Reliance on metrics is a fundamental challenge for AI.

Through a series of case studies, we review how the unthinking pursuit of metric optimization can lead to real-world harms, including recommendation systems promoting radicalization, well-loved teachers fired by an algorithm, and essay grading software that rewards sophisticated garbage. The metrics used are often proxies for underlying, unmeasurable quantities (e.g., "watch time" of a video as a proxy for "user satisfaction"). We propose an evidence-based framework to mitigate such harms by (1) using a slate of metrics to get a fuller and more nuanced picture; (2) conducting external algorithmic audits; (3) combining metrics with qualitative accounts; and (4) involving a range of stakeholders, including those who will be most impacted.

Contrasting effects of allosteric and orthosteric agonists on m1 muscarinic acetylcholine receptor internalization and down-regulation.

A new class of subtype-selective muscarinic acetylcholine (mACh) receptor agonist that activates the receptor through interaction at a site distinct from the orthosteric acetylcholine binding site has been reported recently. Here, we have compared the effects of orthosteric (oxotremorine-M, arecoline, pilocarpine) and allosteric [4-n-butyl-1-[4-(2-methylphenyl)-4-oxo-1-butyl] piperidine (AC-42); 1-[3-(4-butyl-1-piperidinyl)propyl]-3,4-dihydro-2(1H)-quinolinone (77-LH-28-1)] agonists on M(1) mACh receptor internalization and down-regulation, as well as functional coupling in a Chinese hamster ovary (CHO) cell line. In contrast to full and partial orthosteric agonists, which cause significant receptor internalization and down-regulation, prolonged exposure to AC-42 did not significantly alter either cell-surface or total cellular M(1) mACh receptor expression. 77-LH-28-1, an AC-42 homolog, did cause some receptor internalization, but not down-regulation. The presence of atropine completely prevented the orthosteric agonist-induced adaptive changes in receptor populations; however, in contrast, the copresence of atropine and AC-42 significantly increased both cell-surface receptor and total M(1) mACh receptor expression. Maximal phosphoinositide hydrolysis responses to the partial agonist arecoline were similar in CHO-M(1) cells pretreated for 24 h with either AC-42 or vehicle; in contrast, these responses were markedly reduced when cells were pretreated with oxotremorine-M or pilocarpine. These data indicate that, whereas AC-42 binding to the M(1) mACh receptor can initiate signal transduction, the AC-42-liganded receptor is resistant to the usual mechanisms regulating receptor internalization and down-regulation. In addition, our data suggest unusual interactions between allosteric agonists and orthosteric antagonists to regulate cell-surface and total cellular receptor expression.

G protein coupling and signaling pathway activation by m1 muscarinic acetylcholine receptor orthosteric and allosteric agonists.

The M(1) muscarinic acetylcholine (mACh) receptor is among a growing number of G protein-coupled receptors that are able to activate multiple signaling cascades. AC-42 (4-n-butyl-1-[4-(2-methylphenyl)-4-oxo-1-butyl] piperidine) is an allosteric agonist that can selectively activate the M(1) mACh receptor in the absence of an orthosteric ligand. Allosteric agonists have the potential to stabilize unique receptor conformations, which may in turn cause differential activation of signal transduction pathways. In the present study, we have investigated the signaling pathways activated by AC-42, its analog 77-LH-28-1 (1-[3-(4-butyl-1-piperidinyl)propyl]-3,4-dihydro-2(1H)-quinolinone), and a range of orthosteric muscarinic agonists [oxotremorine-M (oxo-M), arecoline, and pilocarpine] in Chinese hamster ovary cells recombinantly expressing the human M(1) mACh receptor. Each agonist was able to activate Galpha(q/11)-dependent signaling, as demonstrated by an increase in guanosine 5'-O-(3-thiotriphosphate) ([(35)S]GTPgammaS) binding to Galpha(q/11) proteins and total [(3)H]inositol phosphate accumulation assays in intact cells. All three orthosteric agonists caused significant enhancements in [(35)S]GTPgammaS binding to Galpha(i1/2) subunits over basal; however, neither allosteric ligand produced a significant response. In contrast, both orthosteric and allosteric agonists are able to couple to the Galpha(s)/cAMP pathway, enhancing forskolin-stimulated cAMP accumulation. These data provide support for the concept that allosteric and orthosteric mACh receptor agonists both stabilize receptor conformations associated with Galpha(q/11)- and Galpha(s)-dependent signaling; however, AC-42 and 77-LH-28-1, unlike oxo-M, arecoline, and pilocarpine, do not seem to promote M(1) mACh receptor-Galpha(i1/2) coupling, suggesting that allosteric agonists have the potential to activate distinct subsets of downstream effectors.

A mathematical model of glutathione metabolism.

BACKGROUND: Glutathione (GSH) plays an important role in anti-oxidant defense and detoxification reactions. It is primarily synthesized in the liver by the transsulfuration pathway and exported to provide precursors for in situ GSH synthesis by other tissues. Deficits in glutathione have been implicated in aging and a host of diseases including Alzheimer's disease, Parkinson's disease, cardiovascular disease, cancer, Down syndrome and autism. APPROACH: We explore the properties of glutathione metabolism in the liver by experimenting with a mathematical model of one-carbon metabolism, the transsulfuration pathway, and glutathione synthesis, transport, and breakdown. The model is based on known properties of the enzymes and the regulation of those enzymes by oxidative stress. We explore the half-life of glutathione, the regulation of glutathione synthesis, and its sensitivity to fluctuations in amino acid input. We use the model to simulate the metabolic profiles previously observed in Down syndrome and autism and compare the model results to clinical data. CONCLUSION: We show that the glutathione pools in hepatic cells and in the blood are quite insensitive to fluctuations in amino acid input and offer an explanation based on model predictions. In contrast, we show that hepatic glutathione pools are highly sensitive to the level of oxidative stress. The model shows that overexpression of genes on chromosome 21 and an increase in oxidative stress can explain the metabolic profile of Down syndrome. The model also correctly simulates the metabolic profile of autism when oxidative stress is substantially increased and the adenosine concentration is raised. Finally, we discuss how individual variation arises and its consequences for one-carbon and glutathione metabolism.