Venetia Pliatsika

Shapley values, derived from cooperative game theory, are widely used in machine learning to design, audit, and recommend recourse for complex models. A wide va- riety of implementations has resulted in explanation multiplicity, the phenomenon where multiple contradictory Shapley value explanations can be derived for the same model and data instance. In this position paper, we argue that an explanation is mathematically underspecified without a formal definition of context of use and a corresponding disclosure of technical assumptions. To enable both goals, we in- troduce a generalized definition of Shapley values and use it to reframe multiplicity as a parametrization problem. Using this unified parametrization, we propose a roadmap for the entire explanation lifecycle, calling focus on two primary research directions. First, we posit that context of use is a formal input to the explanatory process. As such, we require rigorous guidelines to map stakeholder questions to specific mathematical parameters. Second, we argue that an explanation is incomplete without a meta-explanation: a structured disclosure of the underlying assumptions and implementation choices used in its construction. As a step towards this second direction, we propose the Shapley Value Explainability Card. Finally, we demonstrate through several examples how each parameter choice implies specific mathematical assumptions about the data or the explanation’s intended use, and provide a detailed call to action towards a more transparent and context-aware explanation lifecycle

Algorithmic decisions in critical domains such as hiring, college admissions, and lending are often based on rankings. Given the impact of these decisions on individuals, organizations, and population groups, it is essential to understand them - to help individuals improve their ranking position, design better ranking procedures, and ensure legal compliance. In this paper, we argue that explainability methods for classification and regression, such as SHAP, are insufficient for ranking tasks, and present ShaRP - Shapley Values for Rankings and Preferences - a framework that explains the contributions of features to various aspects of a ranked outcome. ShaRP computes feature contributions for various ranking-specific profit functions, such as rank and top-k, and also includes a novel Shapley value-based method for explaining pairwise preference outcomes. We provide a flexible implementation of ShaRP, capable of efficiently and comprehensively explaining ranked and pairwise outcomes over tabular data, in score-based ranking and learning-to-rank tasks. Finally, we develop a comprehensive evaluation methodology for ranking explainability methods, showing through qualitative, quantitative, and usability studies that our rank-aware QoIs offer complementary insights, scale effectively, and help users interpret ranked outcomes in practice.

MINTbase is a repository that comprises nuclear and mitochondrial tRNA-derived fragments (‘tRFs’) found in multiple human tissues. The original version of MINTbase comprised tRFs obtained from 768 transcriptomic datasets. We used our deterministic and exhaustive tRF mining pipeline to process all of The Cancer Genome Atlas datasets (TCGA). We identified 23 413 tRFs with abundance of ≥ 1.0 reads-per-million (RPM). To facilitate further studies of tRFs by the community, we just released version 2.0 of MINTbase that contains information about 26 531 distinct human tRFs from 11 719 human datasets as of October 2017. Key new elements include: the ability to filter tRFs on-the-fly by minimum abundance thresholding; the ability to filter tRFs by tissue keywords; easy access to information about a tRF’s maximum abundance and the datasets that contain it; the ability to generate relative abundance plots for tRFs across cancer types and convert them into embeddable figures; MODOMICS information about modifications of the parental tRNA, etc. Version 2.0 of MINTbase contains 15x more datasets and nearly 4x more distinct tRFs than the original version, yet continues to offer fast, interactive access to its contents. Version 2.0 is available freely at http://cm.jefferson.edu/MINTbase/.

Background: CRISPR/Cas (Clustered Regularly Interspaced Short Palindromic Repeats/CRISPR associated nucleases) is a powerful component of the prokaryotic immune system that has been adapted for targeted genetic engineering in higher organisms. A key element of CRISPR/Cas is the “guide” RNA (gRNA) that is  20 nucleotides (nts) in length and designed to be complementary to the intended target site. An integral requirement of the CRISPR/Cas system is that the target site be followed by a protospacer adjacent motif (PAM). Care needs to be exercised during gRNA design to avoid unintended (“off-target”) interactions. Results: We designed and implemented the Off-Spotter algorithm to assist with the design of optimal gRNAs. When presented with a candidate gRNA sequence and a PAM, Off-Spotter quickly and exhaustively identifies all genomic sites that satisfy the PAM constraint and are identical or nearly-identical to the provided gRNA. Off-Spotter achieves its extreme performance through purely algorithmic means and not through hardware accelerators such as graphical processing units (GPUs). Off-Spotter also allows the user to identify on-the-fly how many and which nucleotides of the gRNA comprise the “seed”. Off-Spotter’s output includes a histogram showing the number of potential off-targets as a function of the number of mismatches. The output also includes for each potential off-target the site’s genomic location, a human genome browser hyperlink to the corresponding location, genomic annotation in the vicinity of the off-target, GC content, etc. Conclusion: Off-Spotter is very fast and flexible and can help in the design of optimal gRNAs by providing several PAM choices, a run-time definition of the seed and of the allowed number of mismatches, and a flexible output interface that allows sorting of the results, optional viewing/hiding of columns, etc. A key element of Off-Spotter is that it does not have a rigid definition of the seed: instead, the user can declare both the seed’s location and extent on-the-fly. We expect that this flexibility in combination with Off-Spotter’s speed and richly annotated output will enable experimenters to interactively and quickly explore different scenarios and gRNA possibilities. Reviewed: This article was reviewed by Dr Eugene Koonin and Dr Frank Eisenhaber.