← Research

Doctoral research · Published foundation

Causal Network Inference in Tumor Microenvironments

I co-developed a method for inferring heterocellular networks from observational tumor measurements and used it to examine how oncogenic CCN4 expression is associated with changes in local cellular relationships in breast cancer and melanoma.

Causal inferenceBayesian networksComputational oncologyEdge stability

Scientific problem

Tumor behavior depends on relationships among malignant, immune, stromal, and vascular states.

Tumors are not collections of malignant cells acting independently. Their behavior depends on interactions among tumor, immune, stromal, and vascular states. Yet many mechanistic network models begin from structures assembled manually from selected literature.

During my doctoral research, I co-developed an approach for inferring heterocellular networks from observational tumor measurements. The method represents cell states and tissue features as nodes and estimates directed relationships among them. Edge direction, sign, and support summarize how the local cellular system is organized under the assumptions of the structure-learning procedure.

Inferred heterocellular network in invasive breast cancer from TCGA BRCA data
Inferred heterocellular network in invasive breast cancer from TCGA BRCA data (n=582). Nodes represent tumor, immune, stromal, and tissue states. Directed edges summarize supported promoting or inhibiting relationships, and line width represents edge confidence. Adapted from Klinke et al., Nature Communications, 2022, under the article's CC BY 4.0 license.

Published result

Oncogenic expression was analyzed as a change in local system structure.

We applied the method to human breast cancer and melanoma data to examine how oncogenic CCN4 expression is associated with the surrounding cellular network. The resulting models connected tumor state with immune and stromal features, allowing the oncogenic signal to be interpreted as a change in local organization rather than as an isolated molecular marker.

The inferred graph is conditional on the measured variables, data quality, and assumptions of the structure-learning procedure. It identifies relationships that can guide mechanistic modeling and experimental follow-up, but it does not establish intervention effects.

CCN4

Placed within a network of tumor and tissue-state relationships.

Immune states

Represented as interacting components rather than isolated abundance estimates.

Stromal context

Included in the same directed system as tumor and immune features.

Edge support

Used to distinguish stronger structural evidence from uncertain relationships.

Methodological extension

BaMANI reports disagreement among plausible network structures.

The doctoral work motivated an ensemble framework for causal structure learning. Rather than relying on one algorithm and one fitted graph, BaMANI compares evidence across algorithms and resampled datasets, then summarizes edge-level stability.

Observational dataCell and tissue-state variables
Multiple algorithmsConstraint and score-based structure learning
Bootstrap resamplingRepeated evidence under data variation
Consensus networkEdge-level support and uncertainty
Appropriate claim

Causal structure learning can prioritize relationships for mechanistic testing. It does not replace intervention studies, and its interpretation remains dependent on identification assumptions.

Takeaway

Interpret biomarkers within the structure of the surrounding cellular system.

The work produced interpretable heterocellular networks in human cancer datasets and established a methodological direction for reporting structural uncertainty. Its strongest use is to identify relationships that warrant mechanistic testing while keeping uncertainty in the graph visible.