(P 295) DEFINING OSTEOSARCOMA HETEROGENEITY: CONSENSUS CLUSTERING REVEALS GENE EXPRESSION SUBTYPES

Objective: Metastasis is a major clinical challenge in osteosarcoma (OS), contributing to significant morbidity and mortality. Approximately 50% of osteosarcoma (OS) cases metastasize within the first years. Patients who present with metastatic disease at diagnosis often have poor prognosis and survival. The biological behavior of metastasis differs from that of the primary tumor with respect to metabolism, karyotype, cell cycle states, and cell motility. These differences are dictated by differentially expressed genes and shifting molecular profiles. Although differential gene expression analysis studies have been conducted on OS tumor samples relative to normal human osteoblasts, no published findings comparing matched primary tumors and metastases in the same patient exist. Furthermore, no study to date has identified metastasis-associated gene expression events. Based on the work described in Sheng et al. 2021 and Ludwig et al. 2021, the biological behavior of primary and metastatic tumors differ significantly. We can uncover transcriptome features in primary osteosarcoma tumors that differ from metastatic tumors and yield candidate targets for novel therapeutic design. Our study seeks to identify the transcriptional differences between primary and metastatic tumors. Using machine learning techniques, we hypothesize that the metastatic sample cluster's transcriptional landscape differs from the matched primary tumor based on gene expression data. This knowledge will provide important insight into the process of metastatic development needed to advance therapeutic design.

Methods: Our study is designed to systematically assess the differences between primary and metastatic tumors across patients using machine-learning approaches. We applied two unsupervised machine-learning clustering algorithms, Consensus Clustering and K-Mean Clustering, which were implemented in R Bioconductor. Consensus clustering enhances the robustness and reliability of clustering results by integrating multiple clustering solutions, re-sampling, and stability analysis. This technique involves performing repeated clustering runs on resampled subsets of the gene expression data using various algorithms or initial conditions. Each clustering run produces a different partition of the genes, and the results are integrated into a consensus matrix, where each element is clustered together across all runs. By aggregating these results, the consensus matrix captures the stability of gene clusters, allowing for the identification of robust gene groups that consistently co-cluster. A final clustering algorithm is applied to this matrix to derive the ultimate clustering solution. This approach mitigates the variability inherent in individual clustering runs, leading to more stable and biologically meaningful clusters. K-means clustering is a robust analytical method applied to gene expression data to identify groups of co-expressed genes. In this approach, gene expression levels across various conditions or time points are first normalized to ensure comparability. The algorithm then partitions the genes into K clusters by iteratively assigning each gene to the nearest cluster centroid based on Euclidean distance and updating the centroids to reflect the mean expression profile of the genes in each cluster. This process repeats until the clusters stabilize. The resulting clusters group together genes with similar expression patterns, which can be linked to specific biological processes or pathways. This technique is particularly valuable in genomics for uncovering co-regulated genes, identifying disease subtypes, and analyzing temporal expression patterns, ultimately contributing to a deeper understanding of the molecular mechanisms underlying biology. We have consolidated OS data from TARGET, Sweet-Cordero Laboratory at UCSF, Stanford, the Hospital for Sick Children, and the St. Jude Cloud, resulting in one of the largest collections of OS transcriptome data. The OS dataset is composed of 202 OS tumor samples sequenced by the Illumina polyA capture method. Among the 202 OS samples, 164 samples stem from primary tumors, and 38 stem from metastatic tumors obtained at diagnosis. All samples have been uniformly processed using the same bioinformatic pipelines to allow for combined analysis. We also compiled associated clinical and genetic information for these samples, including age, sex, histology, and response to treatment. Within this large cohort, nineteen patients have matched primary and metastatic samples.

Results: Using our complete cohort of primary and metastatic OS samples (N=202), we have determined that our metastatic samples cluster apart in the transcriptional space. To do this, we applied Consensus Clustering to determine quantitative evidence for the number of transcriptome clusters and the membership of possible clusters utilizing the consensus clustering algorithm (Figure 1A.) We observed three discrete membership events observed in the dendrogram consensus matrix (Figure 1A.) resulting from distinct transcriptome differences in primary and metastatic tumors. Within the three clusters are two distinct metastatic clusters: cluster 1 (light blue) and cluster 3 ( light green). Cluster 2 (dark blue) describes the class membership of only the primary samples. This preliminary result describes differences between primary and metastatic tumors. It sets the stage for evaluating whether it is possible to identify metastatic signatures within the transcriptomic data from primary tumors.

We utilized a secondary unsupervised clustering approach, K-means clustering (Figure 1B) of primary and metastatic samples (N=202 tumor samples). This approach determines the optimal number of clusters using three statistical analysis metrics: the elbow, average silhouette, and gap methods. In Figure 1B, the Average Silhouette Method was applied to determine the number of clusters and the stability of the clustering solution. The optimal number of clusters for the OS samples (N=202) was selected as k=3. The consensus among all statistical analyses used in the K-means approach revealed that the optimal number of clusters is k=3, representing three transcriptionally distinct clusters. Clusters generated by K-mean clustering where the same clusters generated by consensus clustering. There is sample congruency in each cluster between both clustering approaches.

Within the dataset of 202 OS tumor samples, there were five samples that matched primary and metastatic tumors from the same patient. Rather than the matched OS samples clustering within the same cluster, the matched primary samples cluster with the other non-matched primary samples (N=159, primary non-matched samples).

Conclusion: We demonstrate that unsupervised clustering methods are sensitive enough to detect distinct transcriptional membership between primary and metastatic tumors (Figure 1), lending to the idea that distinct transcriptional differences exist between primary and metastatic osteosarcoma tumors. We have identified three gene expression clusters, one with primary tumors and two with metastatic tumors. Future work will focus on identifying features of the tumors belonging to each cluster.