Qi Long

Qi Long

Auto-generated profile, based on publicly available information

Qi Long

Department of Biostatistics, Epidemiology and Informatics, University of Pennsylvania, Philadelphia, USA

Research Focus & Metrics

Academic Career Length

14Years (2013-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
2 October 2024 in BioData Mining
Verified
Journal Article
Open access
PDF file
JBJenna L. Ballard
Jenna L. Ballard
ZWZexuan Wang
Zexuan Wang
WLWenrui Li
Wenrui Li
Li ShenLi Shen
Li Shen
QLQi Long
Qi Long
Background: The rapid growth of deep learning, as well as the vast and ever-growing amount of available data, have provided ample opportunity for advances in fusion and analysis of complex and heterogeneous data types. Different data modalities provide complementary information that can be leveraged to gain a more complete understanding of each subject. In the biomedical domain, multi-omics data includes molecular (genomics, transcriptomics, proteomics, epigenomics, metabolomics, etc.) and imaging (radiomics, pathomics) modalities which, when combined, have the potential to improve performance on prediction, classification, clustering and other tasks. Deep learning encompasses a wide variety of methods, each of which have certain strengths and weaknesses for multi-omics integration. Method: In this review, we categorize recent deep learning-based approaches by their basic architectures and discuss their unique capabilities in relation to one another. We also discuss some emerging themes advancing the field of multi-omics integration. Results: Deep learning-based multi-omics integration methods were categorized broadly into non-generative (feedforward neural networks, graph convolutional neural networks, and autoencoders) and generative (variational methods, generative adversarial models, and a generative pretrained model). Generative methods have the advantage of being able to impose constraints on the shared representations to enforce certain properties or incorporate prior knowledge. They can also be used to generate or impute missing modalities. Recent advances achieved by these methods include the ability to handle incomplete data as well as going beyond the traditional molecular omics data types to integrate other modalities such as imaging data. Conclusion: We expect to see further growth in methods that can handle missingness, as this is a common challenge in working with complex and heterogeneous data. Additionally, methods that integrate more data types are expected to improve performance on downstream tasks by capturing a comprehensive view of each sample.

20 October 2021 in Proceedings of the National Academy of Sciences
Verified
Journal Article
Open access
PDF file
CFCong Fang
Cong Fang
HHHangfeng He
Hangfeng He
QLQi Long
Qi Long
WSWeijie J. Su
Weijie J. Su
In this paper, we introduce the Layer-Peeled Model, a nonconvex, yet analytically tractable, optimization program, in a quest to better understand deep neural networks that are trained for a sufficiently long time. As the name suggests, this model is derived by isolating the topmost layer from the remainder of the neural network, followed by imposing certain constraints separately on the two parts of the network. We demonstrate that the Layer-Peeled Model, albeit simple, inherits many characteristics of well-trained neural networks, thereby offering an effective tool for explaining and predicting common empirical patterns of deep-learning training. First, when working on class-balanced datasets, we prove that any solution to this model forms a simplex equiangular tight frame, which, in part, explains the recently discovered phenomenon of neural collapse [V. Papyan, X. Y. Han, D. L. Donoho, Proc. Natl. Acad. Sci. U.S.A. 117, 24652–24663 (2020)]. More importantly, when moving to the imbalanced case, our analysis of the Layer-Peeled Model reveals a hitherto-unknown phenomenon that we term Minority Collapse, which fundamentally limits the performance of deep-learning models on the minority classes. In addition, we use the Layer-Peeled Model to gain insights into how to mitigate Minority Collapse. Interestingly, this phenomenon is first predicted by the Layer-Peeled Model before being confirmed by our computational experiments. Significance The remarkable development of deep learning over the past decade relies heavily on sophisticated heuristics and tricks. To better exploit its potential in the coming decade, perhaps a rigorous framework for reasoning about deep learning is needed, which, however, is not easy to build due to the intricate details of neural networks. For near-term purposes, a practical alternative is to develop a mathematically tractable surrogate model, yet maintaining many characteristics of neural networks. This paper proposes a model of this kind that we term the Layer-Peeled Model. The effectiveness of this model is evidenced by, among others, its ability to reproduce a known empirical pattern and to predict a hitherto-unknown phenomenon when training deep-learning models on imbalanced datasets.

5 January 2023 in Journal of Biomedical Informatics
Verified
Journal Article
Closed access
EGEmily Getzen
Emily Getzen
LULyle Ungar
Lyle Ungar
DMDanielle Mowery
Danielle Mowery
XJXiaoqian Jiang
Xiaoqian Jiang
QLQi Long
Qi Long

2 October 2025 in Journal of the American Statistical Association
Verified
Journal Article
Closed access
JXJiancong Xiao
Jiancong Xiao
ZLZiniu Li
Ziniu Li
XXXingyu Xie
Xingyu Xie
EGEmily Getzen
Emily Getzen
CFCong Fang
Cong Fang
WSWeijie J. Su
Weijie J. Su
Accurately aligning large language models (LLMs) with human preferences is crucial for informing fair, economically sound, and statistically efficient decision-making processes. However, we argue that the predominant approach for aligning LLMs with human preferences through a reward model—reinforcement learning from human feedback (RLHF)—suffers from an inherent algorithmic bias due to its Kullback–Leibler-based regularization in optimization. In extreme cases, this bias could lead to a phenomenon we term preference collapse, where minority preferences are virtually disregarded. To mitigate this algorithmic bias, we introduce preference matching (PM) RLHF, a novel approach that provably aligns LLMs with the preference distribution of the reward model under the Bradley–Terry–Luce/Plackett–Luce model. Central to our approach is a PM regularizer that takes the form of the negative logarithm of the LLM’s policy probability distribution over responses, which helps the LLM balance response diversification and reward maximization. Notably, we obtain this regularizer by solving an ordinary differential equation that is necessary for the PM property. For practical implementation, we introduce a conditional variant of PM RLHF that is tailored to natural language generation. Finally, we empirically validate the effectiveness of conditional PM RLHF through experiments on the OPT and Llama-family models, demonstrating a 29% to 41% improvement in alignment with human preferences, as measured by a certain metric, compared to standard RLHF.

3 January 2025 in Nature Medicine
Verified
Journal Article
Open access
PDF file
XOXavier Orcutt
Xavier Orcutt
KCKan Chen
Kan Chen
RMRonac Mamtani
Ronac Mamtani
QLQi Long
Qi Long
RPRavi B. Parikh
Ravi B. Parikh
Randomized controlled trials (RCTs) evaluating anti-cancer agents often lack generalizability to real-world oncology patients. Although restrictive eligibility criteria contribute to this issue, the role of selection bias related to prognostic risk remains unclear. In this study, we developed TrialTranslator, a framework designed to systematically evaluate the generalizability of RCTs for oncology therapies. Using a nationwide database of electronic health records from Flatiron Health, this framework emulates RCTs across three prognostic phenotypes identified through machine learning models. We applied this approach to 11 landmark RCTs that investigated anti-cancer regimens considered standard of care for the four most prevalent advanced solid malignancies. Our analyses reveal that patients in low-risk and medium-risk phenotypes exhibit survival times and treatment-associated survival benefits similar to those observed in RCTs. In contrast, high-risk phenotypes show significantly lower survival times and treatment-associated survival benefits compared to RCTs. Our results were corroborated by a comprehensive robustness assessment, including examinations of specific patient subgroups, holdout validation and semi-synthetic data simulation. These findings suggest that the prognostic heterogeneity among real-world oncology patients plays a substantial role in the limited generalizability of RCT results. Machine learning frameworks may facilitate individual patient-level decision support and estimation of real-world treatment benefits to guide trial design.

8 April 2024 in Journal of Neuroinflammation
Verified
Journal Article
Open access
PDF file
CWChaoran Wu
Chaoran Wu
SZShiwen Zhang
Shiwen Zhang
HSHao Sun
Hao Sun
ALAo Li
Ao Li
FHFengsheng Hou
Fengsheng Hou
HLHong Liao
Hong Liao
Ischemic stroke is the leading cause of adult disability. Ischemia leads to progressive neuronal death and synapse loss. The engulfment of stressed synapses by microglia further contributes to the disruption of the surviving neuronal network and related brain function. Unfortunately, there is currently no effective target for suppressing the microglia-mediated synapse engulfment. Stimulator of interferon genes (STING) is an important participant in innate immune response. In the brain, microglia are the primary cell type that mediate immune response after brain insult. The intimate relationship between STING and microglia-mediated neuroinflammation has been gradually established. However, whether STING affects other functions of microglia remains elusive. In this study, we found that STING regulated microglial phagocytosis of synapses after photothrombotic stroke. The treatment of STING inhibitor H151 significantly improved the behavioral performance of injured mice in grid-walking test, cylinder test, and adhesive removal test after stroke. Moreover, the puncta number of engulfed SYP or PSD95 in microglia was reduced after consecutive H151 administration. Further analysis showed that the mRNA levels of several complement components and phagocytotic receptors were decreased after STING inhibition. Transcriptional factor STAT1 is known for regulating most of the decreased molecules. After STING inhibition, the nucleus translocation of phosphorylated STAT1 was also suppressed in microglia. Our data uncovered the novel regulatory effects of STING in microglial phagocytosis after stroke, and further emphasized STING as a potential drug-able target for post-stroke functional recovery.