Showing papers by "Michael Q. Zhang published in 2023"

Boosted Prompt Ensembles for Large Language Models

Abstract: Methods such as chain-of-thought prompting and self-consistency have pushed the frontier of language model reasoning performance with no additional training. To further improve performance, we propose a prompt ensembling method for large language models, which uses a small dataset to construct a set of few shot prompts that together comprise a ``boosted prompt ensemble''. The few shot examples for each prompt are chosen in a stepwise fashion to be ``hard'' examples on which the previous step's ensemble is uncertain. We show that this outperforms single-prompt output-space ensembles and bagged prompt-space ensembles on the GSM8k and AQuA datasets, among others. We propose both train-time and test-time versions of boosted prompting that use different levels of available annotation and conduct a detailed empirical study of our algorithm.

Classifying Course Discussion Board Questions using LLMs

Abstract: Large language models (LLMs) can be used to answer student questions on course discussion boards, but there is a risk of LLMs answering questions they are unable to address. We propose and evaluate an LLM-based system that classifies student questions into one of four types: conceptual, homework, logistics, and not answerable. We then prompt an LLM using a type-specific prompt. Using GPT-3, we achieve 81% classification accuracy across the four categories. Furthermore, we achieve 93% accuracy on classifying not answerable questions. This indicates that our system effectively ignores questions that it cannot address.

Enterococcus faecalis Strains with Compromised CRISPR-Cas Defense Emerge under Antibiotic Selection for a CRISPR-Targeted Plasmid

Abstract: Enterococcus faecalis is a leading cause of hospital-acquired infections and disseminator of antibiotic resistance plasmids among Gram-positive bacteria. We have previously shown that E. faecalis strains with an active CRISPR-Cas system can prevent plasmid acquisition and thus limit the transmission of antibiotic resistance determinants. ABSTRACT Enterococcus faecalis is a Gram-positive bacterium that natively colonizes the human gastrointestinal tract and opportunistically causes life-threatening infections. Multidrug-resistant (MDR) E. faecalis strains have emerged that are replete with mobile genetic elements (MGEs). Non-MDR E. faecalis strains frequently possess CRISPR-Cas systems, which reduce the frequency of MGE acquisition. We demonstrated in previous studies that E. faecalis populations can transiently maintain both a functional CRISPR-Cas system and a CRISPR-Cas target. In this study, we used serial passage and deep sequencing to analyze these populations. In the presence of antibiotic selection for the plasmid, mutants with compromised CRISPR-Cas defense and enhanced ability to acquire a second antibiotic resistance plasmid emerged. Conversely, in the absence of selection, the plasmid was lost from wild-type E. faecalis populations but not E. faecalis populations that lacked the cas9 gene. Our results indicate that E. faecalis CRISPR-Cas can become compromised under antibiotic selection, generating populations with enhanced abilities to undergo horizontal gene transfer. IMPORTANCE Enterococcus faecalis is a leading cause of hospital-acquired infections and disseminator of antibiotic resistance plasmids among Gram-positive bacteria. We have previously shown that E. faecalis strains with an active CRISPR-Cas system can prevent plasmid acquisition and thus limit the transmission of antibiotic resistance determinants. However, CRISPR-Cas is not a perfect barrier. In this study, we observed populations of E. faecalis with transient coexistence of CRISPR-Cas and one of its plasmid targets. Our experimental data demonstrate that antibiotic selection results in compromised E. faecalis CRISPR-Cas function, thereby facilitating the acquisition of additional resistance plasmids by E. faecalis.

Decomposed Prompting to Answer Questions on a Course Discussion Board

Abstract: We propose and evaluate a question-answering system that uses decomposed prompting to classify and answer student questions on a course discussion board. Our system uses a large language model (LLM) to classify questions into one of four types: conceptual, homework, logistics, and not answerable. This enables us to employ a different strategy for answering questions that fall under different types. Using a variant of GPT-3, we achieve 81% classification accuracy. We discuss our system’s performance on answering conceptual questions from a machine learning course and various failure modes.

Development and Experimental Validation of a Novel Prognostic Signature for Gastric Cancer

Abstract: Simple Summary Gastric cancer (GC) accounts for a considerable amount of morbidity and mortality worldwide. This study developed and experimentally validated a prognostic risk gene signature (PRGS). This is a stable and robust signature for assessing the prognosis of gastric cancer. We performed multiple analyses through consensus clustering and binary classification to assess the robustness of the PRGS in other independent datasets. Additionally, this PRGS exhibited a superior accuracy compared to most traditional clinical markers, including molecular features, and other published signatures. Besides, we also detected the tumor purity, immune cell infiltration, and oncogenic mutation status of high- and low-PRGS groups. Abstract Background: Gastric cancer is a malignant tumor with high morbidity and mortality. Therefore, the accurate recognition of prognostic molecular markers is the key to improving treatment efficacy and prognosis. Methods: In this study, we developed a stable and robust signature through a series of processes using machine-learning approaches. This PRGS was further experimentally validated in clinical samples and a gastric cancer cell line. Results: The PRGS is an independent risk factor for overall survival that performs reliably and has a robust utility. Notably, PRGS proteins promote cancer cell proliferation by regulating the cell cycle. Besides, the high-risk group displayed a lower tumor purity, higher immune cell infiltration, and lower oncogenic mutation than the low-PRGS group. Conclusions: This PRGS could be a powerful and robust tool to improve clinical outcomes for individual gastric cancer patients.

Aging-induced pseudouridine synthase 10 impairs hematopoietic stem cells.

Abstract: Aged hematopoietic stem cells (HSCs) exhibit compromised reconstitution capacity and differentiation-bias towards myeloid lineage. While, the molecular mechanism behind it remains not fully understood. In this study, we observed that the expression of pseudouridine (Ψ) synthase 10 is increased in aged hematopoietic stem and progenitor cells (HSPCs) and enforced PUS10 recapitulates the phenotype of aged HSCs, which is not achieved by its Ψ synthase activity. Consistently, we observed no difference of tRNA pseudouridylation profile between young and aged HSPCs. No significant alteration of hematopoietic homeostasis and HSC function is observed in young Pus10-/- mice, while aged Pus10-/-mice exhibit mild alteration of hematopoietic homeostasis and HSC function. Moreover, we observed that PUS10 is ubiquitinated by E3 ubiquitin ligase CRL4DCAF1 complex and the increase of PUS10 in aged HSPCs is due to aging-declined CRL4DCAF1-mediated ubiquitination degradation signaling. Taken together, this study for the first time evaluated the role of PUS10 in HSC aging and function, and provided novel insight for HSC rejuvenation and clinical application.