We introduce and illustrate a new approach to measuring and mitigating unintended bias in machine learning models. Our definition of unintended bias is parameterized by a test set and a subset of input features. We illustrate how this can be used to evaluate text classifiers using a synthetic test set and a public corpus of comments annotated for toxicity from Wikipedia Talk pages. We also demonstrate how imbalances in training data can lead to unintended bias in the resulting models, and therefore potentially unfair applications. We use a set of common demographic identity terms as the subset of input features on which we measure bias. This technique permits analysis in the common scenario where demographic information on authors and readers is unavailable, so that bias mitigation must focus on the content of the text itself. The mitigation method we introduce is an unsupervised approach based on balancing the training dataset. We demonstrate that this approach reduces the unintended bias without compromising overall model quality.

Obesity and insulin resistance are associated with deposition of triglycerides in tissues other than adipose tissue. Previously, we showed that a missense mutation (I148M) in PNPLA3 (patatin-like phospholipase domain-containing 3 protein) is associated with increased hepatic triglyceride content in humans. Here we examined the effect of the I148M substitution on the enzymatic activity and cellular location of PNPLA3. Structural modeling predicted that the substitution of methionine for isoleucine at residue 148 would restrict access of substrate to the catalytic serine at residue 47. In vitro assays using recombinant PNPLA3 partially purified from Sf9 cells confirmed that the wild type enzyme hydrolyzes emulsified triglyceride and that the I148M substitution abolishes this activity. Expression of PNPLA3-I148M, but not wild type PNPLA3, in cultured hepatocytes or in the livers of mice increased cellular triglyceride content. Cell fractionation studies revealed that ∼90% of wild type PNPLA3 partitioned between membranes and lipid droplets; substitution of isoleucine for methionine at position 148 did not alter the subcellular distribution of the protein. These data are consistent with PNPLA3-I148M promoting triglyceride accumulation by limiting triglyceride hydrolysis.

The upsurge in prevalence of obesity has spawned an epidemic of nonalcoholic fatty liver disease (NAFLD). Previously, we identified a sequence variant (I148M) in patatin-like phospholipase domaincontaining protein 3 (PNPLA3) that confers susceptibility to both hepatic triglyceride (TG) deposition and liver injury. To glean insights into the biological role of PNPLA3, we examined the molecular mechanisms by which nutrient status controls hepatic expression of PNPLA3. PNPLA3 mRNA levels, which were low in fasting animals, increased ∼90-fold with carbohydrate feeding. The increase was mimicked by treatment with a liver X receptor (LXR) agonist and required the transcription factor SREBP-1c. The site of SREBP-1c binding was mapped to intron 1 of Pnpla3 using chromatin immunoprecipitation and electrophoretic mobility shift assays. SREBP-1c also promotes fatty acid synthesis by activating several genes encoding enzymes in the biosynthetic pathway. Addition of fatty acids (C16:0, C18:1, and C18:2) to the medium of cultured hepatocytes (HuH-7) increased PNPLA3 protein mass without altering mRNA levels. The posttranslational increase in PNPLA3 levels persisted after blocking TG synthesis with triascin C. Oleate (400 μM) treatment prolonged the half-life of PNPLA3 from 2.4 to 6.7 h. These findings are consistent with nutritional control of PNPLA3 being effected by a feed-forward loop; SREBP-1c promotes accumulation of PNPLA3 directly by activating Pnpla3 transcription and indirectly by inhibiting PNPLA3 degradation through the stimulation of fatty acid synthesis.adiponutrin | SREBP-1c | triglycerides | hepatic steatosis

A sequence polymorphism (rs738409, I148M) in patatin-like phospholipid domain containing protein 3 (PNPLA3), is strongly associated with nonalcoholic fatty liver disease (NAFLD), but the mechanistic basis for this association remains enigmatic. Neither ablation nor overexpression of wild-type PNPLA3 affects liver fat content in mice, whereas hepatic overexpression of the human 148M transgene causes steatosis. To determine whether the 148M allele causes fat accumulation in the liver when expressed at physiological levels, we introduced a methionine codon at position 148 of the mouse Pnpla3 gene. Knockin mice had normal levels of hepatic fat on a chow diet, but when challenged with a high-sucrose diet their liver fat levels increased 2–3 fold compared to wild-type littermates without any associated changes in glucose homeostasis. The increased liver fat in the knockin mice was accompanied by a 40-fold increase in PNPLA3 on hepatic lipid droplets, with no increase in hepatic PNPLA3 mRNA. Similar results were obtained when the catalytic dyad of PNPLA3 was inactivated by substituting the catalytic serine with alanine (S47A). Conclusion These data provide the first direct evidence that physiological expression of PNPLA3 148M variant causes NAFLD, and that the accumulation of catalytically inactive PNPLA3 on the surfaces of lipid droplets is associated with the accumulation of TG in the liver.

Fsp27, a member of the Cide family proteins, was shown to localize to lipid droplet and promote lipid storage in adipocytes. We aimed to understand the biological role of Fsp27 in regulating adipose tissue differentiation, insulin sensitivity and energy balance. Fsp27 −/− mice and Fsp27/lep double deficient mice were generated and we examined the adiposity, whole body metabolism, BAT and WAT morphology, insulin sensitivity, mitochondrial activity, and gene expression changes in these mouse strains. Furthermore, we isolated mouse embryonic fibroblasts (MEFs) from wildtype and Fsp27 −/− mice, followed by their differentiation into adipocytes in vitro. We found that Fsp27 is expressed in both brown adipose tissue (BAT) and white adipose tissue (WAT) and its levels were significantly elevated in the WAT and liver of leptin-deficient ob/ob mice. Fsp27 −/− mice had increased energy expenditure, lower levels of plasma triglycerides and free fatty acids. Furthermore, Fsp27 −/− and Fsp27/lep double-deficient mice are resistant to diet-induced obesity and display increased insulin sensitivity. Moreover, white adipocytes in Fsp27 −/− mice have reduced triglycerides accumulation and smaller lipid droplets, while levels of mitochondrial proteins, mitochondrial size and activity are dramatically increased. We further demonstrated that BAT-specific genes and key metabolic controlling factors such as FoxC2, PPAR and PGC1α were all markedly upregulated. In contrast, factors inhibiting BAT differentiation such as Rb, p107 and RIP140 were down-regulated in the WAT of Fsp27 −/− mice. Remarkably, Fsp27 −/− MEFs differentiated in vitro show many brown adipocyte characteristics in the presence of the thyroid hormone triiodothyronine (T3). Our data thus suggest that Fsp27 acts as a novel regulator in vivo to control WAT identity, mitochondrial activity and insulin sensitivity.