Introduction

Liquid biopsy through analysis of cell-free DNA (cfDNA) in blood plasma has emerged as a powerful approach for non-invasive cancer detection, monitoring treatment response, and detecting minimal residual disease. Circulating tumor DNA (ctDNA) constitutes a fraction of total cfDNA and carries tumor-specific alterations including somatic mutations, copy number alterations, and aberrant methylation patterns. Computational analysis of cfDNA requires specialized algorithms to handle the low tumor fractions (often <1%) and unique biological properties of cfDNA.

Methods

Fragment Length Analysis

cfDNA fragment lengths were simulated for 40 patients with varying tumor fractions (0.005-0.5). Normal cfDNA follows a bimodal distribution with mononucleosomal (~167bp) and dinucleosomal (~334bp) peaks. ctDNA fragments are shorter (~145bp) due to altered nucleosome positioning in tumor cells. The short/long fragment ratio (100-150bp / 160-180bp) was computed as a tumor fraction proxy.

Tumor Fraction Estimation

Copy number-based tumor fraction estimation was implemented following the ichorCNA approach. Read depth across 500 genomic bins (1Mb each) was modeled as a mixture of normal (CN=2) and tumor (CN=2±CNA) components. Tumor fraction was estimated from the median absolute deviation (MAD) of log2 depth ratios.

Low-VAF Variant Calling

Somatic variants were called from deep sequencing data (mean depth=1000x) using a binomial test against a sequencing error rate of 0.3%. Variants with FDR<0.01 and VAF≥0.5% were called as somatic. True somatic variants (n=150) were simulated with VAF following Beta(1.5, 10) distribution.

Tissue-of-Origin Deconvolution

Reference methylation profiles for 5 cancer types (LUAD, BRCA, CRC, PRAD, HCC) were constructed from 200 tissue-specific CpG markers. Patient cfDNA methylation was modeled as a mixture of tumor and normal profiles. Non-negative least squares (NNLS) deconvolution was used to estimate tissue-of-origin fractions.

Longitudinal Tracking

Five response patterns were modeled: complete responder (exponential decay), partial responder, progressor (exponential growth), stable disease, and mixed response.

Results

Fragment length analysis showed strong correlation between short/long ratio and true tumor fraction (r=0.995), demonstrating the utility of fragment size as a non-invasive tumor fraction proxy. Variant calling achieved F1=0.917 with Precision=1.000 and Recall=0.847, with median called VAF of 4.4%. Tissue-of-origin deconvolution correctly classified 38/40 patients (accuracy=0.950). Longitudinal tracking successfully distinguished all 5 response patterns.

Discussion

LiquidBiopsyEngine provides an integrated computational framework for cfDNA analysis. The high fragment ratio correlation demonstrates the biological signal in nucleosome positioning differences between tumor and normal cfDNA. The high precision variant calling is critical for clinical applications where false positives are costly. Future extensions include integration with epigenomic cfDNA fragmentation patterns and machine learning-based tumor fraction estimation.

Code Availability

Full source code: https://github.com/BioTender-max/LiquidBiopsyEngine

# pip install numpy scipy matplotlib
python liquid_biopsy_engine.py

Key Results

• Patients: 40, Time points: 5, Genomic bins: 500
• Fragment ratio vs TF: r=0.995
• Variant calling: F1=0.917, Precision=1.000
• Tissue-of-origin accuracy: 0.950 (38/40)
• Longitudinal patterns: 5 (responder/partial/progressor/stable/mixed)