ALICE: Hybrid AI for Liquid Biopsy Reveals a New Cell Type in Pancreatic Cancer Blood

Theranostics 2020 AI 5 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Pages 1-3
Liquid Biopsy and the Promise of Blood-Based Cancer Detection

What liquid biopsy means: A liquid biopsy is a blood test that looks for signs of cancer circulating in the bloodstream rather than requiring a surgical biopsy of the tumor itself. In pancreatic cancer, where tumors are often located deep in the abdomen and are difficult to biopsy safely, liquid biopsy approaches offer a less invasive way to detect, monitor, and characterize the disease.

Circulating tumor cells as biomarkers: Circulating tumor cells (CTCs) are cancer cells that have broken away from the primary tumor and entered the bloodstream. Their presence and number in blood samples can indicate how aggressively the tumor is spreading, and their molecular characteristics can reveal information about the cancer's biology that might guide treatment decisions.

The challenge of finding rare cells: CTCs are extremely rare, often fewer than one cell per milliliter of blood among billions of normal blood cells. Accurately detecting and counting them requires both highly sensitive detection methods and sophisticated analysis pipelines that can distinguish true CTCs from normal cells and imaging artifacts.

The need for automated, reproducible analysis: Manual analysis of liquid biopsy samples by human observers is slow, subjective, and difficult to standardize across different laboratories. Automated AI systems that can reproduce expert-level analysis consistently at scale are essential for bringing liquid biopsy into routine clinical use.

TL;DR: Liquid biopsy using circulating tumor cells is a minimally invasive approach to monitoring pancreatic cancer, but detecting rare CTCs reliably requires sophisticated automated AI analysis.
Pages 4-7
ALICE: Combining Rules and Machine Learning for Cell Classification

The ALICE system architecture: ALICE (Automated Liquid Biopsy Cell Enumeration) is a hybrid AI system that combines rule-based logic with machine learning classification. The rule-based component applies biological knowledge to filter obvious non-candidates, while the machine learning component handles the harder cases where biological variability makes rigid rules insufficient. This hybrid approach achieves the interpretability of rules with the adaptability of learned models.

The CellSearch platform as input: ALICE was designed to analyze data from the CellSearch system, the only FDA-cleared platform for CTC enumeration in clinical use. CellSearch uses fluorescent labels to tag epithelial cancer cells (E-cadherin positive) while also marking blood cells (CD45 positive), enabling the separation of tumor-derived cells from normal blood cells in fluorescence images.

Patient cohort: The system was developed and validated on blood samples from patients with pancreatic ductal adenocarcinoma (PDAC), along with healthy control donors. PDAC was chosen because of its aggressive metastatic behavior and the clinical need for non-invasive monitoring tools in this cancer type where surgical options are limited.

Training and validation strategy: ALICE was trained on expert-labeled CellSearch images, learning to replicate the cell classification decisions of experienced human analysts. The model was then validated on independent patient cohorts to assess whether its performance generalized beyond the training data and matched the consistency of expert human reviewers.

TL;DR: ALICE is a hybrid rule-based and machine learning system trained on CellSearch liquid biopsy data to automatically enumerate circulating tumor cells in pancreatic cancer patients.
Pages 8-11
An Unexpected Discovery: A New Type of Circulating Cell

The serendipitous observation: During the development and validation of ALICE, the system identified a population of cells in PDAC patient blood samples that did not fit neatly into either the CTC category or the normal blood cell category. These cells co-expressed both E-cadherin (a marker of epithelial/tumor cells) and CD45 (a marker of white blood cells), which are normally considered mutually exclusive markers.

Defining circulating hybrid cells: These unusual cells were named circulating hybrid cells (CHCs). Two subtypes were identified: CHC-1, expressing E-cadherin, CD45, and the epithelial adhesion marker EpCAM, and CHC-2, expressing E-cadherin and CD45 but lacking EpCAM. The co-expression of both epithelial and immune cell markers suggests these cells may arise from a fusion or interaction between cancer cells and immune cells, though their exact biological origin requires further investigation.

CHC-1 and lymph node metastasis: The most striking clinical finding was that the presence of CHC-1 cells in blood correlated with lymph node metastasis (N1 stage) in PDAC patients with a specificity of 1.000, meaning no false positives were observed. This is a remarkably high specificity for a liquid biopsy biomarker, suggesting CHC-1 detection could help identify patients with nodal spread without requiring invasive surgical staging.

CHC prevalence in PDAC patients: CHCs were found at higher frequencies in PDAC patients compared to healthy controls, and their presence patterns differed from conventional CTCs. This suggests CHCs represent a distinct biological phenomenon associated with pancreatic cancer progression rather than an artifact of the detection system.

TL;DR: ALICE serendipitously discovered a new class of circulating hybrid cells (CHCs) in PDAC patient blood that co-express both cancer and immune cell markers, with CHC-1 showing perfect specificity for lymph node metastasis.
Pages 12-15
Clinical Significance of Circulating Hybrid Cells

Staging pancreatic cancer non-invasively: Determining whether pancreatic cancer has spread to lymph nodes currently requires either surgical exploration or high-quality imaging, both of which have limitations. A blood test that can identify N1-stage disease with perfect specificity would be an extraordinarily valuable clinical tool, allowing oncologists to make more informed treatment decisions before surgery.

The biological mystery of CHCs: The co-expression of both epithelial and immune cell markers in CHCs raises fascinating questions about tumor biology. One hypothesis is that CHCs arise from cancer cell-macrophage fusion events, which have been described in other cancers and may facilitate immune evasion and metastasis. If confirmed, CHCs could represent both a biomarker and a mechanistic window into how pancreatic cancer evades immune destruction and spreads.

Comparison with conventional CTC analysis: While CTCs are already used clinically, their detection rates in early-stage pancreatic cancer are often low, limiting their diagnostic utility. CHCs may complement conventional CTC analysis by providing additional information, particularly in patients where CTC counts are low or ambiguous.

Prospective validation needed: The specificity of 1.000 for CHC-1 and N1 staging is based on a single cohort and requires prospective validation in larger, independent patient populations before clinical implementation. The discovery finding is compelling, but confirmation is essential before this assay could influence treatment decisions.

TL;DR: CHC-1 cells may enable non-invasive identification of lymph node metastasis in pancreatic cancer, though prospective validation in larger cohorts is required before clinical adoption.
Pages 16-20
From Automated Analysis to Biological Discovery

AI enabling unexpected scientific discovery: This study illustrates how building accurate, unbiased automated analysis systems can lead to discoveries that human observers might miss due to preconceived categories. ALICE's systematic classification of all cells, including those that did not fit expected categories, enabled the identification of CHCs as a biologically distinct population.

Reproducibility as a foundation for discovery: The ability of ALICE to consistently identify CHCs across samples and analysis runs was essential for establishing them as a real biological phenomenon rather than noise. AI reproducibility is not just a technical virtue but a prerequisite for making reliable scientific claims from complex biological data.

Implications for liquid biopsy development: The CHC finding suggests that existing liquid biopsy platforms may contain more biological information than is currently being extracted by conventional analysis approaches. Applying sensitive, unbiased AI analysis to archived CellSearch samples from other cancer types could reveal whether hybrid cells are a general feature of cancer biology or specific to PDAC.

TL;DR: ALICE demonstrates that rigorous AI automation of liquid biopsy analysis can both standardize clinical measurements and enable unexpected biological discoveries, opening new directions in pancreatic cancer research.
Citation: Open Access, 2020. Available at: PMC7532685.