NCI Data Jamboree (Project Abstract Submission): Submission #55

Presenter Information
---------------------
First Name: Giuseppe
Middle Initial: {Empty}
Last Name: Tarantino
Degree(s): Ph.D.
Position/Title/Career Status: Instructor in Medicine
Organization: Dana-Farber Cancer Institute
Organization Address:
Boston

Email: giuseppe_tarantino@dfci.harvard.edu

Additional Authors
------------------
List of Additional Authors:
- First Name: Tyler
  Last Name: Aprati
  Affiliation: Dana-Farber Cancer Institute
- First Name: Bojan
  Last Name: Karlas
  Affiliation: Dana-Farber Cancer Institute
- First Name: Paulina
  Last Name: Koehler
  Affiliation: Dana-Farber Cancer Institute
- First Name: Hyeon-Tae
  Last Name: Hwang
  Affiliation: Dana-Farber Cancer Institute
- First Name: Cathering
  Last Name: Feng
  Affiliation: Dana-Farber Cancer Institute
- First Name: Xuelu 
  Last Name: Liu
  Affiliation: Dana-Farber Cancer Institute


Abstract Information
--------------------
Abstract Category: Developing, refining, or validating tools, methods, algorithms, and pipelines
Abstract Keywords: computational pathology; whole-slide imaging; unsupervised tile clustering; tumor cell states; 
Abstract Title: A pipeline and interactive tool for identifying histomorphological tile clusters associated with tumor transcriptional states
Abstract:
Motivation: Transcriptomics has uncovered different biological tumor states that are often associated with disease outcomes and response to therapy; however, transcriptomic sequencing is rarely available in clinical practice. 
On the other hand, whole-slide images (WSIs) such as H&E are routinely performed, but are not currently used to derive tumor state information. Linking morphological patterns from WSIs with existing transcriptional programs would allow these tumor states to be identified in routine clinical practice, integrating transcriptomic findings into the precision medicine pipeline.
We aim to develop a general, end-to-end pipeline and an interactive application that quantifies and tests associations between morphological features and features including any transcriptional signature or patient survival.

Approach: For WSIs, the pipeline will embed and group images into morphological clusters to summarize each slide as a vector. For transcriptomic samples, gene expression signature scores are computed directly. We will then develop a predictive model (adjusted for clinical and genomic confounders) associating these morphological features and signature scores from the matched H&E and bulk RNA-seq samples available in TCGA.
Downstream, cluster presence will be linked to outcomes such as overall survival, therapy response. The approach is cohort- and tumor-type agnostic, providing a pipeline from histomorphology of clinical samples to associations with tumor states and survival for hypothesis generation or publication.

Implementation: The goal is to deliver this application as a simple web app. Users paste a gene list or upload a signature file. The tool computes expression scores and returns per cluster results (survival statistics, visual summaries, etc). Interchangeable configurations (slide type, cluster number, alignment, threshold) are supported, and a precomputed library of signatures is browsable alongside user-defined ones.