XLXiong Liu

Xiong Liu

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Xiong Liu has a total of 101 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Silicon Transcendentalism: This Summer in Concord

    by Xiong Liu

    This treatise proposes and defines the foundational conceptual frameworks of "AI Transcendentalism" (AI 超验主义, or AI 超验) and "Silicon Transcendentalism" (硅基超验主义, or 硅基超验), establishing the technical-philosophical boundaries between computational state representation and biological intervention. As large language models and autonomous agents assume an expanding share of decision-making in pharmaceutical research, modern R&D operates under an implicit technological faith: that continuous architectural scaling and tokenized biological data will automatically yield causal discoveries. Writing from Walden Pond, the historical center of New England Transcendentalism, the author argues that this ideology conflates state classification with dynamic transformation. Observational datasets (such as static single-cell transcriptomic profiles) teach a model the static "nouns" (cellular states) of biology, but fail to inherently resolve the transactional "verbs" (directional perturbations) that govern therapeutic outcomes. No amount of computational scaling substitutes for anchoring algorithmic hypotheses within complementary multi-modal matrices and empirical wet-lab validation. The essay extends this inquiry to clinical outcome prediction, where increasingly accurate forecasts of longitudinal disease trajectories and mortality risk remain safe only as long as an accurate forecast is not mistaken for a validated treatment-selection rule, and where even a correctly used forecast trades human narrative contingency for statistical certainty. The treatise concludes by outlining an Emersonian middle path of evidence-anchored technological practice, positioning Silicon Transcendentalism as a distinct modern philosophical inquiry into the boundaries of human and machine intelligence.

  • Open Access

    Cell-CLIP: Multi-algorithm causal discovery of directed cell program interaction networks from single-cell transcriptomics

    by Mohammed Khodor Firas Al-Tal, Xiong Liu, Joseph Zhou

    Cell-cell communication networks govern tissue function in health and disease, yet existing computational tools rely on static ligand–receptor databases or correlation-based analyses that cannot distinguish direct causal relationships from confounded or indirect associations. We introduce Cell-CLIP (Cellular Causal Learning and Inference Pipeline), a modular framework that combines (i) mask-aware patient-level JASMINE activity scoring with low-percentile imputation in raw activity space, (ii) five complementary causal-discovery algorithms — PC, GES, FCI, DirectLiNGAM, and GRaSP — each fed an algorithm-matched per-column transform of the same masked JASMINE matrix, (iii) direction-preserving bootstrap aggregation with a weighted two-phase consensus, (iv) Joint Causal Inference (JCI) for principled pooling of patient cohorts via context indicators, and (v) a residual-independence (RESIT) direction audit, all evaluated under a systematic three-by-three-by-three hyperparameter sweep. Applied to a 611-patient multi-phenotype lung atlas spanning normal tissue, COVID-19 pneumonia, and lung cancer, the cross-cohort joint-causal-inference graph constructed by stratifying patients into normal, COVID-19, and tumour contexts before pooling recovers 100% of evaluable curated ground-truth edges (8/8; Wilson 95% CI 0.68–1.00) with 75.0% direction accuracy (6/8; 95% CI 0.41–0.95) and zero forbidden-orientation violations (0/8; 95% CI 0.00–0.32) — the only configuration in a comprehensive ablation series spanning 7 cohort/pool configurations, 27 hyperparameter settings, and 3 consensus types to simultaneously reach all three optimal Pareto corners. The single-cohort universal cross-phenotype graph (12 cross-phenotype programs over the full 611-patient atlas) similarly achieves 100% evaluable recall (8/8; 95% CI 0.68–1.00) with 62.5% direction accuracy (5/8; 95% CI 0.31–0.86) and zero forbidden violations (0/3; 95% CI 0.00–0.56) under the same systematic hyperparameter sweep with low-percentile imputation, exceeding an earlier zero-imputation baseline configuration of the pipeline on every reported metric (Table 6, Table 8); given small ground-truth denominators ( n _eval = 8), these point-estimate improvements lie within the Wilson 95% intervals of the baseline and should be read as direction-of-effect rather than as confirmatory comparisons. We further show that single-cohort causal inference on the disease-extended pooled cohorts (which apply the COVID- and cancer-extended program sets to all 611 patients) recovers strong skeleton recall (35.0%–55.2% evaluable) but suffers a structural direction collapse (down to 14.3% direction accuracy and 60% forbidden rate at union) caused by within-cohort averaging over heterogeneous normal-plus-disease patient populations; the patient-stratified joint-causal-inference pool dissolves this collapse without altering the underlying algorithms. Cell-CLIP provides a hypothesis-generating framework for inferring directed cell-program networks from patient-level scRNA-seq atlases, with all reported relationships requiring experimental validation due to observational data limitations.

  • Open Access

    Single Cell Foundation Models Evaluation (scFME) for In-Silico Perturbation

    by James Boylan, Elizaveta Solovyeva, Theophile Bouiller, Xiong Liu, Sebastian Hoersch, Bulent Ataman, Jeremy Jenkins, Murthy Devarakonda

    Foundation models pre-trained on large single-cell RNA atlases offer a compelling alternative to in-vitro experimentation for understanding gene regulatory networks and conducting gene perturbation analyses, with significant implications for target identification. Numerous foundation models have been developed, building upon early efforts such as Geneformer and scGPT. Hyperparameter optimization also results in multiple variants which require comparative analysis. Current benchmarking approaches focus on feature-based assessments or intuitive biological and statistical tasks, which may not align with the models’ training objectives. A recent study proposed a systematic benchmarking framework; however, its scope was limited to pre-trained (zero-shot) models. To address these limitations, we propose Single-Cell Foundation Model Evaluation (scFME)—a systematic method designed to benchmark fine-tuned foundation models for insilico perturbation (ISP). scFME ensures comprehensive and robust assessment by requiring sufficient separation between control and perturbed cells at the outset and by quantifying ISP accuracy against zero and random perturbation baselines. Furthermore, scFME enables exploration of model performance across distinct gene categories, facilitating biological interpretation and functional relevance. Using this framework, we evaluated several commonly used models (and some of their variants) and demonstrated that the methodology effectively characterizes their performance in ISP studies. Our results position scFME as a versatile and rigorous methodology for evaluating and comparing current and future foundation models.


  • Open Access

    Evaluating Foundation Models for In-Silico Perturbation

    by Xiong Liu, James Boylan, Theophile Bouiller, Elizaveta Solovyeva, Bulent Ataman, Sebastian Hoersch, Jeremy Jenkins, Murthy Devarakonda

    In-silico perturbation (ISP) offers a scalable alternative to traditional gene perturbation experiments, yet evaluation of foundation models for ISP remains underexplored. We introduce a novel evaluation framework, the single cell in-silico perturbation framework (scISP), to benchmark ISP models against in-vitro experimental data using biologically meaningful metrics, including cell state separation accuracy, ISP accuracy, and mean reciprocal rank (MRR) for predicting perturbed genes. Complementary functional analyses evaluate model performance across diverse gene categories. Using scISP, we assess two well-known pre-trained foundation models, Geneformer and scGPT, alongside the deep learning model, GEARS, highlighting their respective strengths and limitations in simulating cell state transitions and identifying perturbed genes. These analyses reveal intrinsic differences across models, offering opportunities to optimize foundation models for specific biological contexts or gene categories. Our extensible framework establishes a robust bridge between computational predictions and experimental validation, advancing gene perturbation research and biological discovery.


  • Open Access

    Clinical Trial Information Extraction with BERT

    by Xiong Liu, Greg L. Hersch, Iya Khalil, Murthy Devarakonda

    Background Clinical trials are the gateway to new treatments. Clinical trial designs are documented in unstructured text and insights from natural language processing (NLP) of the documents can be useful in new trial design. Here we identify entity and relation types relevant to clinical trial design and propose a framework called CT-BERT to extract entities and relations from the clinical trial documents. Methods CT-BERT uses BERT models for named entity recognition (NER) and a hybrid approach for relation extraction including rule-based and machine learning-based models. We trained NER models to extract eligibility criteria entities by fine-tuning a set of pre-trained BERT models, including BERT, BioBERT, ClinicalBERT and BlueBERT. We then developed a rule-based relation extraction model to extract relations based on the NER results. For comparison, we also trained machine learning models using BERT to classify entity pairs into relation types. We compared the performance of CT-BERT with recent baseline methods including attention-based BiLSTM (Att-BiLSTM) and Criteria2Query. Results The NER training result shows that BioBERT large fine-tuned on clinical trials data achieves the best F1 score and is thus selected as the optimal NER model for CT-BERT. The result on a benchmark dataset showed that CT-BERT NER outperforms Att-BiLSTM and Criteria2Query. CT-BERT NER achieves 0.844 F1, while Att-BiLSTM and Criteria2Query achieve 0.802 F1 and 0.804 F1, respectively. For relation extraction, the CT-BERT rule-based model achieves 0.870 F1, while the equivalent model in Criteria2Query only achieves 0.793 F1. The CT-BERT rule-based model also outperforms the machine learning models trained on a curated relation data set. This suggests the advantage of the rule-based method in relation extraction. Conclusions The NER and relation extraction results demonstrate the effectiveness of CT-BERT for clinical trial information extraction. These results form the foundation for advancing clinical trials NLP towards AI-enhanced clinical trial design.


  • Open Access

    A Qualitative Analysis of Provider Notes of Atopic Dermatitis-Related Visits Using Natural Language Processing Methods

    by Evangeline J. Pierce, Natalie N. Boytsov, Joe J. Vasey, Theresa C. Sudaria, Xiong Liu, Kevin W. Lavelle, Alina N. Bogdanov, Orin M. Goldblum

    Introduction Real-world disease management of atopic dermatitis (AD) is hampered by a lack of consistency between providers that treat AD regarding assessment of severity, disease activity, and quality of life. Variability and inconsistency in documentation makes it difficult to understand the impact of AD. This study summarizes AD-related symptoms and concerns captured in unstructured qualitative provider notes by healthcare providers during visits with patients with AD. Methods Provider notes were obtained for patients with AD (n = 133,025) from a USA-based ambulatory electronic health records system. The sample included both children (n = 69,551) and adults at least 18 years of age (n = 63,474) receiving treatment from a variety of specialties including primary care, dermatology, and allergy/immunology. Key skin-related words were identified from a review of a sample of notes and natural language processing (NLP) was applied to determine the frequency of the keywords and bigram patterns. Results Provider notes largely focused on symptoms (primarily itch) and symptom relief rather than the impact of AD on work or lifestyle. Despite the known relationship between itch and skin pain, neuralgia was not widely documented. Compared to primary care providers, dermatologists’ and allergist/immunologists’ notes had more documentation of symptom-related issues. Personal and work/life burden issues were not widely documented regardless of specialty. Conclusion The topics documented in case notes by healthcare providers about their patients with AD focus largely on symptoms and, to a lesser extent, treatment, but do not reflect the burden of AD on patients’ lives. This finding highlights a potential care gap that warrants further investigation.


  • Applications of artificial intelligence in drug development using real-world data

    by Zhaoyi Chen, Xiong Liu, William Hogan, Elizabeth Shenkman, Jiang Bian


  • Open Access

    Attention-Based LSTM Network for COVID-19 Clinical Trial Parsing

    by Xiong Liu, Luca A. Finelli, Greg L. Hersch, Iya Khalil

    COVID-19 clinical trial design is a critical task in developing therapeutics for the prevention and treatment of COVID-19. In this study, we apply a deep learning approach to extract eligibility criteria variables from COVID-19 trials to enable quantitative analysis of trial design and optimization. Specifically, we train attention-based bidirectional Long Short-Term Memory (Att-BiLSTM) models and use the optimal model to extract entities (i.e., variables) from the eligibility criteria of COVID-19 trials. We compare the performance of Att-BiLSTM with traditional ontology-based method. The result on a benchmark dataset shows that Att-BiLSTM outperforms the ontology model. Att-BiLSTM achieves a precision of 0.942, recall of 0.810, and F1 of 0.871, while the ontology model only achieves a precision of 0.715, recall of 0.659, and F1 of 0.686. The extracted variables can help characterize patient populations eligible for COVID-19 trials. Our analyses demonstrate that Att-BiLSTM is an effective approach for eligibility criteria parsing.


  • Open Access

    Predicting Heart Failure Readmission from Clinical Notes Using Deep Learning

    by Xiong Liu, Yu Chen, Jay Bae, Hu Li, Joseph Johnston, Todd Sanger

    Heart failure hospitalization is a severe burden on healthcare. How to predict and therefore prevent readmission has been a significant challenge in outcomes research. To address this, we propose a deep learning approach to predict readmission from clinical notes. Unlike conventional methods that use structured data for prediction, we leverage the unstructured clinical notes to train deep learning models based on convolutional neural networks (CNN). We then use the trained models to classify and predict potentially high-risk admissions/patients. For evaluation, we trained CNNs using the discharge summary notes in the MIMIC III database. We also trained regular machine learning models based on random forest using the same datasets. The result shows that deep learning models outperform the regular models in prediction tasks. CNN method achieves a F1 score of 0.756 in general readmission prediction and 0.733 in 30-day readmission prediction, while random forest only achieves a F1 score of 0.674 and 0.656 respectively. We also propose a chi-square test based method to interpret key features associated with deep learning predicted readmissions. It reveals clinical insights about readmission embedded in the clinical notes. Collectively, our method can make the human evaluation process more efficient and potentially facilitate the reduction of readmission rates.

  • Open Access

    A Bayesian gene network reveals insight into the JAK-STAT pathway in systemic lupus erythematosus

    by Yupeng Li, Richard Higgs, Robert Hoffman, Ernst Dow, Xiong Liu, Michelle Petri, Daniel Wallace, Thomas Dörner, Brian Eastwood, Bradley Miller, et al.

    Systemic lupus erythematosus (SLE) is a chronic, remitting, and relapsing, inflammatory disease involving multiple organs, which exhibits abnormalities of both the innate and adaptive immune responses. A limited number of transcriptomic studies have characterized the gene pathways involved in SLE in an attempt to identify the key pathogenic drivers of the disease. In order to further advance our understanding of the pathogenesis of SLE, we used a novel Bayesian network algorithm to hybridize knowledge- and data-driven methods, and then applied the algorithm to build an SLE gene network using transcriptomic data from 1,760 SLE patients’ RNA from the two tabalumab Phase III trials (ILLUMINATE-I & -II), the largest SLE RNA dataset to date. Further, based on the gene network, we carried out hub- and key driver-gene analyses for gene prioritization. Our analyses identified that the JAK-STAT pathway genes, including JAK2, STAT1, and STAT2, played essential roles in SLE pathogenesis, and reaffirmed the recent discovery of pathogenic relevance of JAK-STAT signaling in SLE. Additionally, we showed that other genes, such as IRF1, IRF7, PDIA4, FAM72C, TNFSF10, DHX58, SIGLEC1, and PML, may be also important in SLE and serve as potential therapeutic targets for SLE. In summary, using a hybridized network construction approach, we systematically investigated gene-gene interactions based on their transcriptomic profiles, prioritized genes based on their importance in the network structure, and revealed new insights into SLE activity.