Research Publications

Selected collection of research spanning 30+ years of innovation in biomedical informatics, artificial intelligence, and clinical medicine.

Genetics, Genomics, Rare Diseases

Computational approaches to understanding genetic variation and rare disease mechanisms

Commonalities across computational workflows for uncovering explanatory variants in undiagnosed cases

Genet Med (2021) – Genomic sequencing is a powerful tool for discovering genetic aberrations underlying rare Mendelian conditions.

Surveyed pipelines across 12 academic centers, finding consensus in variant calling and QC but divergence in prioritization and data integration.

Mutual information relevance networks: functional genomic clustering using pairwise entropy measurements

Pac Symp Biocomput (2000) – Increasing numbers of methodologies are available to find functional genomic clusters in RNA expression data.

Pioneering pairwise mutual information thresholding to form biologically meaningful relevance networks.

Linking gene expression data with patient survival times using partial least squares

Bioinformatics (2002) – Linking the large amount of genotypic data, gathered using microarrays for example, with various phenotypic data from patients.

Demonstrates that partial least squares can model censored survival data to identify gene expression signatures correlated with patient outcomes.

Genetic Diagnosis on Patients with Previously Undiagnosed Disease

N Engl J Med (2018) – Many patients remain without a diagnosis despite extensive medical evaluation.

The Undiagnosed Diseases Network applied exome/genome sequencing to 382 patients and achieved a 35% diagnostic yield (132/382), with genomic diagnoses prompting changes in therapy (21%), diagnostic testing (37%), and genetic counseling (36%), and defining 31 new syndromes.

Polygenic risk scores for autoimmune related diseases are significantly different in cancer exceptional responders

NPJ Precis Oncol (2024) – A small number of cancer patients respond exceptionally well to therapies and survive significantly longer than patients with similar diagnoses.

Chen et al. compared PRSs across autoimmune diseases in exceptional responders versus typical cancer patients and found significantly elevated scores for type 1 diabetes, hypothyroidism, and psoriasis.

Simulation of undiagnosed patients with novel genetic conditions

Nat Commun (2023) – Rare Mendelian disorders pose a major diagnostic challenge and collectively affect 300–400 million patients worldwide.

A pipeline simulates realistic patients with novel genetic conditions and shows that common gene prioritization methods underperform on these cases.

Comprehensive Analysis of Tissue-wide Gene Expression and Phenotype Data Reveals Tissues Affected in Rare Genetic Disorders

Cell Syst. (2017) – Integrating tissue-wide gene expression and phenotype data to map tissues impacted by rare disorders.

Feiglin et al. combined multi‑tissue expression profiles with phenotype associations evealing distinct patterns of tissue tropism in rare genetic diseases.

Genetic Misdiagnoses and the Potential for Health Disparities

N Engl J Med (2016) – Misclassification of benign variants as pathogenic highlights need for diverse reference populations.

Manrai et al. analyzed public exome data and clinical testing records, finding that variants once deemed pathogenic were recategorized as benign—predominantly in individuals of African ancestry—underscoring the necessity of sequencing diverse populations and using ancestry‑matched controls for variant interpretation.

Population-Based Penetrance of Deleterious Clinical Variants

JAMA (2022) – Population-based assessment of disease risk associated with gene variants informs clinical decisions.

Forrest et al. evaluated 37,780 pathogenic or loss-of-function variants in 72,434 individuals from the BioMe and UK Biobanks, demonstrating that penetrance is generally low and variable by gene, age, and ancestry—highlighting the importance of population-based penetrance estimates for accurate risk stratification.

An International Effort towards Standards for Best Practices in Clinical Genome Sequencing (CLARITY Challenge)

Genome Biol. (2014) – The CLARITY Challenge assesses current practices in clinical WGS interpretation and reporting.

Brownstein et al. convened multiple teams to analyze standardized genome sequencing cases, revealing convergence in bioinformatic pipelines but notable variability in medical interpretation and clinical reporting, thereby identifying areas requiring further standardization.

Creation and Implications of a Phenome‑Genome Network

Nat Biotechnol. (2006) – Integrating phenotypic, environmental, and gene expression data to identify gene–phenome relationships.

Butte & Kohane built a network linking UMLS‑annotated phenotypic and environmental concepts from GEO with genes showing differential expression, clustering data sets by phenotype and uncovering novel gene associations such as aging regulators—paving the way toward a Human Phenome Project.

A Gene Expression Profile of Stem Cell Pluripotentiality and Differentiation Is Conserved across Diverse Solid and Hematopoietic Cancers

Genome Biol. (2012) – A 189‑gene “stemness” signature stratifies tissues and cancers by pluripotentiality.

Palmer et al. defined a 189‑gene stem‑cell gene set (SCGS) using an unbiased filter on Affymetrix data, showing that SCGS activity orders human and murine samples by plasticity and correlates with tumor grade in multiple cancers—offering a quantitative measure of cancer stem‑like transcriptional activity.

Analysis of Gene Expression in a Developmental Context Emphasizes Distinct Biological Leitmotifs in Human Cancers

Genome Biol. (2008) – Projecting tumors onto developmental timelines reveals three cancer classes with unique transcriptional patterns.

Naxerova et al. mapped gene expression from 32 tumor types onto developmental timelines derived from ten embryonic processes, identifying three classes of cancers with distinct developmental signatures.

Gene regulation and DNA damage in the ageing human brain

Nature (2004) – The ageing of the human brain is a cause of cognitive decline in the elderly and the major risk factor for Alzheimer's disease.

Lu et al. used transcriptional profiling of human frontal cortex samples from individuals aged 26–106 years to identify genes whose expression declines after age 40. They demonstrate that oxidative DNA damage selectively accumulates in promoters of age‑downregulated genes.

Using electronic health records to drive discovery in disease genomics

Nat Rev Genet (2011) – If genomic studies are to be a clinically relevant and timely reflection of the relationship between genetics and health status …

Kohane proposes leveraging the codified and narrative data in EHRs to cost‑effectively accelerate genomic research at population scale, reproducing and extending GWAS findings across diverse populations, while highlighting regulatory and consent challenges to broad EDGR adoption.

Conserved mechanisms across development and tumorigenesis revealed by a mouse development perspective of human cancers

Genes Dev (2004) – Identification of common mechanisms underlying organ development and primary tumor formation should yield new insights into tumor biology …

Kohane et al. projected human medulloblastoma gene‑expression profiles onto a mouse cerebellar development timeline (P1–P60), finding that tumors most closely mirror early postnatal stages, with metastatic MBs aligning to a narrow developmental window, thereby demonstrating that developmental context can illuminate tumor biology and model generation.

Mutual information relevance networks: functional genomic clustering using pairwise entropy measurements

Pac Symp Biocomput (2000) – Increasing numbers of methodologies are available to find functional genomic clusters in RNA expression data.

Butte & Kohane introduce a method computing pairwise mutual information across all genes in an expression dataset, applying a threshold to construct “Relevance Networks” of 22 gene clusters from 2,467 genes in 79 samples, and demonstrate the biological significance of each network for functional genomics analysis.

Artificial Intelligence/Machine Learning

Pioneering AI applications in healthcare and biomedical research

Few shot learning for phenotype-driven diagnosis of patients with rare genetic diseases

NPJ Digit Med (2025) – Research Article

SHEPHERD, a knowledge-graph–grounded few‑shot learning framework that accurately diagnoses patients with rare genetic diseases across multiple cohorts (Undiagnosed Diseases Network, MyGene2, and Deciphering Developmental Disorders).

Artificial Intelligence in Medicine

NEJM (2023) - Editorial launching NEJM AI

Broad view of the future of medical AI research and announcing the launch of NEJM AI journal.

Medical Artificial Intelligence and Human Values

NEJM AI (2023) - Ethics and values in medical AI

Framework for incorporating human values and ethical considerations into AI clinical decision-support systems.

Systematic Characterization of the Effectiveness of Alignment in Large Language Models for Categorical Decisions

arXiv preprint (2024) – As large language models (LLMs) are deployed in high‑stakes domains like healthcare, understanding how well their decision‑making aligns with human preferences and values becomes crucial.

Introduces the Alignment Compliance Index and demonstrates variable alignment effectiveness across three LLMs in medical triage.

The potential of Generative Pre-trained Transformer 4 (GPT-4) to analyse medical notes in three different languages: a retrospective model-evaluation study

Lancet Digit Health (2025) – Retrospective model-evaluation study

Across eight university hospitals in the USA, Colombia, Singapore, and Italy, GPT-4 demonstrated robust multilingual clinical text understanding.

Artificial intelligence in healthcare

Nature Biomedical Engineering (2018) - Comprehensive review

Seminal review of AI applications and implications for healthcare, covering technical advances and implementation challenges.

A unified inference procedure for a class of measures to assess improvement in risk prediction systems with survival data

Statistics in Medicine (2013) – Risk prediction procedures aid treatment selection and disease management.

Presents a unified inferential framework for measures like NRI and IDI with censored survival data.

Machine Learning in Medicine

N Engl J Med (2019) – Machine learning is not just a new tool, such as a new drug or medical device.

Rajkomar, Dean & Kohane review how ML algorithms can process vast healthcare data to support prognosis, diagnosis, treatment, and clinician workflows, discuss integration challenges including data quality and clinical workflow fit, and envision a future where ML meaningfully augments medical practice.

Heterogeneity of continuous glucose monitoring features and their clinical associations in a type 2 diabetes population

Diabetes Obes Metab (2025) – Cohort study

Analysed CGM and electronic health record data from 6,533 individuals with type 2 diabetes. Clustering revealed four distinct feature patterns with heterogeneous associations to clinical covariates, underscoring the potential of CGM‑derived metrics to inform precision diabetes management.

Adversarial attacks on medical machine learning

Science (2019) – With public and academic attention increasingly focused on the new role of machine learning in the health information economy, an unusual and no-longer-esoteric category of vulnerabilities in machine-learning systems could prove important.

Finlayson et al. outline how small, carefully designed perturbations (“adversarial examples”) can subvert state‑of‑the‑art medical deep‑learning classifiers across multiple clinical domains, warn of healthcare‑specific incentives for such attacks, and call for research into defenses to safeguard clinical deployments.

Framing the challenges of artificial intelligence in medicine

BMJ Qual Saf (2019) – Use of artificial intelligence (AI) and computer algorithms as tools to improve diagnosis have both risks and benefits.

Yu & Kohane discuss key hurdles for safe AI integration in clinical settings—data quality, algorithm reliability, workflow compatibility, and patient trust—and emphasize that addressing these challenges is critical for realizing AI’s potential in medicine.

Big Data and Machine Learning in Health Care

JAMA (2018) – Big data and machine learning can create algorithms that perform on par with human physicians.

Beam & Kohane highlight how large-scale healthcare datasets and ML techniques can scale in performance and data set size.

Longitudinal histories as predictors of future diagnoses of domestic abuse

BMJ (2009) – Commonly available longitudinal diagnostic data can be useful for predicting a patient’s future risk of receiving a diagnosis of abuse.

Developed Bayesian models using routine diagnostic codes to predict domestic abuse diagnoses 10–30 months in advance highlighting the potential for early identification and intervention .

Biases in electronic health record data due to processes within the healthcare system: retrospective observational study

BMJ (2018) – The presence and timing of laboratory test orders have significant predictive value for patient outcomes independent of test results.

Analyzing 669 452 patients across two Boston hospitals, the authors show that ordering patterns (time of day, day of week, frequency) predict three‑year survival more accurately than actual test results in 68% of tests, underscoring the need to model healthcare processes in EHR research.

Bayesian approach to discovering pathogenic SNPs in conserved protein domains

Human Mutation (2004) – The probability of a missense mutation causing a significant phenotype change depends on phylogenetic, biochemical, and structural features.

Bayesian algorithm that integrates evolutionary and biochemical features to predict pathogenic nsSNPs in conserved domains, achieving 90% specificity when tested on OMIM and dbSNP variants.

Fuzzy logic controller for weaning neonates from mechanical ventilation

Proc AMIA Symp (2002) – Weaning is gradual detachment from ventilatory support until spontaneous breathing resumes.

Developed a fuzzy logic controller using heart rate, respiratory rate, tidal volume, and oxygen saturation trends to adjust SIMV settings for newborns.

Temporal reasoning in medical expert systems

MIT LCS Technical Report (1987) – Diseases develop and change over time.

Methods for temporal abstraction, constraint propagation, and diagnostic evaluation, established frameworks for handling time‑dependent medical data in clinical decision support and influencing subsequent research in biomedical temporal reasoning.

Clinical Informatics

Electronic health records, clinical decision support, and healthcare information systems

Predicting seizure recurrence after an initial seizure-like episode from routine clinical notes using large language models: a retrospective cohort study

The Lancet Digital Health (2023) – The evaluation and management of first-time seizure-like events in children can be difficult. Can LLM help?

LLMs outperformed structured-data models in evaluating first time seizures.

SMART on FHIR: a standards‑based, interoperable apps platform for electronic health records

J Am Med Inform Assoc (2016) – A platform that enables medical applications to be written once and run unmodified across different healthcare IT systems.

Mandel et al. describe the creation and industry prototyping of SMART on FHIR, demonstrating feasibility across multiple EHR vendors.

Postsurgical prescriptions for opioid naive patients and association with overdose and misuse: retrospective cohort study

BMJ (2018) – Quantify the effects of varying opioid prescribing patterns after surgery on dependence, overdose, or abuse.

Brat et al. show that each refill is associated with a 44.0% increase in misuse and each additional week of use with a 19.9% increase in hazard.

Enabling phenotypic big data with PheNorm

J Am Med Inform Assoc (2018) – Electronic health record (EHR)‑based phenotyping infers whether a patient has a disease based on EHR information without a large gold‑standard training set.

Yu et al. introduce PheNorm, which uses anchor features and a mixture model with denoising self‑regression to generate phenotype predictions from unlabeled EHR data.

Development of an Algorithm to Identify Patients with Physician‑Documented Insomnia

Sci Rep (2018) – We developed an insomnia classification algorithm by interrogating an EMR database of 314,292 patients.

Kartoun et al. combine structured ICD‑9 codes with unstructured clinical note mentions, achieving an AUROC of 0.83 versus 0.55 for billing codes alone, and identify 36,810 insomnia patients—fewer than 17% of whom had insomnia billing codes.

Finding the missing link for big biomedical data

JAMA (2014) – Linking big data will enable physicians and researchers to test new hypotheses and identify areas of possible intervention.

Explore the technical and social challenges of integrating de‑identified datasets across institutions, advocating for federated data linkage frameworks.

Medicine's uncomfortable relationship with math: calculating positive predictive value

JAMA Intern Med (2014) – A survey of internists revealed dramatic overestimation of test PPV compared to the true probability.

Manrai et al. highlight pervasive challenges in applying statistical reasoning to clinical decision‑making and calling for enhanced quantitative training in medicine.

Autism

Computational approaches to understanding autism spectrum disorders

Comorbidity clusters in autism spectrum disorders: an electronic health record time-series analysis

Pediatrics (2014) – The distinct trajectories of patients with autism spectrum disorders (ASDs) have not been extensively studied, particularly regarding clinical manifestations beyond the neurobehavioral criteria in DSM.

Hierarchical clustering of longitudinal ICD‑9 codes in ASD patients revealed four subgroups (seizure, multisystem, psychiatric, unspecific), indicating heterogeneous comorbidity patterns with etiologic and therapeutic implications.

Characteristics and predictive value of blood transcriptome signature in males with autism spectrum disorders

PLOS One (2012) – Autism Spectrum Disorders (ASD) is a spectrum of highly heritable neurodevelopmental disorders in which known mutations contribute to disease risk in 20% of cases.

Developed a 55‑gene blood expression signature that classified male ASD cases with AUC 0.70 and early evidence of immunological dysregulation in a subgroup

Gene expression analysis in Fmr1KO mice identifies an immunological signature in brain tissue and mGluR5-related signaling in primary neuronal cultures

Mol Autism (2015) – Fragile X syndrome (FXS) is a neurodevelopmental disorder whose biochemical manifestations involve dysregulation of mGluR5-dependent pathways in neuronal models. :contentReference[oaicite:0]{index=0}

Finds an immunological pathway signature in Fmr1KO embryonic brain tissue contrasting with synaptic signatures in primary neuronal cultures.

Association of Sex With Recurrence of Autism Spectrum Disorder Among Siblings

JAMA Pediatr (2017) – Autism spectrum disorder (ASD) is known to be more prevalent among males than females in the general population.

In 3.17 million children from 1.58 million families, recurrence rates were 12.9% in male and 4.2% in female siblings if the older sibling was male, and 16.7% and 7.6% respectively if the older sibling was female. :contentReference[oaicite:0]{index=0}

Integrative analysis of genetic data sets reveals a shared innate immune component in autism spectrum disorder and its co-morbidities

Genome Biol. (2016) – ASD and most co-morbidities share dysregulated innate immune pathways.

Three-tier transcriptomic meta-analysis identifies Toll‑like receptor and chemokine signaling as common to ASD and its co‑morbid diseases.

Finding a new balance between a genetics-first or phenotype-first approach to the study of disease

Neuron (2021) – Successes in neuroscience using a genetics-first approach to characterizing disorders such as autism have eclipsed the value of a comprehensive phenotype-first approach. :contentReference[oaicite:0]{index=0}

Argues for integrating genetics-first and phenotype-first strategies with high-throughput phenotyping to enhance disease understanding. :contentReference[oaicite:1]{index=1}

Therapeutics

Drug discovery, repurposing, and therapeutic intervention research

The Tell-Tale Heart: Population-Based Surveillance Reveals an Association of Rofecoxib and Celecoxib with Myocardial Infarction

PLOS One (2007) – COX-2 selective inhibitors are associated with myocardial infarction (MI).

Population-level surveillance at two Boston hospitals found an 18.5% rise in MI hospitalizations concurrent with COX-2 inhibitor prescriptions. The hospitals were not aware during the exposure.

Drug target-gene signatures that predict teratogenicity are enriched for developmentally related genes

Reprod Toxicol (2011) – Drugs prescribed during pregnancy affect two populations simultaneously: fetuses and their mothers.

Identifies developmentally enriched target‑gene signatures that predict teratogenic risk with 79% accuracy.

Discovering functional relationships between RNA expression and chemotherapeutic susceptibility using relevance networks

Proc Natl Acad Sci U S A (2000) – Novel method to find gene regulatory networks and clusters of genes that affect cancer susceptibility to anticancer agents.

Joined expression profiles of 7,245 genes in 60 cancer cell lines with sensitivity data for 5,084 drugs to build “relevance networks,” identifying candidate single‑gene determinants of chemotherapeutic response.}

<

Rapid Identification of Myocardial Infarction Risk Associated With Diabetes Medications Using Electronic Medical Records

Diabetes Care (2010) – To assess the ability to identify potential association(s) of diabetes medications with myocardial infarction using usual care clinical data obtained from the electronic medical record.

EHR analysis of thousands of patients led to "black-boxing" of a hypoglycemic agent.

Policy

Healthcare policy, regulation, and ethical frameworks for medical AI

Medicine. Reestablishing the researcher-patient compact

Science (2007) – Well-intentioned regulations protecting privacy are denying important information to patient subjects.

Provides early blueprint to leverage modern information technology to restore bidirectional communication between researchers and participants while preserving privacy and autonomy

Multidimensional results reporting to participants in genomic studies: getting it right

Sci Transl Med (2010) – Recent surveys about participation in cohort studies reconfirm that participants value and desire the return of research results, contradicting restrictive ethics recommendations.

Introduces communicability as a research variable to structure ethics-based reporting and align participant preferences with results disclosure frameworks. :contentReference[oaicite:0]{index=0}

Medical Artificial Intelligence and Human Values

N Engl J Med (2024) – Review

In this review, Yu et al. examine how and where human values and ethics do (and do not) inform AI programs.

To do no harm – and the most good – with AI in health care

Nat Med (2024) – Perspective on ensuring AI in health care does minimal harm and maximal benefit.

Goldberg et al. discuss ethical and practical frameworks to guide safe, equitable, and transparent AI deployment in medicine.

"\The AI Revolution in Medicine: GPT-4 and Beyond

Pearson (2023) - Comprehensive guide to medical AI policy

Early book on LLM transformation of healthcare practice, research including policy implications and regulatory considerations.

Designing a public square for research computing

Sci Transl Med (2012) – A set of principles proposed for sponsors and developers of research computing applications that can increase the likelihood of successful adoption by researchers.

Proposes principles to boost adoption of research computing.

No small change for the health information economy

N Engl J Med (2009) – President Obama’s economic stimulus package includes a $19 billion investment in health information technology.}

Mandl and Kohane discuss how monolithic EHRs can be replaced by a thriving app economy.

Understanding Covid Vaccine Efficacy over Time - Bridging a Gap Between Public Health and Health Care

N Engl J Med (2022) – It's now feasible to track, for each patient whichvaccine they received, when and what theirclinical evaluation revealed.

Dicusses how linking vaccination dates with clinical data enables near–real‑time monitoring of vaccine efficacy to inform both public health policy and clinical care.

Perspective: The Human Values Project

Proceedings of Machine Learning for Health Symposium (2025) – Alignment of AI models to ensure that they are safe and useful decision‑aids or decision‑makers in human society is close to the top of the technical concerns of many if not most major AI deployment efforts.

Motivates an international Human Values Project by exploring AI‑driven medical triage decisions and proposing studies to capture descriptive and normative clinical decision dynamics.

Explore Our Research Impact

Dive deeper into 30+ years of groundbreaking research that has shaped the field of biomedical informatics and clinical AI.