Deep learning methods, including deep multiple instance learning methods, have been criticized for their limited ability to incorporate domain knowledge. A reason that knowledge incorporation is challenging in deep learning is that the models usually lack a mapping between their model components and the entities of the domain, making it a non-trivial task to incorporate probabilistic prior information. In this work, we show that such a mapping between domain entities and model components can be defined for a multiple instance learning setting and propose a framework DeeMILIP that encompasses multiple strategies to exploit this mapping for prior knowledge incorporation. We motivate and formalize these strategies from a probabilistic perspective. Experiments on an immune-based diagnostics case show that our proposed strategies allow to learn generalizable models even in settings with weak signals, limited dataset size, and limited compute.

MOTIVATION: Many high-throughput methods produce sets of genomic regions as one of their main outputs. Scientists often use genomic colocalization analysis to interpret such region sets, for example to identify interesting enrichments and to understand the interplay between the underlying biological processes. Although widely used, there is little standardization in how these analyses are performed. Different practices can substantially affect the conclusions of colocalization analyses. RESULTS: Here, we describe the different approaches and provide recommendations for performing genomic colocalization analysis, while also discussing common methodological challenges that may influence the conclusions. As illustrated by concrete example cases, careful attention to analysis details is needed in order to meet these challenges and to obtain a robust and biologically meaningful interpretation of genomic region set data. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

BACKGROUND: The current versions of reference genome assemblies still contain gaps represented by stretches of Ns. Since high throughput sequencing reads cannot be mapped to those gap regions, the regions are depleted of experimental data. Moreover, several technology platforms assay a targeted portion of the genomic sequence, meaning that regions from the unassayed portion of the genomic sequence cannot be detected in those experiments. We here refer to all such regions as inaccessible regions, and hypothesize that ignoring these regions in the null model may increase false findings in statistical testing of colocalization of genomic features. RESULTS: Our explorative analyses confirm that the genomic regions in public genomic tracks intersect very little with assembly gaps of human reference genomes (hg19 and hg38). The little intersection was observed only at the beginning and end portions of the gap regions. Further, we simulated a set of synthetic tracks by matching the properties of real genomic tracks in a way that nullified any true association between them. This allowed us to test our hypothesis that not avoiding inaccessible regions (as represented by assembly gaps) in the null model would result in spurious inflation of statistical significance. We contrasted the distributions of test statistics and p-values of Monte Carlo-based permutation tests that either avoided or did not avoid assembly gaps in the null model when testing colocalization between a pair of tracks. We observed that the statistical tests that did not account for assembly gaps in the null model resulted in a distribution of the test statistic that is shifted to the right and a distribution of p-values that is shifted to the left (indicating inflated significance). We observed a similar level of inflated significance in hg19 and hg38, despite assembly gaps covering a smaller proportion of the latter reference genome. CONCLUSION: We provide empirical evidence demonstrating that inaccessible regions, even when covering only a few percentages of the genome, can lead to a substantial amount of false findings if not accounted for in statistical colocalization analysis.