Mass spectrometry is a powerful tool used in proteomic analysis. By combining with different isotope labeling methods, protein quantification at the entire proteome scale can be achieved. In this dissertation, I employed an LC-MS/MS coupled with stable isotope labeling by amino acid in cell culture (SILAC) strategy, to explore the molecular mechanisms of action of anti-cancer drugs and an environmental toxicant, arsenic.

In Chapters 2 and 3, I performed a quantitative assessment of global proteome changes upon treatment with two anti-cancer drugs, both are nucleoside analogues, i.e. 5-aza-2'-deoxytidine (5-Aza-CdR) and 6-thioguanine (SG). We found, for the first time, that 5-Aza-CdR exerts its cytotoxic effects in leukemia and melanoma cells through epigenetic reactivation of DPP4 gene and the resultant inhibition of cholesterol biosynthesis. In addition, proteome-wide analysis suggested that SG may exert its cytotoxic effect by inducing mitochondrial dysfunction and reactive oxygen species formation in acute lymphoblastic leukemia cells.

In Chapters 4 and 5, we utilized the same MS-based quantitative proteomic method, to study the mechanisms underlying trivalent arsenic-induced carcinogenic effect. Our results confirmed the previous findings and provided a more complete picture about the biological pathways that are altered upon arsenite treatment in human skin fibroblast cells. To further understand how arsenic affects RING-finger proteins that are known to be important in DNA damage repair, we performed in vitro and in vivo arsenite binding experiments. Our studies revealed, for the first time, that arsenite may exert its carcinogenic effect by targeting cysteine residues in the RING finger domains of the RNF20-RNF40 histone E3 ubiquitin ligase, thereby altering histone epigenetic mark and compromising DNA DSB repair.

Classifier combination is an effective and popular method to improve the predictive performance of classification models. It has been employed in various fields, including pattern recognition and biometrics. This thesis proposes a novel classifier combination method based on the uniformness, a statistical measurement of the predicted probabilities of base classifiers. By choosing different measurement functions, three combination schemes are explored. The new method is designed to achieve improved accuracy and efficiency on the classification. It is tested on a real multi-class classification problem of plant species using leaf image features, which proves the advantage and robustness of this combination method.

The dissertation has three chapters. In the first chapter, I estimate consumer search cost with purchase outcome data. I analyze a model where consumers search sequentially for the best option based on advertised price and partial product information. I verify the model's ability to recover the structural parameters by a numerical experiment. Then using Nielsen Consumer Panel Data, I estimate the product-specific search costs for the top five brands in 32oz refrigerated yogurt market. I find that after controlling for the prices, the private label brand has the lowest search cost. Counterfactual analysis shows that eliminating the search cost increases overall purchase and decreases the price sensitivity. Incorrectly ignoring search frictions leads to an overestimation of own-price elasticity.

In the second chapter, we state conditions under which choice data suffices to identify preferences when consumers may not be fully informed about attributes of goods. Our approach can be used to test for full information, to forecast how consumers will respond to information, and to conduct welfare analysis when consumers are imperfectly informed. In a lab experiment, we successfully forecast the response to new information when consumers engage in costly search. In data from Expedia, our method identifies which attribute was not immediately visible to consumers in search results, and we then use the model to compute the value of additional information.

In the third chapter, we study consumers' variety-seeking preferences and explore their implications for targeted marketing using proprietary data from a food delivery platform. We document that a substantial fraction of consumers have variety-seeking preferences. Consumers, on average, are willing to pay 20% more to switch to a different seller. In the counterfactual analysis, we find that optimizing rankings by taking into account variety-seeking preferences increases revenue, consumer welfare, and purchase probability. Furthermore, we find that consumers' variety-seeking preferences soften price competition. Optimal targeted pricing implies an increase in prices for rival sellers' consumers and a decrease in prices for the sellers' own consumers.