Postprediction Inference for Clinical Characteristics Extracted With Machine Learning on Electronic Health Records.

PURPOSE: Real-world data (RWD) derived from electronic health records (EHRs) are often used to understand population-level relationships between patient characteristics and cancer outcomes. Machine learning (ML) methods enable researchers to extract characteristics from unstructured clinical notes, and represent a more cost-effective and scalable approach than manual expert abstraction. These extracted data are then used in epidemiologic or statistical models as if they were abstracted observations. Analytical results derived from extracted data in this way may differ from those given by abstracted data, and the magnitude of this difference is not directly informed by standard ML performance metrics. METHODS: In this paper, we define the task of postprediction inference, which is to recover similar estimation and inference from an ML-extracted variable that would be obtained from abstracting the variable. We consider fitting a Cox proportional hazards model that uses a binary ML-extracted variable as a covariate and evaluate four approaches for postprediction inference in this setting. The first two approaches only require the ML-predicted probability, while the latter two additionally require a labeled (human abstracted) validation data set. RESULTS: Our results for both simulated data and EHR-derived RWD from a national cohort demonstrate that we can improve inference from ML-extracted variables by leveraging a limited amount of labeled data. CONCLUSION: We describe and evaluate methods for fitting statistical models using ML-extracted variables subject to model error. We show that estimation and inference is generally valid when using extracted data from high-performing ML models. More complex methods that incorporate auxiliary labeled data provide further improvements.

The effect of population-based blood pressure screening on long-term cardiometabolic morbidity and mortality in Germany: A regression discontinuity analysis.

BACKGROUND: Hypertension represents one of the major risk factors for cardiovascular morbidity and mortality globally. Early detection and treatment of this condition is vital to prevent complications. However, hypertension often goes undetected, and even if detected, not every patient receives adequate treatment. Identifying simple and effective interventions is therefore crucial to fight this problem and allow more patients to receive the treatment they need. Therefore, we aim at investigating the impact of a population-based blood pressure (BP) screening and the subsequent "low-threshold" information treatment on long-term cardiovascular disease (CVD) morbidity and mortality. METHODS AND FINDINGS: We examined the impact of a BP screening embedded in a population-based cohort study in Germany and subsequent personalized "light touch" information treatment, including a hypertension diagnosis and a recommendation to seek medical attention. We pooled four waves of the KORA study, carried out between 1984 and 1996 (N = 14,592). Using a sharp multivariate regression discontinuity (RD) design, we estimated the impact of the information treatment on CVD mortality and morbidity over 16.9 years. Additionally, we investigated potential intermediate outcomes, such as hypertension awareness, BP, and behavior after 7 years. No evidence of effect of BP screening was observed on CVD mortality (hazard ratio (HR) = 1.172 [95% confidence interval (CI): 0.725, 1.896]) or on any (fatal or nonfatal) long-term CVD event (HR = 1.022 [0.636, 1.641]) for individuals just above (versus below) the threshold for hypertension. Stratification for previous self-reported diagnosis of hypertension at baseline did not reveal any differential effect. The intermediate outcomes, including awareness of hypertension, were also unaffected by the information treatment. However, these results should be interpreted with caution since the analysis might not be sufficiently powered to detect a potential intervention effect. CONCLUSIONS: The study does not provide evidence of an effect of the assessed BP screening and subsequent information treatment on BP, health behavior, or long-term CVD mortality and morbidity. Future studies should consider larger datasets to detect possible effects and a shorter follow-up for the intermediate outcomes (i.e., BP and behavior) to detect short-, medium-, and long-term effects of the intervention along the causal pathway.

A machine learning framework supporting prospective clinical decisions applied to risk prediction in oncology.

We present a general framework for developing a machine learning (ML) tool that supports clinician assessment of patient risk using electronic health record-derived real-world data and apply the framework to a quality improvement use case in an oncology setting to identify patients at risk for a near-term (60 day) emergency department (ED) visit who could potentially be eligible for a home-based acute care program. Framework steps include defining clinical quality improvement goals, model development and validation, bias assessment, retrospective and prospective validation, and deployment in clinical workflow. In the retrospective analysis for the use case, 8% of patient encounters were associated with a high risk (pre-defined as predicted probability ≥20%) for a near-term ED visit by the patient. Positive predictive value (PPV) and negative predictive value (NPV) for future ED events was 26% and 91%, respectively. Odds ratio (OR) of ED visit (high- vs. low-risk) was 3.5 (95% CI: 3.4-3.5). The model appeared to be calibrated across racial, gender, and ethnic groups. In the prospective analysis, 10% of patients were classified as high risk, 76% of whom were confirmed by clinicians as eligible for home-based acute care. PPV and NPV for future ED events was 22% and 95%, respectively. OR of ED visit (high- vs. low-risk) was 5.4 (95% CI: 2.6-11.0). The proposed framework for an ML-based tool that supports clinician assessment of patient risk is a stepwise development approach; we successfully applied the framework to an ED visit risk prediction use case.

The limits of learning: Exploration, generalization, and the development of learning traps.

Learning usually improves the accuracy of beliefs through the accumulation of experience. But are there limits to learning that prevent us from accurately understanding our world? In this article we investigate the concept of a "learning trap"-the formation of a stable false belief even with extensive experience. Our review highlights how these traps develop through the interaction of learning and decision making in unknown environments. We further document a particularly pernicious learning trap driven by selective attention, a mechanism often assumed to facilitate learning in complex environments. Using computer simulation, we demonstrate the key attributes of the agent and environment that lead to this new type of learning trap. Then, in a series of experiments we present evidence that people robustly fall into this trap, even in the presence of various interventions predicted to meliorate it. These results highlight a fundamental limit to learning and adaptive behavior that impacts individuals, organizations, animals, and machines. (PsycINFO Database Record (c) 2018 APA, all rights reserved).

psiTurk: An open-source framework for conducting replicable behavioral experiments online.

Online data collection has begun to revolutionize the behavioral sciences. However, conducting carefully controlled behavioral experiments online introduces a number of new of technical and scientific challenges. The project described in this paper, psiTurk, is an open-source platform which helps researchers develop experiment designs which can be conducted over the Internet. The tool primarily interfaces with Amazon's Mechanical Turk, a popular crowd-sourcing labor market. This paper describes the basic architecture of the system and introduces new users to the overall goals. psiTurk aims to reduce the technical hurdles for researchers developing online experiments while improving the transparency and collaborative nature of the behavioral sciences.

Epigenetic regulation by Z-DNA silencer function controls cancer-associated ADAM-12 expression in breast cancer: cross-talk between MeCP2 and NF1 transcription factor family.

A disintegrin and metalloprotease domain-containing protein 12 (ADAM-12) is upregulated in many human cancers and promotes cancer metastasis. Increased urinary level of ADAM-12 in breast and bladder cancers correlates with disease progression. However, the mechanism of its induction in cancer remains less understood. Previously, we reported a Z-DNA-forming negative regulatory element (NRE) in ADAM-12 that functions as a transcriptional suppressor to maintain a low-level expression of ADAM-12 in most normal cells. We now report here that overexpression of ADAM-12 in triple-negative MDA-MB-231 breast cancer cells and breast cancer tumors is likely due to a marked loss of this Z-DNA-mediated transcriptional suppression function. We show that Z-DNA suppressor operates by interaction with methyl-CpG-binding protein, MeCP2, a prominent epigenetic regulator, and two members of the nuclear factor 1 family of transcription factors, NF1C and NF1X. While this tripartite interaction is highly prevalent in normal breast epithelial cells, both in vitro and in vivo, it is significantly lower in breast cancer cells. Western blot analysis has revealed significant differences in the levels of these 3 proteins between normal mammary epithelial and breast cancer cells. Furthermore, we show, by NRE mutation analysis, that interaction of these proteins with the NRE is necessary for effective suppressor function. Our findings unveil a new epigenetic regulatory process in which Z-DNA/MeCP2/NF1 interaction leads to transcriptional suppression, loss of which results in ADAM-12 overexpression in breast cancer cells.

Solution structure of the Zbeta domain of human DNA-dependent activator of IFN-regulatory factors and its binding modes to B- and Z-DNAs.

The DNA-dependent activator of IFN-regulatory factors (DAI), also known as DLM-1/ZBP1, initiates an innate immune response by binding to foreign DNAs in the cytosol. For full activation of the immune response, three DNA binding domains at the N terminus are required: two Z-DNA binding domains (ZBDs), Zα and Zß, and an adjacent putative B-DNA binding domain. The crystal structure of the Zß domain of human DAI (hZß(DAI)) in complex with Z-DNA revealed structural features distinct from other known Z-DNA binding proteins, and it was classified as a group II ZBD. To gain structural insights into the DNA binding mechanism of hZß(DAI), the solution structure of the free hZß(DAI) was solved, and its bindings to B- and Z-DNAs were analyzed by NMR spectroscopy. Compared to the Z-DNA-bound structure, the conformation of free hZß(DAI) has notable alterations in the α3 recognition helix, the "wing," and Y145, which are critical in Z-DNA recognition. Unlike some other Zα domains, hZß(DAI) appears to have conformational flexibility, and structural adaptation is required for Z-DNA binding. Chemical-shift perturbation experiments revealed that hZß(DAI) also binds weakly to B-DNA via a different binding mode. The C-terminal domain of DAI is reported to undergo a conformational change on B-DNA binding; thus, it is possible that these changes are correlated. During the innate immune response, hZß(DAI) is likely to play an active role in binding to DNAs in both B and Z conformations in the recognition of foreign DNAs.

Hearing loss alters quantal release at cochlear nucleus stellate cells.

OBJECTIVES/HYPOTHESIS: Auditory nerve synapses in ventral cochlear nucleus end on two principal cell types, bushy and stellate cells. Although the effects of hearing loss on bushy cells have been well studied, little is known about the effects of hearing loss on synaptic input to the stellate cells. Based on prior observations in bushy cells, we hypothesized that noise-induced hearing loss (NIHL) would decrease quantal release onto stellate cells. STUDY DESIGN: Prospective, randomized animal study. METHODS: CBA/CaJ mice were exposed for 2 hours to 98 dB sound pressure level (SPL) 8- to 16-kHz noise to produce a temporary threshold shift (TTS) or 114 dB SPL to produce a permanent threshold shift (PTS). Spontaneous miniature excitatory postsynaptic currents (mEPSCs) were then measured in stellate cells in brain slices of the cochlear nucleus. RESULTS: Click auditory brainstem evoked response thresholds were elevated by 35 dB in both TTS and PTS mice. Spontaneous mEPSC frequency was found to be five-fold higher than normal in stellate cells of TTS mice and three-fold higher in PTS mice. The mEPSC amplitude was also larger in PTS mice. The mEPSC time course was not different between experimental and control groups. CONCLUSIONS: The dramatic increase in mEPSC frequency after NIHL was not expected. The increase in mEPSC amplitude in PTS mice suggests a postsynaptic remodeling process. Both of these effects could contribute to increased spontaneous firing in the cochlear nucleus in the absence of sound. Our results also suggest that hearing loss may have different effects at auditory nerve synapses on bushy and stellate cells in the VCN.

Crystal structure of a junction between two Z-DNA helices.

The double helix of DNA, when composed of dinucleotide purine-pyrimidine repeats, can adopt a left-handed helical structure called Z-DNA. For reasons not entirely understood, such dinucleotide repeats in genomic sequences have been associated with genomic instability leading to cancer. Adoption of the left-handed conformation results in the formation of conformational junctions: A B-to-Z junction is formed at the boundaries of the helix, whereas a Z-to-Z junction is commonly formed in sequences where the dinucleotide repeat is interrupted by single base insertions or deletions that bring neighboring helices out of phase. B-Z junctions are shown to result in exposed nucleotides vulnerable to chemical or enzymatic modification. Here we describe the three-dimensional structure of a Z-Z junction stabilized by Zalpha, the Z-DNA binding domain of the RNA editing enzyme ADAR1. We show that the junction structure consists of a single base pair and leads to partial or full disruption of the helical stacking. The junction region allows intercalating agents to insert themselves into the left-handed helix, which is otherwise resistant to intercalation. However, unlike a B-Z junction, in this structure the bases are not fully extruded, and the stacking between the two left-handed helices is not continuous.

The era of RNA awakening: structural biology of RNA in the early years.

In the mid-1950s, RNA was a somewhat mysterious molecule with unknown three-dimensional structure and little hard evidence of biological function. Changes began with the 1956 discoveries of the RNA double helix and the phenomenon of nucleic acid hybridization. Discovery of the DNA-RNA hybrid helix in 1960 opened the door to understanding biological information transfer. Single-crystal X-ray diffraction analysis made it possible to precisely define the RNA double helix, discover the novel L-shaped fold of transfer RNA (tRNA), and finally reveal the complete three-dimensional tRNA structure by 1974. By then, a functional understanding of protein synthesis had developed with an appreciation of the various roles of different RNA species. This was the era of RNA awakening.

Base extrusion is found at helical junctions between right- and left-handed forms of DNA and RNA.

Base extrusion is a major structural feature at the junction between B- and Z-DNA (the B-Z junction) where a base pair is broken, and the two bases are extruded from the double helix. Despite the demonstration of base extrusion at the B-Z junction, it is not clear whether a similar base extrusion occurs at other types of junctions involving the left-handed Z conformation. Here, we investigate structural changes of bases at three Z-form junctions: DNA B-Z and Z-Z and RNA A-Z junctions. By monitoring fluorescently labeled duplex nucleic acids using 2-aminopurines at various positions relative to the junction point, we show that base extrusion occurs not only at the DNA B-Z junction, but also at the RNA A-Z and DNA Z-Z junctions. Our data suggest that base extrusion is a general feature of Z-form nucleic-acid junctions.

The structures of non-CG-repeat Z-DNAs co-crystallized with the Z-DNA-binding domain, hZ alpha(ADAR1).

The Z-DNA conformation preferentially occurs at alternating purine-pyrimidine repeats, and is specifically recognized by Z alpha domains identified in several Z-DNA-binding proteins. The binding of Z alpha to foreign or chromosomal DNA in various sequence contexts is known to influence various biological functions, including the DNA-mediated innate immune response and transcriptional modulation of gene expression. For these reasons, understanding its binding mode and the conformational diversity of Z alpha bound Z-DNAs is of considerable importance. However, structural studies of Z alpha bound Z-DNA have been mostly limited to standard CG-repeat DNAs. Here, we have solved the crystal structures of three representative non-CG repeat DNAs, d(CACGTG)(2), d(CGTACG)(2) and d(CGGCCG)(2) complexed to hZ alpha(ADAR1) and compared those structures with that of hZ alpha(ADAR1)/d(CGCGCG)(2) and the Z alpha-free Z-DNAs. hZ alpha(ADAR1) bound to each of the three Z-DNAs showed a well conserved binding mode with very limited structural deviation irrespective of the DNA sequence, although varying numbers of residues were in contact with Z-DNA. Z-DNAs display less structural alterations in the Z alpha-bound state than in their free form, thereby suggesting that conformational diversities of Z-DNAs are restrained by the binding pocket of Z alpha. These data suggest that Z-DNAs are recognized by Z alpha through common conformational features regardless of the sequence and structural alterations.

The crystal structure of the second Z-DNA binding domain of human DAI (ZBP1) in complex with Z-DNA reveals an unusual binding mode to Z-DNA.

Mammalian DAI (DNA-dependent activator of IFN-regulatory factors), an activator of the innate immune response, senses cytosolic DNA by using 2 N-terminal Z-DNA binding domains (ZBDs) and a third putative DNA binding domain located next to the second ZBD. Compared with other previously known ZBDs, the second ZBD of human DAI (hZbeta(DAI)) shows significant variation in the sequence of the residues that are essential for DNA binding. In this article, the crystal structure of the hZbeta(DAI)/Z-DNA complex reveals that hZbeta(DAI) has a similar fold to that of other ZBDs, but adopts an unusual binding mode for recognition of Z-DNA. A residue in the first beta-strand rather than residues in the beta-loop contributes to DNA binding, and part of the (alpha3) recognition helix adopts a 3(10) helix conformation. The role of each residue that makes contact with DNA was confirmed by mutational analysis. The 2 ZBDs of DAI can together bind to DNA and both are necessary for full B-to-Z conversion. It is possible that binding 2 DAIs to 1 dsDNA brings about dimerization of DAI that might facilitate DNA-mediated innate immune activation.

Characterization of DNA-binding activity of Z alpha domains from poxviruses and the importance of the beta-wing regions in converting B-DNA to Z-DNA.

The E3L gene is essential for pathogenesis in vaccinia virus. The E3L gene product consists of an N-terminal Z alpha domain and a C-terminal double-stranded RNA (dsRNA) binding domain; the left-handed Z-DNA-binding activity of the Z alpha domain of E3L is required for viral pathogenicity in mice. E3L is highly conserved among poxviruses, including the smallpox virus, and it is likely that the orthologous Z alpha domains play similar roles. To better understand the biological function of E3L proteins, we have investigated the Z-DNA-binding behavior of five representative Z alpha domains from poxviruses. Using surface plasmon resonance (SPR), we have demonstrated that these viral Z alpha domains bind Z-DNA tightly. Ability of Z alpha(E3L) converting B-DNA to Z-DNA was measured by circular dichroism (CD). The extents to which these Z alphas can stabilize Z-DNA vary considerably. Mutational studies demonstrate that residues in the loop of the beta-wing play an important role in this stabilization. Notably the Z alpha domain of vaccinia E3L acquires ability to convert B-DNA to Z-DNA by mutating amino acid residues in this region. Differences in the host cells of the various poxviruses may require different abilities to stabilize Z-DNA; this may be reflected in the observed differences in behavior in these Zalpha proteins.

A left-handed RNA double helix bound by the Z alpha domain of the RNA-editing enzyme ADAR1.

The A form RNA double helix can be transformed to a left-handed helix, called Z-RNA. Currently, little is known about the detailed structural features of Z-RNA or its involvement in cellular processes. The discovery that certain interferon-response proteins have domains that can stabilize Z-RNA as well as Z-DNA opens the way for the study of Z-RNA. Here, we present the 2.25 A crystal structure of the Zalpha domain of the RNA-editing enzyme ADAR1 (double-stranded RNA adenosine deaminase) complexed to a dUr(CG)(3) duplex RNA. The Z-RNA helix is associated with a unique solvent pattern that distinguishes it from the otherwise similar conformation of Z-DNA. Based on the structure, we propose a model suggesting how differences in solvation lead to two types of Z-RNA structures. The interaction of Zalpha with Z-RNA demonstrates how the interferon-induced isoform of ADAR1 could be targeted toward selected dsRNAs containing purine-pyrimidine repeats, possibly of viral origin.

Characterization of Z-DNA as a nucleosome-boundary element in yeast Saccharomyces cerevisiae.

In this article, the effect of a d(CG) DNA dinucleotide repeat sequence on RNA polymerase II transcription is examined in yeast Saccharomyces cerevisiae. Our previous report shows that a d(CG)n dinucleotide repeat sequence located proximally upstream of the TATA box enhances transcription from a minimal CYC1 promoter in a manner that depends on its surrounding negative supercoiling. Here, we demonstrate that the d(CG)9 repeat sequence stimulates gene activity by forming a Z-DNA secondary structure. Furthermore, the extent of transcriptional enhancement by Z-DNA is promoter-specific and determined by its separation distance relative to the TATA box. The stimulatory effect exerted by promoter proximal Z-DNA is not affected by helical phasing relative to the TATA box, suggesting that Z-DNA effects transcription without interacting with the general transcription machinery by looping-out the intervening DNA. A nucleosome-scanning assay reveals that the d(CG)9 repeat sequence in the Z conformation blocks nucleosome formation, and it is found in the linker DNA with two flanking nucleosomes. This result suggests that Z-DNA formation proximally upstream of a promoter is sufficient to demarcate the boundaries of its neighboring nucleosomes, which produces transcriptionally favorable locations for the TATA box near the nucleosomal DNA-entry site and at dyad positions on the nucleosome. These findings suggest that Z-DNA formation in chromatin is a part of the "genomic code" for nucleosome positioning in vivo.

A peptide with alternating lysines can act as a highly specific Z-DNA binding domain.

Many nucleic acid binding proteins use short peptide sequences to provide specificity in recognizing their targets, which may be either a specific sequence or a conformation. Peptides containing alternating lysine have been shown to bind to poly(dG-d5meC) in the Z conformation, and stabilize the higher energy form [H. Takeuchi, N. Hanamura, H. Hayasaka and I. Harada (1991) FEBS Lett., 279, 253-255 and H. Takeuchi, N. Hanamura and I. Harada (1994) J. Mol. Biol., 236, 610-617.]. Here we report the construction of a Z-DNA specific binding protein, with the peptide KGKGKGK as a functional domain and a leucine zipper as a dimerization domain. The resultant protein, KGZIP, induces the Z conformation in poly(dG-d5meC) and binds to Z-DNA stabilized by bromination with high affinity and specificity. The binding of KGZIP is sufficient to convert poly(dG-d5meC) from the B to the Z form, as shown by circular dichroism. The sequence KGKGKGK is found in many proteins, although no functional role has been established. KGZIP also has potential for engineering other Z-DNA specific proteins for future studies of Z-DNA in vitro and in vivo.