Back

Multimodal AI for Single cfDNA Profiling and Cancer Screening

Wang, B.; Song, L.; Li, H.; Lin, N.; Xin, Y.; He, X.; Liu, W.; Liu, L.; Cui, J.; Li, X.; Mei, Y.; You, Q.; Zhu, H.; Zhao, G.; Chen, G.; Liu, J.; Zhu, B.; Sun, X.; Wu, X.; Gao, Z.; Ye, Y.

2025-12-30 bioinformatics
10.64898/2025.12.29.696856 bioRxiv
Show abstract

Cell-free DNA (cfDNA) serves as a non-invasive biomarker for cancer detection, but conventional methods face challenges due to the ultra-low abundance of tumor-derived cfDNA (ctDNA) among normal cfDNA. Though nucleosome-bound cfDNA harbors rich epigenomic features that could enable ctDNA identification by single-molecule multi-omics cross-validation, this remains unexplored due to methodological limits. Here, we developed a cfDNA sequencing approach integrating methylation, fragmentomics, and histone modifications at the single-molecule level; together with gene semantics and epigenomic annotations, these modalities were vectorized and fused to represent each cfDNA molecule. We trained a Transformer-based model (cfAI) to profile and evaluate ctDNA likelihood at molecule, gene, and sample levels. cfAI achieved [~]10-fold enrichment of cancer-derived signals over noise and reached 72.6% sensitivity at 93.1% specificity for multi-cancer detection. Our study establishes an innovative framework that overcomes the inherent signal-to-noise limitations of conventional assays and reveals biological features at molecular resolution for cancer detection.

Matching journals

The top 4 journals account for 50% of the predicted probability mass.

1
Nature Communications
○ 5641 papers in training set
Top 6%
22.0%
Similar papers in this journal
2
Nature Methods
○ 385 papers in training set
Top 0.6%
15.1%
Similar papers in this journal
3
Nature Biotechnology
○ 172 papers in training set
Top 0.1%
12.7%
Similar papers in this journal
4
Nature Machine Intelligence
○ 70 papers in training set
Top 0.6%
4.3%
Similar papers in this journal
50% of probability mass above
5
Cell
○ 431 papers in training set
Top 2%
4.0%
Similar papers in this journal
6
Nature Biomedical Engineering
○ 47 papers in training set
Top 0.4%
3.3%
7
Science Advances
● 1243 papers in training set
Top 10%
3.2%
8
Nature
○ 645 papers in training set
Top 4%
3.2%
9
Cell Systems
○ 201 papers in training set
Top 1%
3.2%
10
Genome Biology
○ 637 papers in training set
Top 4%
2.8%
11
Cell Genomics
○ 172 papers in training set
Top 2%
1.9%
12
Advanced Science
○ 286 papers in training set
Top 4%
1.7%
13
Nucleic Acids Research
◐ 1281 papers in training set
Top 8%
1.7%
14
Science
● 477 papers in training set
Top 6%
1.3%
15
Nature Genetics
○ 286 papers in training set
Top 4%
1.3%
16
Cell Reports Methods
○ 165 papers in training set
Top 2%
1.1%
17
Proceedings of the National Academy of Sciences
● 2444 papers in training set
Top 34%
1.1%
18
eLife
● 5828 papers in training set
Top 64%
0.9%
19
Communications Biology
○ 993 papers in training set
Top 29%
0.8%
20
Briefings in Bioinformatics
◐ 354 papers in training set
Top 7%
0.8%
21
Bioinformatics
◐ 1204 papers in training set
Top 9%
0.8%
22
Molecular Systems Biology
○ 162 papers in training set
Top 3%
0.8%
23
Molecular Cell
○ 350 papers in training set
Top 5%
0.8%
24
Cell Reports Medicine
○ 153 papers in training set
Top 5%
0.8%
25
Genome Research
● 468 papers in training set
Top 7%
0.6%
26
Cell Reports
○ 1498 papers in training set
Top 29%
0.6%
27
Genome Medicine
○ 183 papers in training set
Top 6%
0.6%
28
Nature Chemical Biology
○ 119 papers in training set
Top 3%
0.6%

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.