We study the problem of generating counterfactual text for a classifier as a means for understanding and debugging classification. Given a textual input and a classification model, we aim to minimally alter the text to change the model's prediction. White-box approaches have been successfully applied to similar problems in vision where one can directly optimize the continuous input. Optimization-based approaches become difficult in the language domain due to the discrete nature of text. We bypass this issue by directly optimizing in the latent space and leveraging a language model to generate candidate modifications from optimized latent representations. We additionally use Shapley values to estimate the combinatoric effect of multiple changes. We then use these estimates to guide a beam search for the final counterfactual text. We achieve favorable performance compared to recent whitebox and black-box baselines using human and automatic evaluations. Ablation studies show that both latent optimization and the use of Shapley values improve success rate and the quality of the generated counterfactuals.

The biological half-life (t1/2) of methylmercury (MeHg) shows considerable individual variability (t1/2 < 30 to > 120 days), highlighting the importance of mechanisms controlling MeHg metabolism and elimination. Building on a prior physiologically based pharmacokinetic (PBPK) model, we elucidate parameters that have the greatest influence on variability of MeHg t1/2 in the human body. Employing a dataset of parameters for mean organ volumes and blood flow rates appropriate for man and woman (25–35 years) and child (4 − 6 years), we demonstrate model fitness by simulating data from our prior controlled study of MeHg elimination in people. Model predictions give MeHg t1/2 of 46.9, 38.9, and 31.5 days and steady-state blood MeHg of 2.6, 2.6, and 2.3 µg/l in man, woman, and child, respectively, subsequent to a weekly dose of 0.7 µg/kg body weight. The major routes of elimination are biotransformation to inorganic Hg in the gut lumen (73% in adults, 61% in child) and loss of MeHg via excretion within growing hair (13% in adults, 24% in child). Local and global sensitivity analyses of model parameters reveal that variation in biotransformation rate in the gut lumen, and rates of transport between gut lumen and gut tissue, have the greatest influence on MeHg t1/2. Volume and partition coefficients for skeletal muscle (SM) and gut tissue also show significant sensitivity affecting model output of MeHg t1/2. Our results emphasize the role of gut microbiota in MeHg biotransformation, transport kinetics at the level of the gut, and SM mass as moderators of MeHg kinetics in the human body.

A growing body of research has established that the microbiome can mediate the dynamics and functional capacities of diverse biological systems. Yet, we understand little about what governs the response of these microbial communities to host or environmental changes.

We use open source human gut microbiome data to learn a microbial “language” model by adapting techniques from Natural Language Processing (NLP). Our microbial “language” model is trained in a self-supervised fashion (i.e., without additional external labels) to capture the interactions among different microbial species and the common compositional patterns in microbial communities. The learned model produces contextualized taxa representations that allow a single bacteria species to be represented differently according to the specific microbial environment it appears in. The model further provides a sample representation by collectively interpreting different bacteria species in the sample and their interactions as a whole. We show that, compared to baseline representations, our sample representation consistently leads to improved performance for multiple prediction tasks including predicting Irritable Bowel Disease (IBD) and diet patterns. Coupled with a simple ensemble strategy, it produces a highly robust IBD prediction model that generalizes well to microbiome data independently collected from different populations with substantial distribution shift. We visualize the contextualized taxa representations and find that they exhibit meaningful phylum-level structure, despite never exposing the model to such a signal. Finally, we apply an interpretation method to highlight bacterial species that are particularly influential in driving our model’s predictions for IBD.

No abstract