Rewriting China's Breast Cancer Early Screening Paradigm: cfDNA Fragmentomics + AI Tackles the 'Last Mile' Challenge

In China, breast cancer is the most prevalent malignant tumor among women. Data from a research paper reveal a stark reality: between 2018 and 2019, only 22.3% of Chinese women in the eligible age group underwent breast cancer screening; furthermore, 24.1% of Chinese breast cancer patients were diagnosed at Stage III or later, a proportion more than double that of the United States (11.8%).

“The earlier the detection, the better the treatment outcome” is the most fundamental consensus in oncology. However, within China’s primary healthcare system, insufficient coverage of mammography equipment, the operator-dependent nature of ultrasound screening, and limited sensitivity for dense breast tissue collectively constitute a bottleneck in the “last mile” of early breast cancer screening.

Recently, a prospective multicenter study conducted by the Second Affiliated Hospital of Zhejiang University School of Medicine in collaboration with AI technology company OmixScience and seven other medical institutions was published in the prestigious international journal Nature Communications. The study enrolled 550 breast cancer patients and 289 benign controls, and developed TuFEst, a machine learning model based on cfDNA fragmentomics.^®(Tumor Fraction Estimator), which for the first time achieved ultra-early detection, non-invasive molecular subtyping, and lymph node metastasis prediction of breast cancer simultaneously from a single blood draw. This also represents the largest-scale global study on cfDNA fragmentomics combined with AI for breast cancer to date.

Behind this study stand two corresponding authors with distinctly different roles but deep complementarity:

Ni Chao, Deputy Director of the Department of Breast Surgery at the Second Affiliated Hospital of Zhejiang University School of Medicine, Chief Physician, and Doctoral Supervisor. As a frontline clinician performing 30 surgeries per week, he spearheaded this study from its inception, overseeing the entire process from sample collection and quality control systems to clinical validation, driven by insights from clinical pain points.

Lin Ziao, Founder & CEO of Aoming Xingcheng, holds dual Ph.D. degrees in Computer Science and Biomedical Informatics from Harvard University, and is the only Chinese doctoral graduate of Professor Gad Getz, a pioneer in the field of cancer genomics. He led his team to build TuFEst^®underlying algorithm platform, and is driving the transition of this technology from the laboratory to industrialization.

Breast cancer is one of the primary threats to women's health worldwide. Statistics show that the number of new breast cancer cases in China has risen from over 300,000 in 2015 to approximately 420,000 in 2020, making it the fourth most common cancer in the country. Of greater concern is that the average age of onset for breast cancer among Chinese women is significantly younger compared to their counterparts in Europe and the United States.

“Compared with Europe and the United States, the average age of onset in the Chinese population is ten years earlier, and the proportion of patients presenting at a more advanced stage at initial diagnosis is significantly higher than that in Western countries,” Director Ni Chao told VCBeat. As a frontline physician in the Department of Breast Surgery at the Second Affiliated Hospital of Zhejiang University School of Medicine, he deals daily with such cases of “late detection.”

In China, breast cancer screening primarily relies on two imaging modalities: ultrasound and mammography. However, the practical challenge lies in the fact that many primary healthcare institutions lack mammography equipment; furthermore, ultrasound examinations are operator-dependent, have limited sensitivity for dense breast tissue, and suffer from inconsistent diagnostic accuracy. These structural issues directly result in inadequate screening coverage and low rates of early detection.

So, can liquid biopsy, which has gained prominence in recent years, offer a new pathway? Currently, mainstream liquid biopsy approaches fall into two major categories: mutation detection and methylation analysis. Mutation detection requires large blood volumes and ultra-deep sequencing, and its sensitivity is limited in early-stage tumors due to extremely low variant allele frequencies (typically around 0.01%). Although methylation assays offer higher specificity, the bisulfite conversion process degrades approximately 90% of input DNA, resulting in sensitivities of only 50%–76% for early-stage tumors. More critically, cost constraints hinder the large-scale adoption of these approaches.

“Tests such as methylation or mutation site detection are prohibitively expensive, with per-test costs abroad ranging from approximately $300 to $1,000. This is clearly impractical in China,” explained Director Ni Chao. “Fragmentomics, by contrast, is relatively inexpensive, costing roughly in the range of hundreds of yuan, and thus offers greater feasibility. It may even be anticipated that the government will eventually incorporate it into routine screening protocols for the two major cancers.”

Furthermore, genetic testing faces issues regarding applicability across different ethnicities. Director Ni Chao emphasized, “Due to distinct genetic backgrounds and significant ethnic differences, foreign models are not applicable in China. In light of biosafety concerns, we must develop a solution tailored specifically for the Chinese population.”

Against this clinical backdrop, Director Ni Chao’s team, in collaboration with Aoming Xingcheng, launched a prospective, multicenter study encompassing eight centers and more than 800 samples.

Unlike the retrospective designs of previous liquid biopsy studies, this study employed a rigorous prospective cohort design from the outset (Registration No. NCT06016790), encompassing seven tertiary hospitals and one community health service center. The study enrolled 550 patients with pathologically confirmed breast cancer (77.1% of whom had early-stage disease, i.e., stage 0–IIa) and 289 benign controls. Low-pass whole-genome sequencing (approximately 2× coverage) was performed on peripheral blood samples from all participants.

“The greatest taboo in cfDNA research is the use of substandard samples.” Reflecting on quality control challenges, Director Ni Chao recalled lessons learned from their biobank: “We previously attempted to retrieve blood samples from the central biobank and found that the eligibility rate was less than 30%. If initial research is based on non-conforming samples, the resulting conclusions may only be sufficient for publishing a paper, lacking practical value for future clinical adoption.”

Therefore, all samples in this study were collected de novo: specialized cfDNA collection tubes were used, transported in dedicated temperature-controlled containers, and all centers adhered to unified standard operating procedures. This rigorous quality control “from scratch” serves as a crucial safeguard for the credibility of the study data.

At the algorithmic level, TuFEst^®The model employs a multimodal fusion strategy, integrating multi-dimensional fragmentomics features—including cfDNA fragment size distribution, end-sequence characteristics, breakage site patterns, repetitive element fragmentation information, and transcription factor binding site coverage—and identifies the optimal Stacked Ensemble Model from six machine learning architectures.

“TuFEst^®“Breaking through the dimensional limitations of traditional testing,” explained Dr. Lin Ziao, “enables high signal-to-noise ratio extraction of faint tumor signals at an extremely low cost, making large-scale population screening possible.”

Level 1 Answer: Is there cancer?

TuFEst^®The model demonstrated excellent performance in differentiating between benign and malignant breast cancer. In the training set, the model achieved a sensitivity of 95%, a specificity of 78.3%, and an AUC of 0.937; in the external validation set, the sensitivity was 92%, the specificity was 86.9%, and the AUC further increased to 0.968.

More compelling is the validation from a series of special cases: 26 patients with early-stage breast cancer from four hospitals were initially classified as benign (BI-RADS category 3) by both ultrasound and mammography during routine screening. However, within six months, biopsies performed due to lesion enlargement pathologically confirmed invasive carcinoma. In single-blind validation using blood samples collected during the initial imaging assessment of these patients, TuFEst^®Twenty-five cases were successfully identified, with a sensitivity of up to 96.2%, and all these cases were early-stage cancers or those without lymph node metastasis.

Second-level response: What type?

The treatment of breast cancer is highly dependent on molecular subtyping (ER/PR/HER2 status), with significant differences in biological behavior and therapeutic strategies among different subtypes. The research team at TuFEst^®Building on this framework, the TuFEst-MS model was further developed to distinguish luminal (ER+/PR+/HER2-), HER2-positive, and triple-negative breast cancer (TNBC) subtypes using only blood-based cfDNA data, achieving AUCs of 0.939, 0.925, and 0.893 in the validation set, respectively. Among 21 patients with metastatic disease, the model achieved an accuracy of 85.7% in predicting the molecular subtype of metastatic lesions.

This means that in cases where tissue biopsy is inaccessible or difficult (such as when metastatic lesions are located in challenging sites), TuFEst-MS can provide a non-invasive subtyping reference for clinical practice, thereby assisting in treatment decision-making.

The Third Question: Has Metastasis Occurred?

Whether axillary lymph node metastasis has occurred is one of the key factors determining the treatment plan for breast cancer. Traditional imaging modalities have a sensitivity of only approximately 50%–60% in lymph node assessment, with limited accuracy, often leading to unnecessary lymph node dissection surgeries or undertreatment. The TuFEst-LN model achieved an AUC of 0.874 in the validation set. Among 85 “challenging” discordant cases that were imaging-positive but pathology-negative, the model’s negative predictive value reached as high as 97.6%.

The clinical significance of this finding lies in its potential to help more patients clinically assessed as node-negative (cN0) confidently avoid unnecessary axillary surgery, thereby enabling more precise treatment de-escalation.

The paper’s data further reveal deeper biological associations: by correlating cfDNA “cancer scores” with transcriptomic data from paired tumor tissues (n=79), the study found that tumors with high scores exhibited more active signaling pathways related to cell proliferation, inflammatory responses, and immune evasion, as well as greater immune cell infiltration. This suggests that TuFEst^®Not only a capturer of signals, but also to some extent reflects the invasive biological behavior of tumors.

“We are not testing for the sake of testing,” summarized Director Ni Chao. “There are still many unresolved issues in clinical practice. If a single blood sample could simultaneously address these three issues, it would directly transform many future clinical diagnostic and treatment workflows.”

Director Ni Chao is, first and foremost, a surgeon.

“I am a frontline clinical surgeon, performing approximately 30 surgeries per week,” he told VCBeat. This means his understanding of breast cancer is not derived from literature and databases, but from the real patients he encounters daily. It is precisely this frontline clinical experience that has shaped his research philosophy of being “guided by clinical questions.”

In addition to early screening and diagnosis of breast cancer, Ni Chao’s research also focuses on the tumor immune microenvironment. His team has identified several new druggable targets and is advancing clinical studies. However, in his view, the most transformative force reshaping his entire experimental framework in recent years has been artificial intelligence technology.

“Take cfDNA fragmentomics in this case: we extracted more than 3,000 features. How could one possibly screen them manually? That’s impossible,” he candidly stated. “It is essential to pursue interdisciplinary collaboration between medicine and engineering, leveraging artificial intelligence or machine learning techniques to accomplish this.”

Interdisciplinary collaboration was the key to making this research possible. According to Director Ni Chao, the relationship between clinicians and the technical team is complementary. “They do not always fully understand the true pain points in current clinical practice, but we do. However, they possess the technological expertise. Only through such a strong alliance can we produce research outcomes that truly benefit the public and align with national priorities.”

Just prior to the publication of the paper, Director Ni Chao’s team had already launched the next phase of their prospective population-based screening study. He revealed that the team has undertaken a National Key R&D Program project (a major special project of the Ministry of Science and Technology) focused on early screening for breast cancer, with the goal of delivering validation results applicable to real-world screening scenarios within three years.

His vision is clear and specific: “We aim to implement precision-stratified screening, moving away from the previous one-size-fits-all approach.” High-risk populations should undergo screening at shorter intervals, while screening intervals for low-risk populations can be appropriately extended. This strategy not only reduces the burden on national health economics and alleviates the financial burden of medical care on the public, but also enables earlier detection of patients who truly require intervention.

“Ultimately,” he said, “we still hope to have the opportunity for our approach to be adopted by the government as the national protocol for future female breast cancer screening in China.”

If Dr. Ni Chao represents the clinical perspective, then Dr. Lin Ziao, born in the 1990s, embodies another narrative thread at the technological foundation of this research; he is TuFEst^®The Core Builder of the Algorithm Platform and the Driver of Translating Scientific Research Achievements into Applications.

Dr. Lin Ziao earned his bachelor’s degree from Zhejiang University and subsequently pursued dual doctoral degrees in Computer Science and Biomedical Informatics at Harvard University, where he studied under the supervision of the internationally renowned scholar Professor Gad Getz. As one of the pioneers in cancer genomics and a core architect of international benchmark genomic databases such as The Cancer Genome Atlas (TCGA), Professor Getz counts Dr. Lin as his only Chinese doctoral graduate.

During his tenure at Harvard, Dr. Lin Ziao was deeply involved in major international scientific initiatives such as the ICGC, TCGA, and PCAWG. He gained comprehensive experience in building an end-to-end technical infrastructure, encompassing biological sample processing, full-process quality control, high-throughput data generation, and core algorithm development. This experience directly laid the foundation for thisNature CommunicationsTuFEst in the paper^®The algorithmic design of the model provides a methodological foundation.

“The greatest impact Professor Gad Getz had on me was not any specific technique, but rather a comprehensive methodology for scientific research and a data-driven mindset,” recalled Dr. Lin Ziao. “From sample processing and end-to-end quality control to algorithm development, model validation, and cross-validation of multi-center data, every step emphasized rigor, reproducibility, and generalizability across large-scale datasets.”

It is this methodology that constitutes TuFEst.^®This provides the technical foundation for efficient screening and modeling across more than 3,000 fragmentomics features. In this study, Dr. Lin Ziao’s team was responsible for the core algorithmic architecture, including the design of multimodal feature fusion strategies, a systematic comparison of six machine learning architectures, and the selection and optimization of the final Stacked Ensemble Model.

“TuFEst^®“Breaking through the dimensional limitations of traditional testing,” explained Dr. Lin Ziao, “it enables high signal-to-noise ratio extraction of weak tumor signals at extremely low cost, making large-scale population screening possible.”

The transition of this research from algorithmic development to clinical validation would not have been possible without close collaboration with Ni Chao’s team. In an interview, Ni Chao remarked, “They do not always fully understand the true pain points in current clinical practice, whereas we do. However, they possess the technological expertise. Only through such a strong partnership can we produce research outcomes that truly benefit the public.”

In Dr. Lin Ziao’s view, this complementarity between clinical practice and technology represents precisely the collaborative paradigm most needed for medical innovation in China. Clinicians provide requirement definitions based on real-world scenarios and enforce rigorous sample quality control, while technical teams are responsible for algorithm development and model validation. The two parties supervise each other, thereby ensuring both data credibility and the clinical applicability of research findings.

“Technology-driven enterprises are, in essence, machines that amplify scientific research outcomes. Scientific research addresses the question of ‘feasibility,’ whereas we aim to resolve whether value can be created at scale and on a sustainable basis.” Dr. Lin Ziao defines this as the core challenge in translating scientific achievements into practical applications. He believes that these two aspects are not mutually exclusive but rather two stages of the same process. “Enabling scientific research to truly step out of the laboratory, benefit the public, and generate greater social value—this is what we are truly committed to achieving.”

After completing their Ph.D. studies, Dr. Lin Ziao, together with his Harvard classmates Dr. Zhao Hanchen and Dr. Hao Jin, returned to China to co-found OmixScience. OmixScience is an AI-driven technology enterprise grounded in the foundational paradigm of AI for Science (AI4S), specializing in large-model-based individual health trend prediction and disease intervention. With AI as its core engine, the company is dedicated to promoting the deep integration of AI with frontier life sciences, building a strategic core platform that supports the entire upstream and downstream biomedical industry chain, and providing paradigmatic leadership and foundational support for the development of new-quality productive forces and global technological innovation. The company has participated in undertaking multi-level projects, including the construction of national AI innovation highlands, National Key R&D Programs (the "Four Major Chronic Diseases" special project), Zhejiang Provincial Major Science and Technology Projects, and Shenzhen Major Projects. It has accumulated over 60 intellectual property rights and contributed to the formulation of two industry standards. To date, the company has secured over RMB 100 million in Series A strategic financing, with investors including Shenzhen Capital Group, Fosun Pharma (Fosun Health Capital), China Taiping, Guangdong Traditional Chinese Medicine Big Health Fund (jointly established by China Development Bank Financial Corporation and Hengjian Holdings), and Hangzhou Industrial Investment Group, among other renowned industrial capital firms and government investment funds. Furthermore, it has established deep mechanisms for collaborative achievement transformation with multiple top-tier international research institutions, including the Shenzhen Bay Laboratory, forming a full-chain layout covering technology R&D, clinical application, and industrial ecosystem empowerment.

“We are not developing a standalone screening tool,” he emphasized. “At its core, we are building a life-status decoding system capable of continuous learning.”

In Dr. Lin Ziao’s view, it is the dynamic changes at the molecular level, rather than the “result presentation” at the imaging stage, that truly determine the onset and progression of diseases. Multi-omics data serve as the production materials that directly reflect these “underlying signals.” The value of AI lies in integrating data from different modalities—including multi-omics data such as genomics, transcriptomics, and proteomics, as well as multi-source heterogeneous information such as medical imaging, digital pathology, and electronic health records—to capture more complex associations.

This study, published inNature CommunicationsThis research marks a new phase in the application of liquid biopsy for breast cancer management, transitioning from single-dimensional detection to an integrated assessment model of “one blood draw, triple answers.” However, for Director Ni Chao and Dr. Lin Ziao, this is merely a starting point.

Director Ni Chao’s team has launched a follow-up prospective population screening study, with the goal of developing a productized solution applicable to real-world screening scenarios within three years. Dr. Lin Ziao is dedicated to advancing TuFEst^®The platform’s application scenarios are continuously expanding from validated areas such as breast cancer to a broader range of disease indications. He believes that the underlying methodology established by the platform possesses high generalizability, capable of providing systematic, full-cycle support for complex diseases—covering screening, diagnosis, treatment, intervention, prognosis management, and monitoring—and ultimately achieving personalized health management through AI agents, thereby forming a complete closed loop. Furthermore, by integrating AI agents with CRO (Contract Research Organization) services, the platform can offer comprehensive support to diverse stakeholders, including pharmaceutical companies, biopharmaceutical firms, hospitals, insurance companies, and government entities.

Dr. Ni Chao, who focuses on clinical pain points, and Dr. Lin Ziao, who masters an internationally leading original AI technology system, have joined forces—a partnership that serves as a vivid microcosm of China’s medical innovation process. Cutting-edge technologies are transcending laboratories and academic papers, transforming into smart healthcare products and robust clinical solutions that serve the vast population of Chinese patients through the powerful synergy of “internationally advanced technologies” and “China’s clinical needs.” This represents not only the deep integration of “AI + Healthcare,” but also a paradigm of co-created value by “global wisdom” and “local practice.”

Paper Link:https://www.nature.com/articles/s41467-026-70204-w