Search

Safe deployment of deep learning systems in critical real world applications requires models to make very few mistakes, and only under predictable circumstances. In this work, we address this problem using an abstaining classifier that is tuned to have $>$95% accuracy, and then identify the determinants of abstention using LIME. Essentially, we are training our model to learn the attributes of pathology reports that are likely to lead to incorrect classifications, albeit at the cost of reduced sensitivity. We demonstrate an abstaining classifier in a multitask setting for classifying cancer pathology reports from the NCI SEER cancer registries on six tasks of interest. For these tasks, we reduce the classification error rate by factors of 2--5 by abstaining on 25--45% of the reports. For the specific task of classifying cancer site, we are able to identify metastasis, reports involving lymph nodes, and discussion of multiple cancer sites as responsible for many of the classification mistakes, and observe that the extent and types of mistakes vary systematically with cancer site (e.g., breast, lung, and prostate). When combining across three of the tasks, our model classifies 50% of the reports with an accuracy greater than 95% for three of the six tasks\edit, and greater than 85% for all six tasks on the retained samples. Furthermore, we show that LIME provides a better determinant of classification than measures of word occurrence alone. By combining a deep abstaining classifier with feature identification using LIME, we are able to identify concepts responsible for both correctness and abstention when classifying cancer sites from pathology reports. The improvement of LIME over keyword searches is statistically significant, presumably because words are assessed in context and have been identified as a local determinant of classification.

Using a sparsity inducing penalty in artificial neural networks (ANNs) avoids over-fitting, especially in situations where noise is high and the training set is small in comparison to the number of features. For linear models, such an approach provably also recovers the important features with high probability in regimes for a well-chosen penalty parameter. The typical way of setting the penalty parameter is by splitting the data set and performing the cross-validation, which is (1) computationally expensive and (2) not desirable when the data set is already small to be further split (for example, whole-genome sequence data). In this study, we establish the theoretical foundation to select the penalty parameter without cross-validation based on bounding with a high probability the infinite norm of the gradient of the loss function at zero under the zero-feature assumption. Our approach is a generalization of the universal threshold of Donoho and Johnstone (1994) to nonlinear ANN learning. We perform a set of comprehensive Monte Carlo simulations on a simple model, and the numerical results show the effectiveness of the proposed approach., Comment: 8 pages, 2 figures

Summary The dense arrangement of N-glycans masking antigenic surfaces on the HIV-1 envelope (Env) protein acts as a shield from the adaptive immune system. The molecular complexity of glycan modifications and their inherent dynamic heterogeneity on a protein surface make experimental studies of glycoprotein structures a challenge. Here we have integrated a high-throughput atomistic modeling with graph-theory based method to capture the native glycan shield topological network and identify concerted behavior of these glycans. This is the first time that a complete computational model of an HIV-1 Env trimeric SOSIP structure has been generated with a native glycosylation pattern including both oligomannose and complex glycans, thus obtaining results which are immunologically more relevant. Important global and local feature differences due to the native-like glycosylation pattern have been identified, that stem from the charged sialic acid tips, fucose rings at the base, and different branching patterns of the complex glycans. Analyses of network attributes have aided in detailed description of the shield in a biological context. We have also derived a measure to quantify the shielding effect based on the number of glycan heavy atoms encountered over the antigenic protein surface that can define regions of relative vulnerability and resilience on the shield, and can be harnessed for potential immunogen design.

Gram-negative bacteria are a serious health concern due to the strong multidrug resistance that they display, partly due to the presence of a permeability barrier comprising two membranes with active efflux. New approaches are urgently needed to design antibiotics effective against these pathogens. In this work, we present a novel topological fragment-based approach ("Hunting Fragments Of X" or "Hunting FOX") to rationally "hunt for" chemical fragments that promote compound ability to permeate the outer membrane. Our approach generalizes to other drug design applications. We measure minimum inhibitory concentrations of compounds in two strains of Pseudomonas aeruginosa with variable permeability barriers and use them as an input to the Hunting FOX algorithm to identify molecular fragments responsible for enhanced outer membrane permeation properties and candidate molecules from an external library that demonstrate good permeation ability. Overall, we present proof of concept for a novel method that is expected to be valuable for rational design of hybrid drugs., 15 pages, 5 figures, 4 pages of supporting information, 3 supporting figures, 2 ancillary files

We analyze the component evolution in inhomogeneous random intersection graphs when the average degree is close to 1. As the average degree increases, the size of the largest component in the random intersection graph goes through a phase transition. We give bounds on the size of the largest components before and after this transition. We also prove that the largest component after the transition is unique. These results are similar to the phase transition in Erd\H{o}s-R\'enyi random graphs; one notable difference is that the jump in the size of the largest component varies in size depending on the parameters of the random intersection graph., Comment: 18 pages

Random intersection graphs (RIGs) are an important random structure with applications in social networks, epidemic networks, blog readership, and wireless sensor networks. RIGs can be interpreted as a model for large randomly formed non-metric data sets. We analyze the component evolution in general RIGs, and give conditions on existence and uniqueness of the giant component. Our techniques generalize existing methods for analysis of component evolution: we analyze survival and extinction properties of a dependent, inhomogeneous Galton-Watson branching process on general RIGs. Our analysis relies on bounding the branching processes and inherits the fundamental concepts of the study of component evolution in Erd\H{o}s-R\'enyi graphs. The major challenge comes from the underlying structure of RIGs, which involves its both the set of nodes and the set of attributes, as well as the set of different probabilities among the nodes and attributes.