An explainable artificial intelligence approach for predicting cardiovascular outcomes using electronic health records

Understanding the conditionally-dependent clinical variables that drive cardiovascular health outcomes is a major challenge for precision medicine. Here, we deploy a recently developed massively scalable comorbidity discovery method called Poisson Binomial based Comorbidity discovery (PBC), to analyze Electronic Health Records (EHRs) from the University of Utah and Primary Children's Hospital (over 1.6 million patients and 77 million visits) for comorbid diagnoses, procedures, and medications. Using explainable Artificial Intelligence (AI) methodologies, we then tease apart the intertwined, conditionally-dependent impacts of comorbid conditions and demography upon cardiovascular health, focusing on the key areas of heart transplant, sinoatrial node dysfunction and various forms of congenital heart disease. The resulting multimorbidity networks make possible wide-ranging explorations of the comorbid and demographic landscapes surrounding these cardiovascular outcomes, and can be distributed as web-based tools for further community-based outcomes research. The ability to transform enormous collections of EHRs into compact, portable tools devoid of Protected Health Information solves many of the legal, technological, and data-scientific challenges associated with large-scale EHR analyses.

The history and geographic distribution of a KCNQ1 atrial fibrillation risk allele

The genetic architecture of atrial fibrillation (AF) encompasses low impact, common genetic variants and high impact, rare variants. Here, we characterize a high impact AF-susceptibility allele, KCNQ1 R231H, and describe its transcontinental geographic distribution and history. Induced pluripotent stem cell-derived cardiomyocytes procured from risk allele carriers exhibit abbreviated action potential duration, consistent with a gain-of-function effect. Using identity-by-descent (IBD) networks, we estimate the broad- and fine-scale population ancestry of risk allele carriers and their relatives. Analysis of ancestral migration routes reveals ancestors who inhabited Denmark in the 1700s, migrated to the Northeastern United States in the early 1800s, and traveled across the Midwest to arrive in Utah in the late 1800s. IBD/coalescent-based allele dating analysis reveals a relatively recent origin of the AF risk allele (~5000 years). Thus, our approach broadens the scope of study for disease susceptibility alleles to the context of human migration and ancestral origins.

A Poisson binomial-based statistical testing framework for comorbidity discovery across electronic health record datasets

Discovering the concomitant occurrence of distinct medical conditions in a patient, also known as comorbidities, is a prerequisite for creating patient outcome prediction tools. Current comorbidity discovery applications are designed for small datasets and use stratification to control for confounding variables such as age, sex or ancestry. Stratification lowers false positive rates, but reduces power, as the size of the study cohort is decreased. Here we describe a Poisson binomial-based approach to comorbidity discovery (PBC) designed for big-data applications that circumvents the need for stratification. PBC adjusts for confounding demographic variables on a per-patient basis and models temporal relationships. We benchmark PBC using two datasets to compute comorbidity statistics on 4,623,841 pairs of potentially comorbid medical terms. The results of this computation are provided as a searchable web resource. Compared with current methods, the PBC approach reduces false positive associations while retaining statistical power to discover true comorbidities.

Differences in the genomic landscape of advanced prostate cancer (aPC) patients (pts) with BRCA1 versus BRCA2 mutations as detected by machine learning analysis of the comprehensive genomic profile (CGP) of cell-free DNA (cfDNA)

162

Background: PARP inhibitors (PARPi) provide significant clinical benefit for men with aPC with BRCA 1 and BRCA 2 mutations. However, in clinical trials, pts with BRCA1 mutations appeared to derive less benefit than pts with BRCA2 (De Bono et al., 2020). Probabilistic Graphical Models (PGMs) are artificial intelligence (AI) algorithms that capture multivariate, multi-level dependencies in complex patterns in large datasets while retaining human interpretability. We hypothesize that PGMs can reveal variants in BRCA1 and 2 that co-segregate with other known pathogenic variants and may explain the difference in response to PARPi therapy. Methods: Multilevel gene interdependencies between BRCA1 or BRCA2 were assessed using a Bayesian Network (BN) machine learning approach and Fisher’s exact test. CGP was performed by a validated cfDNA NGS panel that sequenced 74 clinically relevant cancer genes (Guardant360, Redwood City, CA). Only variants of known significance and those of unknown significance with a pathogenic REVEL score were included in the analysis. Results: Of 4671 men with aPC undergoing cfDNA CGP, 1248 men with somatic mutations in BRCA1, BRCA2, ATM, or combinations of the three were included in the analysis. The Bayesian network analysis demonstrated positive interdependencies between pathogenic variants in BRCA1 and 7 other genes. A positive interdependency between BRCA2 and 2 genes was present (table). ATM displayed negative interdependency with both BRCA 1 and 2. Conclusions: Our results demonstrate a decreased association of BRCA2 versus BRCA1 with known or predicted pathogenic variants at other loci. This may explain increased sensitivity of aPC with BRCA2 mutations to PARPi due to fewer concurrent resistance pathways. For example, alteration of ERBB2, which segregates strongly with BRCA1, is known to induce tumor progression and invasion in aPC and is associated with castration-resistance. These hypothesis-generating data reveal differential genomic signatures associated with BRCA1 as compared to BRCA2 and may inform development of future combinatorial treatment regimens for these pts. [Table: see text]

Antigenic evolution of H9N2 chicken influenza viruses isolated in China during 2009-2013 and selection of a candidate vaccine strain with broad cross-reactivity

We previously demonstrated that H9N2 subtype avian influenza viruses (AIVs) isolated from 1994 to 2008 evolved into distinct antigenic groups (C, D, and E) and then underwent antigenic drift from commercial vaccines, causing a country-wide outbreak during 2010-2013. In this study, H9N2 AIVs isolated from chickens during 2009-2013 were antigenically analyzed by performing hemagglutination inhibition and neutralization assays using a panel of polyclonal antibodies. Our findings confirmed the antigenic drift of recent H9N2 viruses from the commercial vaccine and showed that most of these antigenic variants form a novel HI antigenic group, F, with a few belonging to groups D and E. Slight antigenic variation was observed in group F viruses. Genetic analysis of amino acid sequences deduced from hemagglutinin (HA) gene sequences indicated that 9 of 15 mutations predominant in the 2009-2013 viruses can be mapped to known antigenic sites, which might be responsible for the novel antigenicity of group F. These antigenic changes make it necessary to modify the influenza vaccine to ensure efficient protection. A vaccine candidate, Ck/HeB/YT/10, was selected and provided significant protection against viruses from different antigenic groups in terms of reduction in virus shedding, suggesting broad cross-reactivity. Taken together, our results indicate that the H9N2 chicken influenza viruses in China have evolved from distinct antigenic groups into a novel group F that became dominant during the country-wide outbreak and now seems to be undergoing new antigenic divergence. Systematic surveillance and timely updating of vaccine strains are important for viral prevention and control in the future.

Genomic landscape of Ewing sarcoma defines an aggressive subtype with co-association of STAG2 and TP53 mutations

Ewing sarcoma is a primary bone tumor initiated by EWSR1-ETS gene fusions. To identify secondary genetic lesions that contribute to tumor progression, we performed whole-genome sequencing of 112 Ewing sarcoma samples and matched germline DNA. Overall, Ewing sarcoma tumors had relatively few single-nucleotide variants, indels, structural variants, and copy-number alterations. Apart from whole chromosome arm copy-number changes, the most common somatic mutations were detected in STAG2 (17%), CDKN2A (12%), TP53 (7%), EZH2, BCOR, and ZMYM3 (2.7% each). Strikingly, STAG2 mutations and CDKN2A deletions were mutually exclusive, as confirmed in Ewing sarcoma cell lines. In an expanded cohort of 299 patients with clinical data, we discovered that STAG2 and TP53 mutations are often concurrent and are associated with poor outcome. Finally, we detected subclonal STAG2 mutations in diagnostic tumors and expansion of STAG2-immunonegative cells in relapsed tumors as compared with matched diagnostic samples.

SIGNIFICANCE

Whole-genome sequencing reveals that the somatic mutation rate in Ewing sarcoma is low. Tumors that harbor STAG2 and TP53 mutations have a particularly dismal prognosis with current treatments and require alternative therapies. Novel drugs that target epigenetic regulators may constitute viable therapeutic strategies in a subset of patients with mutations in chromatin modifiers.

The landscape of somatic mutations in epigenetic regulators across 1,000 paediatric cancer genomes

Studies of paediatric cancers have shown a high frequency of mutation across epigenetic regulators. Here we sequence 633 genes, encoding the majority of known epigenetic regulatory proteins, in over 1,000 paediatric tumours to define the landscape of somatic mutations in epigenetic regulators in paediatric cancer. Our results demonstrate a marked variation in the frequency of gene mutations across 21 different paediatric cancer subtypes, with the highest frequency of mutations detected in high-grade gliomas, T-lineage acute lymphoblastic leukaemia and medulloblastoma, and a paucity of mutations in low-grade glioma and retinoblastoma. The most frequently mutated genes are H3F3A, PHF6, ATRX, KDM6A, SMARCA4, ASXL2, CREBBP, EZH2, MLL2, USP7, ASXL1, NSD2, SETD2, SMC1A and ZMYM3. We identify novel loss-of-function mutations in the ubiquitin-specific processing protease 7 (USP7) in paediatric leukaemia, which result in decreased deubiquitination activity. Collectively, our results help to define the landscape of mutations in epigenetic regulatory genes in paediatric cancer and yield a valuable new database for investigating the role of epigenetic dysregulations in cancer.

Simultaneous prediction of binding free energy and specificity for PDZ domain-peptide interactions

Interactions between protein domains and linear peptides underlie many biological processes. Among these interactions, the recognition of C-terminal peptides by PDZ domains is one of the most ubiquitous. In this work, we present a mathematical model for PDZ domain-peptide interactions capable of predicting both affinity and specificity of binding based on X-ray crystal structures and comparative modeling with ROSETTA. We developed our mathematical model using a large phage display dataset describing binding specificity for a wild type PDZ domain and 91 single mutants, as well as binding affinity data for a wild type PDZ domain binding to 28 different peptides. Structural refinement was carried out through several ROSETTA protocols, the most accurate of which included flexible peptide docking and several iterations of side chain repacking and backbone minimization. Our findings emphasize the importance of backbone flexibility and the energetic contributions of side chain-side chain hydrogen bonds in accurately predicting interactions. We also determined that predicting PDZ domain-peptide interactions became increasingly challenging as the length of the peptide increased in the N-terminal direction. In the training dataset, predicted binding energies correlated with those derived through calorimetry and specificity switches introduced through single mutations at interface positions were recapitulated. In independent tests, our best performing protocol was capable of predicting dissociation constants well within one order of magnitude of the experimental values and specificity profiles at the level of accuracy of previous studies. To our knowledge, this approach represents the first integrated protocol for predicting both affinity and specificity for PDZ domain-peptide interactions.

Computational design of protein-small molecule interfaces

The computational design of proteins that bind small molecule ligands is one of the unsolved challenges in protein engineering. It is complicated by the relatively small size of the ligand which limits the number of intermolecular interactions. Furthermore, near-perfect geometries between interacting partners are required to achieve high binding affinities. For apolar, rigid small molecules the interactions are dominated by short-range van der Waals forces. As the number of polar groups in the ligand increases, hydrogen bonds, salt bridges, cation-π, and π-π interactions gain importance. These partial covalent interactions are longer ranged, and additionally, their strength depends on the environment (e.g. solvent exposure). To assess the current state of protein-small molecule interface design, we benchmark the popular computer algorithm Rosetta on a diverse set of 43 protein-ligand complexes. On average, we achieve sequence recoveries in the binding site of 59% when the ligand is allowed limited reorientation, and 48% when the ligand is allowed full reorientation. When simulating the redesign of a protein binding site, sequence recovery among residues that contribute most to binding was 52% when slight ligand reorientation was allowed, and 27% when full ligand reorientation was allowed. As expected, sequence recovery correlates with ligand displacement.

Towards ligand docking including explicit interface water molecules

Small molecule docking predicts the interaction of a small molecule ligand with a protein at atomic-detail accuracy including position and conformation the ligand but also conformational changes of the protein upon ligand binding. While successful in the majority of cases, docking algorithms including RosettaLigand fail in some cases to predict the correct protein/ligand complex structure. In this study we show that simultaneous docking of explicit interface water molecules greatly improves Rosetta’s ability to distinguish correct from incorrect ligand poses. This result holds true for both protein-centric water docking wherein waters are located relative to the protein binding site and ligand-centric water docking wherein waters move with the ligand during docking. Protein-centric docking is used to model 99 HIV-1 protease/protease inhibitor structures. We find protease inhibitor placement improving at a ratio of 9∶1 when one critical interface water molecule is included in the docking simulation. Ligand-centric docking is applied to 341 structures from the CSAR benchmark of diverse protein/ligand complexes [1]. Across this diverse dataset we see up to 56% recovery of failed docking studies, when waters are included in the docking simulation.

Prediction of HIV-1 protease/inhibitor affinity using RosettaLigand

Predicting HIV-1 protease/inhibitor binding affinity as the difference between the free energy of the inhibitor bound and unbound state remains difficult as the unbound state exists as an ensemble of conformations with various degrees of flap opening. We improve computational prediction of protease/inhibitor affinity by invoking the hypothesis that the free energy of the unbound state while difficult to predict is less sensitive to mutation. Thereby the HIV-1 protease/inhibitor binding affinity can be approximated with the free energy of the bound state alone. Bound state free energy can be predicted from comparative models of HIV-1 protease mutant/inhibitor complexes. Absolute binding energies are predicted with R = 0.71 and SE = 5.91 kJ/mol. Changes in binding free energy upon mutation can be predicted with R = 0.85 and SE = 4.49 kJ/mol. Resistance mutations that lower inhibitor binding affinity can thereby be recognized early in HIV-1 protease inhibitor development.

ABCB1 and GST polymorphisms associated with TP53 status in breast cancer

BACKGROUND AND OBJECTIVE

Many environmental and genetic factors influence the development of chemoresistance. The goal of this study was to characterize the genetic variation in the ABCB1, GSTM1, GSTT1 and GSTP1 genes, as well as the haplotype structure in the ABCB1 gene.

METHODS

Variants in these genes were studied in 109 healthy controls and 93 breast cancer cases, both of Caucasian origin. The cases were analyzed in relation to TP53 mutation status and response to doxorubicin. Both single and multiple single nucleotide polymorphism analyses were performed.

RESULTS

Chi-square analyses revealed a significant association between TP53 mutation status and both the GA genotype of ABCB1 exon 11 (Ser400Asn) and the GG genotype of GSTP1 (Ile105Val; P<0.01 and P<0.05, respectively). Multifactor dimensionality reduction showed that carriers of the combined GG genotype for GSTP1 and the GG for ABCB1 exon 11 had the highest chance of acquiring a mutation in the TP53 gene (P<0.02). Haplotype analysis of ABCB1 revealed a significantly different distribution of haplotypes between the breast cancer cases and the controls (P<0.01). A specific haplotype association to TP53 mutation (P<0.01) distant metastases (P<0.05) and estrogen receptor status (P<0.05) was also observed in the case group.

CONCLUSION

An association between polymorphisms in GSTP1 and ABCB1 and risk of acquiring intratumoral TP53 mutations suggests the existence of putative predisposing genotype backgrounds. The degree of linkage disequilibrium in the ABCB1 gene was higher in healthy individuals, whereas haplotypes in the cases seemed degenerated by a number of low frequency variants. This observation may either point to the existence of a protective haplotype in the controls or may underline the importance of the accumulation of low frequency variants as susceptibility factors.

Cervical spine instability: clearance using dynamic fluoroscopy

Cle aring the cervical spine in a multiply injured trauma patient is a dilemma because clinical examination for ligamentous instability cannot be performed, and the standard cervical spine series can miss isolated ligamentous injury. Static flexion/extension views are unsafe, as the obtunded patient has no protective reflexes and cannot complain of pain during the exam. This results in a need for prolonged spinal immobilization and its attendant complications. Dynamic fluoroscopy may be useful in the detection of otherwise occult injuries.We performed a prospective study of a cervical spine clearance algorithm incorporating dynamic fluoroscopy with flexion/extension views. Inpatient records over a 3-year period were reviewed. Patient demographic data, results of cervical spine films and fluoroscopic exams, interventions based on positive results, and missed injuries were recorded.One hundred ten patients with normal spine plain films underwent dynamic fluoroscopy with flexion and extension views of the cervical spine. The average Glasgow Coma Score was 9.2 and the average revised Trauma Score was 9.5. Nine patients had evidence of cervical instability on exam. Six of these were deemed stable by the orthopedic or neurosurgical spine consultants, and these patients had their hard collars removed. One patient with positive findings had cervical immobilization with hard collar continued, a second had halo placement, and a third underwent spinal fusion for atlanto-occipital disassociation. No patients undergoing dynamic fluoroscopy were subsequently found to have missed cervical spine injury.With our protocol, 3 patients had significant cervical instability that would have been missed without dynamic fluoroscopy. Given the significant medical and legal ramifications of missed cervical spine injury and the benefits of early removal of cervical collars, more widespread use of dynamic fluoroscopy of the cervical spine is warranted.

The role of glutamine in skeletal muscle ischemia/reperfusion injury in the rat hind limb model

OBJECTIVES

Ischemia/reperfusion injury is a commonly occurring event with severe pathologic consequences. Reperfusion initiates both the local and systematic damage in part through rapid oxygen generation. The glutathione system is a major mechanism of reducing this oxidative stress. If this system can be maintained or augmented during this stress then less damage may occur. Glutamine provides the source of glutamate to this system and has been shown to preserve total glutathione levels after injury/ischemia to both hepatic and gut models. To test this effect, we looked at glutamine and its role in ischemia/reperfusion injury in a rat hind limb model.

METHODS

Fifty male HSD/Holtzman rats weighing 350-400 g were randomized to receive glutamine (3% sol) or normal saline via intraperitoneal injections. The groups were then subjected to 2 hours of ischemia to their hind limbs using the Tourni-Cot method. Animals were then randomized to reperfusion groups of 30 minutes, 2 hours, and 4 hours. Muscle tissue assays were performed for lipid peroxidation (LPO), total glutathione (GSH), and myeloperoxidase (MPO). Peripheral blood was analyzed for creatinephosphokinase levels (CPK).

RESULTS

Animals that received glutamine showed a general trend of less lipid peroxidation products than the normal saline groups. In animals that received glutamine and underwent 2 hours of ischemia and reperfusion times of 0 minutes, 30 minutes, and 2 hours, there were significantly less percent changes in lipid peroxidation products from controls (4.6% vs 48.2%, P <0.05), (18.9% vs 123%, P <0.05), (12.6% vs 115%, P <0.05). A general trend upward was noted in CPK levels in both groups. In animals receiving 2 hours of ischemia and 30 minutes of reperfusion, there was a significantly greater level of creatinephosphokinase (CPK) calculated as percent change from control in the normal saline group as compared with the glutamine group (209.2% vs 92.7%). Myeloperoxidase assay of muscle tissue revealed a progressive increase as the reperfusion times grew. In animals receiving 2 hours of ischemia and 30 minutes of reperfusion, the normal saline group had a significantly larger percent increase from controls than the group that received glutamine (1126.4% vs 108%, P <0.05). Also, in those animals receiving 4 hours of reperfusion, the normal saline group had a significantly higher percent increase in MPO content than the glutamine group (6245% vs 108%, P <0.05). Total glutathione levels decreased rapidly as reperfusion occurred in both the normal saline and glutamine groups. No significant difference between the groups was noted.

CONCLUSIONS

Total glutathione levels during reperfusion were not significantly different in the groups receiving glutamine versus normal saline. Glutamine may provide an initial protective effect on reperfusion injury after moderate reperfusion times in the hind limb model as defined by CPK and LPO levels. Glutamine may blunt neutrophil recruitment after longer reperfusion times (4 hours) in the ischemic hind limb. Total glutathione levels decreased significantly after moderate levels of ischemia (2 hours) and reperfusion (30 minutes, 2 hours).

Clinical experience with the Greenfield filter in 193 patients and description of a new technique for operative insertion

Transvenous inferior vena cava (IVC) interruption using the Greenfield filter was performed upon 193 patients from January 1982 to December 1988. Placement of a filter in the IVC was indicated for prophylaxis (23.8 percent), contraindication to anticoagulation (22.8 percent), pulmonary embolism despite anticoagulation (18.7 percent), complications of anticoagulation (26.9 percent) and free-floating thrombus (7.8 percent). Filters were placed in an infrarenal location in 92 percent of the patients. The remaining 8 percent of patients had placement of a suprarenal filter for specific indications, without complication. Most (97.4 percent) of the patients had filters placed through internal jugular or femoral vein cutdown. Five patients required filter placement through a retroperitoneal approach to the right common iliac vein and IVC junction. This new technique of filter insertion is described. The operative morbidity rate was 4.7 percent, with an additional 8.8 percent having postoperative thrombotic complications. The 30 day operative mortality rate (6.7 percent) was related to preexisting associated disease. Nonfatal, late, recurrent pulmonary embolism occurred in 2.6 percent of the patients despite filter placement. Caval patency remains at 97.9 percent in long term follow-up evaluation. The Greenfield filter is an effective and safe adjunct in the treatment of venous thromboembolic disease and a satisfactory prophylactic measure in specific high-risk patients.