Shanthamallu, Uday Shankar ; Jones, Alexander ; Panés, Julián ; Corraliza Márquez, Ana Maria ; Salas, Azucena ; Akmaev, Slava ; Ghiassian, Susan Dina
Background
Non-response to TNF inhibitor (TNFi) therapies remains a challenge in UC/IBD
management. Identification of non-responders before initiating biologic therapy can reduce costs and bolster
the principles of precision medicine. In this study, we employ a network-based approach to derive biomarkers
of response and use whole blood pre-treatment gene expression data from severe UC patients to differentiate
TNFi therapy responders and non-responders.
Methods
Clinical and molecular data for a cohort of 15 UC patients with moderate-to-severe disease activity
and active ulcers at baseline were analyzed. Prior to TNFi therapy, whole blood RNAseq data were aligned to
hg38 by STAR and transcriptome profiles were generated by RSEM. The Global Mayo UC score for each
patient was recorded before and 14 weeks after initiation of a TNFi. The pre-/post-treatment difference in
Global Mayo UC score was used to assess patient clinical response. A score reduction exceeding 3 points
after the 14-week period was interpreted as a meaningful response. To obtain the top predictive biomarkers,
PRoBeNet, a previously developed network-based algorithm, was run on the map of the Human Interactome.
This identified the most important intermediary nodes between TNF (source node), drug target, and UC
disease signature cytokines (readout nodes). We evaluated the predictive power of PRoBeNet biomarkers by
running unsupervised clustering algorithm UMAP [7] on baseline transcriptomic data of UC patients and
assessing clusters and their enrichment with response clinical outcome. For patient class assignment and
calculating significance of enrichment, we implemented Hierarchical Clustering Algorithm with L1-norm as
distance metric on top two UMAP components.
Results
Implementing UMAP on baseline transcriptomic profile of the top 100 predicted biomarkers captured
two distinct subgroups in the UC patient cohort. UMAP clustering was performed with no prior information on
patient clinical outcome. Patient clinical outcomes on the UMAP visualization demonstrate a clear separation of
responders and non-responders (Figure 1A). Figure 1B shows patient class assignments as well as number of
responders and non-responders in each class. Cluster 1 is enriched with TNFi non-responders (p-value =
0.0089, Hypergeometric test).
Conclusion
Baseline whole-blood transcriptomic data contains information predictive of response to TNFi
therapy. PRoBeNet is a validated approach for identifying predictive biomarkers within transcriptomic data. The
predictive power of identified biomarkers was demonstrated in clear separation of UC responders and nonresponders.
Such predictive power is pivotal for tailoring therapeutic regimens, emphasizing the importance of
precision medicine in IBD treatment. Further studies are recommended to validate and expand upon these
findings.