrome (RLS) and insomnia. The causality and direction of the relationships between PLMS and RLS was evaluated using Mendelian randomization. Results We found 2 independent loci were significantly associated with PLMS: rs113851554 (p = 3.51 × 10−12, β = 0.486), an SNP located in a putative regulatory element of intron eight of MEIS1 (2p14); and rs9369062 (p = 3.06 × 10−22, β = 0.2093), a SNP located in the intron region of BTBD9 (6p12); both of which were also lead signals in RLS GWAS. PLMS is genetically correlated with insomnia, risk of stroke, and RLS, but not with iron deficiency. Pleiotropy adjusted Mendelian randomization analysis identified a causal effect of RLS on PLMS. Conclusions Because PLMS is more common than RLS, PLMS may have multiple causes and additional studies are needed to further validate these findings.

Multiplex immunofluorescence (mIF) assays multiple protein biomarkers on a single tissue section. Recently, high-plex CODEX (co-detection by indexing) systems enable simultaneous imaging of 40+ protein biomarkers, unlocking more detailed molecular phenotyping, leading to richer insights into cellular interactions and disease. However, high-plex imaging can be slower and more costly to collect, limiting its applications, especially in clinical settings. We propose a machine learning framework, 7-UP, that can computationally generate in silico 40-plex CODEX at single-cell resolution from a standard 7-plex mIF panel by leveraging cellular morphology. We demonstrate the usefulness of the imputed biomarkers in accurately classifying cell types and predicting patient survival outcomes. Furthermore, 7-UP’s imputations generalize well across samples from different clinical sites and cancer types. 7-UP opens the possibility of in silico CODEX, making insights from high-plex mIF more widely available.

There is a growing focus on making clinical trials more inclusive but the design of trial eligibility criteria remains challenging1-3. Here we systematically evaluate the effect of different eligibility criteria on cancer trial populations and outcomes with real-world data using the computational framework of Trial Pathfinder. We apply Trial Pathfinder to emulate completed trials of advanced non-small-cell lung cancer using data from a nationwide database of electronic health records comprising 61,094 patients with advanced non-small-cell lung cancer. Our analyses reveal that many common criteria, including exclusions based on several laboratory values, had a minimal effect on the trial hazard ratios. When we used a data-driven approach to broaden restrictive criteria, the pool of eligible patients more than doubled on average and the hazard ratio of the overall survival decreased by an average of 0.05. This suggests that many patients who were not eligible under the original trial criteria could potentially benefit from the treatments. We further support our findings through analyses of other types of cancer and patient-safety data from diverse clinical trials. Our data-driven methodology for evaluating eligibility criteria can facilitate the design of more-inclusive trials while maintaining safeguards for patient safety.

Background: Olfactory impairment is evident in Alzheimer’s disease (AD); however, its precise relationships with clinical biomarker measures of tau pathology and neuroinflammation are not well understood. Objective: To determine if odor identification performance measured with the University of Pennsylvania Smell Identification Test (UPSIT) is related to in vivo measures of tau pathology and neuroinflammation. Methods: Cognitively normal and cognitively impaired participants were selected from an established research cohort of adults aged 50 and older who underwent neuropsychological testing, brain MRI, and amyloid PET. Fifty-four participants were administered the UPSIT. Forty-one underwent 18F-MK-6240 PET (measuring tau pathology) and fifty-three underwent 11C-PBR28 PET (measuring TSPO, present in activated microglia). Twenty-three participants had lumbar puncture to measure CSF concentrations of total tau (t-tau), phosphorylated tau (p-tau), and amyloid-β (Aβ42). Results: Low UPSIT performance was associated with greater18F-MK-6240 binding in medial temporal cortex, hippocampus, middle/inferior temporal gyri, inferior parietal cortex, and posterior cingulate cortex (p

ImportanceTelemedicine use accelerated during the COVID-19 pandemic, and skin conditions were a common use case. However, many images submitted may be of insufficient quality for making a clinical determination.ObjectiveTo determine whether an artificial intelligence (AI) decision support tool, a machine learning algorithm, could improve the quality of images submitted for telemedicine by providing real-time feedback and explanations to patients.Design, Setting, and ParticipantsThis quality improvement study with an AI performance component and single-arm clinical pilot study component was conducted from March 2020 to October 2021. After training, the AI decision support tool was tested on 357 retrospectively collected telemedicine images from Stanford telemedicine from March 2020 to June 2021. Subsequently, a single-arm clinical pilot study was conducted to assess feasibility with 98 patients in the Stanford Department of Dermatology across 2 clinical sites from July 2021 to October 2021. For the clinical pilot study, inclusion criteria for patients included being adults (aged ≥18 years), presenting to clinic for a skin condition, and being able to photograph their own skin with a smartphone.InterventionsDuring the clinical pilot study, patients were given a handheld smartphone device with a machine learning algorithm interface loaded and were asked to take images of any lesions of concern. Patients were able to review and retake photos prior to submitting, so each submitted photo met the patient’s assumed standard of clinical acceptability. A machine learning algorithm then gave the patient feedback on whether the image was acceptable. If the image was rejected, the patient was provided a reason by the AI decision support tool and allowed to retake the photos.Main Outcomes and MeasuresThe main outcome of the retrospective image analysis was the receiver operator curve area under the curve (ROC-AUC). The main outcome of the clinical pilot study was the image quality difference between the baseline images and the images approved by AI decision support.ResultsOf the 98 patients included, the mean (SD) age was 49.8 (17.6) years, and 50 (51%) of the patients were male. On retrospective telemedicine images, the machine learning algorithm effectively identified poor-quality images (ROC-AUC of 0.78) and the reason for poor quality (blurry ROC-AUC of 0.84; lighting issues ROC-AUC of 0.70). The performance was consistent across age and sex. In the clinical pilot study, patient use of the machine learning algorithm was associated with improved image quality. An AI algorithm was associated with reduction in the number of patients with a poor-quality image by 68.0%.Conclusions and RelevanceIn this quality improvement study, patients use of the AI decision support with a machine learning algorithm was associated with improved quality of skin disease photographs submitted for telemedicine use.

Dimensionality reduction (DR) is commonly used to project highdimensional data into lower dimensions for visualization, which could then generate new insights and hypotheses. However, DR algorithms necessarily introduce distortions in the visualization and cannot faithfully represent all relations in the data. As such, there is a need for methods to assess the reliability of DR visualizations. Here we present DynamicViz, a framework for generating dynamic visualizations that capture the sensitivity of DR visualizations to perturbations in the data. DynamicVic can be applied to all commonly used DR methods. We show the utility of dynamic visualizations in diagnosing common interpretative pitfalls of static visualizations and extending existing single-cell analyses. We introduce the variance score to quantify the dynamic variability of observations in these visualizations. The variance score characterizes natural variability in the data and can be used to optimize DR algorithm implementations. We have made DynamicViz freely available to assist with the evaluation of DR visualizations.

Multiplexed immunofluorescence imaging enables high-dimensional molecular profiling at subcellular resolution. However, learning disease-relevant cellular environments from these rich imaging data is an open challenge. We developed SPAtial CEllular Graphical Modeling (SPACE-GM), a geometric deep learning framework that flexibly models tumor microenvironments (TMEs) as cellular graphs. We applied SPACE-GM to 658 head-and-neck and colorectal human cancer samples assayed with 40-plex immunofluorescence imaging to identify spatial motifs associated with cancer recurrence and patient survival after immunotherapy. SPACE-GM is substantially more accurate in predicting patient outcomes than previous approaches for modeling spatial data using neighborhood cell-type compositions. Computational interpretation of the disease-relevant microenvironments identified by SPACE-GM generates insights into the effect of spatial dispersion of tumor cells and granulocytes on patient prognosis.

Clinical trials are the gatekeepers and bottlenecks of progress in medicine. In recent years, they have become increasingly complex and expensive, driven by a growing number of stakeholders requiring more endpoints, more diverse patient populations, and a stringent regulatory environment. Trial designers have historically relied on investigator expertise and legacy norms established within sponsor companies to improve operational efficiency while achieving study goals. As such, data-driven forecasts of operational metrics can be a useful resource for trial design and planning. We develop a machine learning model to predict clinical trial operational efficiency using a novel dataset from Roche containing over 2,000 clinical trials across 20 years and multiple disease areas. The data includes important operational metrics related to patient recruitment and trial duration, as well as a variety of trial features such as the number of procedures, eligibility criteria, and endpoints. Our results demonstrate that operational efficiency can be predicted robustly using trial features, which can provide useful insights to trial designers on the potential impact of their decisions on patient recruitment success and trial duration.

Current methods for spatial transcriptomics are limited by low spatial resolution. Here we introduce a method that integrates spatial gene expression data with histological image data from the same tissue section to infer higher-resolution expression maps. Using a deep generative model, our method characterizes the transcriptome of micrometer-scale anatomical features and can predict spatial gene expression from histology images alone. ispartof: NATURE BIOTECHNOLOGY vol:40 issue:4 pages:476-+ ispartof: location:United States status: published

Large language models, which are increasingly used in AI applications, display undesirable stereotypes such as persistent associations between Muslims and violence. New approaches are needed to systematically reduce the harmful bias of language models in deployment.

Background. Transparent and accessible reporting of COVID-19 data is critical for public health efforts. Each state and union territory (UT) of India has its own mechanism for reporting COVID-19 data, and the quality of their reporting has not been systematically evaluated. We present a comprehensive assessment of the quality of COVID-19 data reporting done by the Indian state and union territory governments. This assessment informs the public health efforts in India and serves as a guideline for pandemic data reporting by other governments. Methods. We designed a semi-quantitative framework to assess the quality of COVID-19 data reporting done by the states and union territories of India. This framework captures four key aspects of public health data reporting - availability, accessibility, granularity, and privacy. We then used this framework to calculate a COVID-19 Data Reporting Score (CDRS, ranging from 0 to 1) for 29 states based on the quality of COVID-19 data reporting done by the state during the two-week period from 19 May to 1 June, 2020. States that reported less than 10 total confirmed cases as of May 18 were excluded from the study. Findings. Our results indicate a strong disparity in the quality of COVID-19 data reporting done by the state governments in India. CDRS varies from 0.61 (good) in Karnataka to 0.0 (poor) in Bihar and Uttar Pradesh, with a median value of 0.26. Only ten states provide a visual representation of the trend in COVID-19 data. Ten states do not report any data stratified by age, gender, comorbidities or districts. In addition, we identify that Punjab and Chandigarh compromised the privacy of individuals under quarantine by releasing their personally identifiable information on the official websites. Across the states, the CDRS is positively associated with the state's sustainable development index for good health and well-being (Pearson correlation: r=0.630, p=0.0003). Interpretation. The disparity in CDRS across states highlights three important findings at the national, state, and individual level. At the national level, it shows the lack of a unified framework for reporting COVID-19 data in India, and highlights the need for a central agency to monitor or audit the quality of data reporting done by the states. Without a unified framework, it is difficult to aggregate the data from different states, gain insights from them, and coordinate an effective nationwide response to the pandemic. Moreover, it reflects the inadequacy in coordination or sharing of resources among the states in India. Coordination among states is particularly important as more people start moving across states in the coming months. The disparate reporting score also reflects inequality in individual access to public health information and privacy protection based on the state of residence. Funding. J.Z. is supported by NSF CCF 1763191, NIH R21 MD012867-01, NIH P30AG059307, NIH U01MH098953 and grants from the Silicon Valley Foundation and the Chan-Zuckerberg Initiative.

We sought to determine if upstream amyloid accumulation and downstream cognitive impairment have independent relationships with microglial activation and tau pathology. Fifty-eight older adults were stratified by amyloid and cognitive status based on (18)F-florbetaben PET, history, and neuropsychological testing. Of these, 57 had (11)C-PBR28 PET to measure microglial activation and 43 had (18)F-MK-6240 PET to measure tau pathology. Amyloid and cognitive status were associated with increased overall binding for both (11)C-PBR28 and (18)F-MK-6240 (p’s < 0.01). While there was no interaction between amyloid and cognitive status in their association with (11)C-PBR28 binding (p = 0.6722), there was an interaction in their association with (18)F-MK-6240 binding (p = 0.0115). Binding of both radioligands was greater in amyloid-positive controls than in amyloid-negative controls; however, this difference was seen in neocortical regions for (11)C-PBR28 and only in medial temporal cortex for (18)F-MK-6240. We conclude that, in the absence of cognitive symptoms, amyloid deposition has a greater association with microglial activation than with tau pathology.

The cell’s shape and motion represent fundamental aspects of the cell identity, and can be highly predictive of the function and pathology. However, automated analysis of the morphodynamic states remains challenging for most cell types, especially primary human cells where genetic labeling may not be feasible. To enable automated and quantitative analysis of morphodynamic states, we developed DynaMorph – a computational framework that combines quantitative live cell imaging with self-supervised learning. To demonstrate the fidelity and robustness of this approach, we used DynaMorph to annotate morphodynamic states observed with label-free measurements of density and anisotropy of live microglia isolated from human brain tissue. These cells show complex behavior and have varied responses to disease-relevant stimuli. DynaMorph generates quantitative morphodynamic representations that can be used to evaluate the effects of disease-relevant perturbations. Using DynaMorph, we identify distinct morphodynamic states of microglia polarization and detect rare transition events between states. The methodologies presented here can facilitate automated discovery of functional states of diverse cellular systems.

Quantifying the effectiveness of different cancer therapies in patients with specific tumor mutations is critical for improving patient outcomes and advancing precision medicine. Here we perform a large-scale computational analysis of 40,903 US patients with cancer who have detailed mutation profiles, treatment sequences and outcomes derived from electronic health records. We systematically identify 458 mutations that predict the survival of patients on specific immunotherapies, chemotherapy agents or targeted therapies across eight common cancer types. We further characterize mutation-mutation interactions that impact the outcomes of targeted therapies. This work demonstrates how computational analysis of large real-world data generates insights, hypotheses and resources to enable precision oncology.

Eye tracking, or oculography, provides insight into where a person is looking. Recent advances in camera technology and machine learning have enabled prevalent devices like smart-phones to track gaze and visuo-motor behavior at near clinical-quality resolution. A critical gap in using oculography to diagnose visuo-motor dysfunction on a large scale is in the design of visual task paradigms, algorithms for diagnosis, and sufficiently large datasets. In this study, we used a 500 Hz infrared oculography dataset in healthy controls and patients with various neurological diseases causing visuo-motor abnormality due to eye movement disorder or vision loss. We used novel visuo-motor tasks involving rapid reading of 40 single-digit numbers per page and developed a machine learning algorithm for predicting disease state. We show that oculography data acquired while a person reads one page of 40 single-digit numbers (15-30 seconds duration) is predictive of of visuo-motor dysfunction (ROC-AUC = 0:973). Remarkably, we also find that short recordings of about 2.5 seconds (6-12× reduction in time) are sufficient for disease detection (ROC-AUC = 0:831). We identify which tasks are most informative for identifying visuo-motor dysfunction (those with the most visual crowding), and more specifically, which aspects of the task are most predictive (the recording segments where gaze moves vertically across lines). In addition to segregating disease and controls, our novel visuo-motor paradigms can discriminate among diseases impacting eye movement, diseases associated with vision loss, and healthy controls (81% accuracy compared with baseline of 33%).

Single-cell RNA sequencing (scRNA-seq) has the potential to provide powerful, high-resolution signatures to inform disease prognosis and precision medicine. This paper takes an important first step towards this goal by developing an interpretable machine learning algorithm, CloudPred, to predict individuals' disease phenotypes from their scRNA-seq data. Predicting phenotype from scRNA-seq is challenging for standard machine learning methods -- the number of cells measured can vary by orders of magnitude across individuals and the cell populations are also highly heterogeneous. Typical analysis creates pseudo-bulk samples which are biased toward prior annotations and also lose the single cell resolution. CloudPred addresses these challenges via a novel end-to-end differentiable learning algorithm which is coupled with a biologically informed mixture of cell types model. CloudPred automatically infers the cell subpopulation that are salient for the phenotype without prior annotations. We developed a systematic simulation platform to evaluate the performance of CloudPred and several alternative methods we propose, and find that CloudPred outperforms the alternative methods across several settings. We further validated CloudPred on a real scRNA-seq dataset of 142 lupus patients and controls. CloudPred achieves AUROC of 0.98 while identifying a specific subpopulation of CD4 T cells whose presence is highly indicative of lupus. CloudPred is a powerful new framework to predict clinical phenotypes from scRNA-seq data and to identify relevant cells., Preprint of an article published in Pacific Symposium on Biocomputing \copyright\ 2021 World Scientific Publishing Co., Singapore, http://psb.stanford.edu/

The T-cell receptor (TCR) allows T-cells to recognize and respond to antigens presented by infected and diseased cells. However, due to TCRs’ staggering diversity and the complex binding dynamics underlying TCR antigen recognition, it is challenging to predict which antigens a given TCR may bind to. Here, we present TCR-BERT, a deep learning model that applies self-supervised transfer learning to this problem. TCR-BERT leverages unlabeled TCR sequences to learn a general, versatile representation of TCR sequences, enabling numerous downstream applications. We demonstrate that TCR-BERT can be used to build state-of-the-art TCR-antigen binding predictors with improved generalizability compared to prior methods. TCR-BERT simultaneously facilitates clustering sequences likely to share antigen specificities. It also facilitates computational approaches to challenging, unsolved problems such as designing novel TCR sequences with engineered binding affinities. Importantly, TCR-BERT enables all these advances by focusing on residues with known biological significance. TCR-BERT can be a useful tool for T-cell scientists, enabling greater understanding and more diverse applications, and provides a conceptual framework for leveraging unlabeled data to improve machine learning on biological sequences.

Introduction: Echocardiographic assessment of left ventricular (LV) systolic function by ejection fraction (EF) is crucial for diagnosis and monitoring in pediatric patients. EF requires manual tracing of the LV but measurement reproducibility is affected by diverse phenotypes, image quality, and operator skill. We sought to train a deep learning model (EchoNet-Peds) on a pediatric-specific video dataset to automatically segment the LV and estimate EF by 5/6 Area Length (AL) using two common views, the apical 4 chamber (A4C) and parasternal short axis (PSAX). Methods: EchoNet-Peds evaluated raw A4C and PSAX videos to segment the LV and predict the EF from the A4C and PSAX views. Eighty percent of the dataset was used for algorithm training, 10% for validation, and 10% for testing. We used the training set to predict the EF by AL, determined hyperparameters from the validation set, and reported results using 10-fold cross-validation on the test set. Results: The dataset included 1923 children (43% female) with 4400 echocardiogram studies. EFs ranged from 10% to 81% (61% ± 10.1) and ages from 0 to 18 years (10 ± 5.4 years). EchoNet-Peds achieved Dice Similarity Coefficients of 0.901 (0.895 - 0.907) and 0.887 (0.879 - 0.895) for segmenting the LV in A4C and PSAX views respectively. The A4C view predicts an EF by AL with an R2 of 0.73 (0.67 - 0.77) and the PSAX view predicts an R2 of 0.74 (0.70 - 0.77). Combining both predictions results in an overall R2 of 0.78 (0.74 - 0.81). The model performs well when stratified by sex, age, and body surface area (BSA), and improves in younger patients and lower BSA (Figure 1 b-d). Conclusions: EchoNet-Peds is the first demonstration that AI trained on a pediatric-specific dataset utilizing multiple views can accurately segment the LV and predict EF. The automated functionality of EchoNet-Peds can be used for real-time assessment of cardiac function despite pediatric heterogeneity.

Gene expression at the individual cell-level resolution, as quantified by single-cell RNA-sequencing (scRNA-seq), can provide unique insights into the pathology and cellular origin of diseases and complex traits. Here, we introduce single-cell Disease Relevance Score (scDRS), an approach that links scRNA-seq with polygenic risk of disease at individual cell resolution; scDRS identifies individual cells that show excess expression levels for genes in a disease-specific gene set constructed from GWAS data. We determined via simulations that scDRS is well-calibrated and powerful in identifying individual cells associated to disease. We applied scDRS to GWAS data from 74 diseases and complex traits (average N=341K) in conjunction with 16 scRNA-seq data sets spanning 1.3 million cells from 31 tissues and organs. At the cell type level, scDRS broadly recapitulated known links between classical cell types and disease, and also produced novel biologically plausible findings. At the individual cell level, scDRS identified subpopulations of disease-associated cells that are not captured by existing cell type labels, including subpopulations of CD4+ T cells associated with inflammatory bowel disease, partially characterized by their effector-like states; subpopulations of hippocampal CA1 pyramidal neurons associated with schizophrenia, partially characterized by their spatial location at the proximal part of the hippocampal CA1 region; and subpopulations of hepatocytes associated with triglyceride levels, partially characterized by their higher ploidy levels. At the gene level, we determined that genes whose expression across individual cells was correlated with the scDRS score (thus reflecting co-expression with GWAS disease genes) were strongly enriched for gold-standard drug target and Mendelian disease genes.

We develop a deep learning approach

Machine learning models have great potential in biomedical applications. A new platform called GradioHub offers an interactive and intuitive way for clinicians and biomedical researchers to try out models and test their reliability on real-world, out-of-training data.

The goal of sex and gender analysis is to promote rigorous, reproducible and responsible science. Incorporating sex and gender analysis into experimental design has enabled advancements across many disciplines, such as improved treatment of heart disease and insights into the societal impact of algorithmic bias. Here we discuss the potential for sex and gender analysis to foster scientific discovery, improve experimental efficiency and enable social equality. We provide a roadmap for sex and gender analysis across scientific disciplines and call on researchers, funding agencies, peer-reviewed journals and universities to coordinate efforts to implement robust methods of sex and gender analysis. The authors discuss the potential for sex and gender analysis to foster scientific discovery, improve experimental efficiency and enable social equality.

In order for machine learning to be deployed and trusted in many applications, it is crucial to be able to reliably explain why the machine learning algorithm makes certain predictions. For example, if an algorithm classifies a given pathology image to be a malignant tumor, then the doctor may need to know which parts of the image led the algorithm to this classification. How to interpret black-box predictors is thus an important and active area of research. A fundamental question is: how much can we trust the interpretation itself? In this paper, we show that interpretation of deep learning predictions is extremely fragile in the following sense: two perceptively indistinguishable inputs with the same predicted label can be assigned very different interpretations. We systematically characterize the fragility of several widely-used feature-importance interpretation methods (saliency maps, relevance propagation, and DeepLIFT) on ImageNet and CIFAR-10. Our experiments show that even small random perturbation can change the feature importance and new systematic perturbations can lead to dramatically different interpretations without changing the label. We extend these results to show that interpretations based on exemplars (e.g. influence functions) are similarly fragile. Our analysis of the geometry of the Hessian matrix gives insight on why fragility could be a fundamental challenge to the current interpretation approaches., Comment: Published as a conference paper at AAAI 2019

Generative adversarial networks (GANs) represent an attractive and novel approach to generate realistic data, such as genes, proteins or drugs, in synthetic biology. Here, we apply GANs to generate synthetic DNA sequences encoding for proteins of variable length. We propose a novel feedback-loop architecture, feedback GAN (FBGAN), to optimize the synthetic gene sequences for desired properties using an external function analyser. The proposed architecture also has the advantage that the analyser does not need to be differentiable. We apply the feedback-loop mechanism to two examples: generating synthetic genes coding for antimicrobial peptides, and optimizing synthetic genes for the secondary structure of their resulting peptides. A suite of metrics, calculated in silico, demonstrates that the GAN-generated proteins have desirable biophysical properties. The FBGAN architecture can also be used to optimize GAN-generated data points for useful properties in domains beyond genomics. Generative machine learning models are used in synthetic biology to find new structures such as DNA sequences, proteins and other macromolecules with applications in drug discovery, environmental treatment and manufacturing. Gupta and Zou propose and demonstrate in silico a feedback-loop architecture to optimize the output of a generative adversarial network that generates synthetic genes to produce ones specifically coding for antimicrobial peptides.

Recent development in inferring RNA velocity from single-cell RNA-seq opens up exciting new vista into developmental lineage and cellular dynamics. However, the estimated velocity only gives a snapshot of how the transcriptome instantaneously changes in individual cells, and it does not provide quantitative predictions and insights about the whole system. In this work, we develop RNA-ODE, a principled computational framework that extends RNA velocity to quantify systems level dynamics and improve single-cell data analysis. We model the gene expression dynamics by an ordinary differential equation (ODE) based formalism. Given a snapshot of gene expression at one time, RNA-ODE is able to predict and extrapolate the expression trajectory of each cell by solving the dynamic equations. Systematic experiments on simulations and on new data from developing brain demonstrate that RNA-ODE substantially improves many aspects of standard single-cell analysis. By leveraging temporal dynamics, RNA-ODE more accurately estimates cell state lineage and pseudo-time compared to previous state-of-the-art methods. It also infers gene regulatory networks and identifies influential genes whose expression changes can decide cell fate. We expect RNA-ODE to be a Swiss army knife that aids many facets of single-cell RNA-seq analysis.

Spontaneous tumors in dogs are highly similar to human cancers histologically, genetically, molecularly and clinically. In this presentation we explore similarities found across species when associating cancer biomarkers and outcomes. Our findings validate concordance across the cancer journeys. This is important to deepen the field of comparative oncology and affirming spontaneous cancer in dogs as an ideal ‘model’ for precision oncology, building upon recent work in tumor biology concordance. Our machine learning analysis prognosis and predictive effect of genomic alterations in 1303 client-owned tumor-bearing dogs identified TP53 and PTPRD as prognostic markers for poor survival in dogs with an overall survival hazard ratio (OS HR) of 1.732, 1.914, respectively (P Citation Format: Kevin Wu, Lucas Rodrigues, Gerald Post, Garrett Harvey, Aubrey Miller, Lindsay Lambert, Christina Lopes, Benjamin Lewis, James Zou. Concordance between dogs and humans: The use of AI in evaluating clinical cancer genomic datasets [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 635.