The sarcoidosis genetic analysis (SAGA) study previously identified eight chromosomal regions with suggestive evidence for linkage to sarcoidosis susceptibility in African-American sib pairs. Since the clinical course of sarcoidosis is variable and likely under genetic control, we used the affected relative pair portion of the SAGA sample (n ¼ 344 pairs) to perform multipoint linkage analyses with covariates based on pulmonary and organ involvement phenotypes. Chest radiographic resolution was the pulmonary phenotype with the highest LOD (logarithm of the backward odds, or likelihood ratio) score of 5.11 at D1S3720 on chromosome 1p36 (P ¼ 4 Â 10 À5). In general, higher LOD scores were attained for covariates that modeled clustered organ system involvement rather than individual organ systems, with the cardiac/renal group having the highest LOD score of 6.65 at chromosome 18q22 (P ¼ 2 Â 10 À5 ). The highest LOD scores for the other three organ involvement groups of liver/spleen/bone marrow, neuro/lymph and ocular/skin/joint were 3.72 at 10p11 (P ¼ 0.0004), 5.16 at 7p22 (P ¼ 4 Â 10 À5 ) and 2.93 at 10q26 (P ¼ 0.001), respectively. Most of the phenotype linkages did not overlap with the regions previously found linked to susceptibility. Our results suggest that genes influencing clinical presentation of sarcoidosis in African Americans are likely to be different from those that underlie disease susceptibility.

Artificially intelligent computer systems are used extensively in medical sciences. Common applications include diagnosing patients, end-to-end drug discovery and development, improving communication between physician and patient, transcribing medical documents, such as prescriptions, and remotely treating patients. While computer systems often execute tasks more efficiently than humans, more recently, state-of-the-art computer algorithms have achieved accuracies which are at par with human experts in the field of medical sciences. Some speculate that it is only a matter of time before humans are completely replaced in certain roles within the medical sciences. The motivation of this article is to discuss the ways in which artificial intelligence is changing the landscape of medical science and to separate hype from reality.

Sarcoidosis, a systemic granulomatous disease, likely results from both environmental agents and genetic susceptibility. Sarcoidosis is more prevalent in women and, in the United States, African Americans are both more commonly and more severely affected than Caucasians. We report a follow up of the first genome scan for sarcoidosis susceptibility genes in African Americans. Both the genome scan and the present study comprise 229 African American nuclear families ascertained through two or more sibs with sarcoidosis. Regions studied included those which reached a significance in the genome scan of 0.01 (2p25, 5q11, 5q35, 9q34, 11p15 and 20q13), 0.05 (3p25 and 5p15-13) or which replicated previous findings (3p14-11). We performed genotyping with additional markers in the same families used in the genome scan. We examined multi-locus models for epistasis and performed model-based linkage analysis on subsets of the most linked families to characterize the underlying genetic model. The strongest signal was at marker D5S407 (P=0.005) on 5q11.2, using both full and half sibling pairs. Our results support, in an African American population, a sarcoidosis susceptibility gene on chromosome 5q11.2, and a gene protective for sarcoidosis on 5p15.2. These fine mapping results further prioritize the importance of candidate regions on chromosomes 2p25, 3p25, 5q35, 9q34, 11p15 and 20q13 for African Americans. Additionally, our results suggest joint action of the effects of putative genes on chromosome 3p14-11 and 5p15.2. We conclude that multiple susceptibility loci for sarcoidosis exist in African Americans and that some may have interdependent effects on disease pathogenesis.

No abstract

Construction of precise confidence sets of disease gene locations after initial identification of linked regions can improve the efficiency of the ensuing fine mapping effort. We took the confidence set inference, a framework proposed and implemented using the Mean test statistic (CSI-Mean) and improved the efficiency substantially by using a likelihood ratio test statistic (CSI-MLS). The CSI framework requires knowledge of some disease-model-related parameters. In the absence of prior knowledge of these parameters, a two-step procedure may be employed: 1) the parameters are estimated using a coarse map of markers; 2) CSI-Mean or CSI-MLS are applied to construct the confidence sets of the disease gene locations using a finer map of markers, assuming the estimates from Step 1 for the required parameters. In this article we show that the advantages of CSI-MLS over CSI-Mean, previously demonstrated when the required parameters are known, are preserved in this two-step procedure, using both the simulated and real data contributed to Problems 2 and 3 of Genetic Analysis Workshop 15. In addition, our result suggests that microsatellite data, when available, should be used in Step 1. Also explored in detail is the effect of the absence of parental genotypes on the performance of CSI-MLS.

The simultaneous testing of a large number of hypotheses in a genome scan, using individual thresholds for significance, inherently leads to inflated genome-wide false positive rates. There exist various approaches to approximating the correct genomewide p-values under various assumptions, either by way of asymptotics or simulations. We explore a philosophically different criterion, recently proposed in the literature, which controls the false discovery rate. The test statistics are assumed to arise from a mixture of distributions under the null and non-null hypotheses. We fit the mixture distribution using both a nonparametric approach and commingling analysis, and then apply the local false discovery rate to select cut-off points for regions to be declared interesting. Another criterion, the minimum total error, is also explored. Both criteria seem to be sensible alternatives to controlling the classical type I and type II error rates.

Genomewide linkage studies are tending toward the use of single-nucleotide polymorphisms (SNPs) as the markers of choice. However, linkage disequilibrium (LD) between tightly linked SNPs violates the fundamental assumption of linkage equilibrium (LE) between markers that underlies most multipoint calculation algorithms currently available, and this leads to inflated affected-relative-pair allele-sharing statistics when founders' multilocus genotypes are unknown. In this study, we investigate the impact that the degree of LD, marker allele frequency, and association type have on estimating the probabilities of sharing alleles identical by descent in multipoint calculations and hence on type I error rates of different sib-pair linkage approaches that assume LE. We show that marker-marker LD does not inflate type I error rates of affected sib pair (ASP) statistics in the whole parameter space, and that, in any case, discordant sib pairs (DSPs) can be used to control for marker-marker LD in ASPs. We advocate the ASP/DSP design with appropriate sib-pair statistics that test the difference in allele sharing between ASPs and DSPs.

We introduce models for saliency prediction for mobile user interfaces. A mobile interface may include elements like buttons, text, etc. in addition to natural images which enable performing a variety of tasks. Saliency in natural images is a well studied area. However, given the difference in what constitutes a mobile interface, and the usage context of these devices, we postulate that saliency prediction for mobile interface images requires a fresh approach. Mobile interface design involves operating on elements, the building blocks of the interface. We first collected eye-gaze data from mobile devices for free viewing task. Using this data, we develop a novel autoencoder based multi-scale deep learning model that provides saliency prediction at the mobile interface element level. Compared to saliency prediction approaches developed for natural images, we show that our approach performs significantly better on a range of established metrics.