AML has been extensively studied in a gene-to-gene and pathway-to-pathway fashion over the years, unraveling insightful local patterns that capture heterogeneity in patients and identify potential drug targets. However, our understanding of AML from a global and systems perspective is still lacking. A global proteomic pathway map is yet to be drawn to integrate local activity patterns and to translate patient classifications across pathways. This will not only improve our scientific understanding of how different functional pathways are inter-related, but will also enable us to develop more robust and effective therapeutic regimens based on pathway cross-talks.
A proteomic profile, containing the expression pattern of 231 proteins in each of the 415 newly diagnosed AML patients at UT MD Anderson Cancer Center, was generated using the Reverse Phase Protein Array (RPPA) technology. We grouped these proteins into 23 functional pathways based on protein association known in literature and correlation shown in the proteomic profile. Principal component analysis and scree plot were used to refine the pathway construction. The AML patients were clustered by their protein expression patterns within each individual pathway, and were then compared across pathways. The association of patient clusters between different pathways was measured by Goodman-Kruskal's (GK) tau method, indicating the predictability of patient clustering in one pathway given that in the other. This association between pathways and interchangeability of patient groupings were visualized in a circos plot (Figure 1), depicting a global proteomic pathway map.
The global proteomic pathway map illustrates how strongly protein expression patterns of different pathways are associated, and how patient classifications under different pathways could be translated from one to another. Here, we highlight some of the key insights surfaced from this analysis. First, we identified ‘social' pathways that have intensive cross-talks with multiple other pathways, including some of the cell signal transduction pathways (MEK, PI3K, mTOR), genetic information processing pathways (transcription, histone methylation), and cell survival/death pathways (apoptosis, autophagy). We also identified ‘orphan' pathways that are more independent and are poorly associated with others. These include a subset of signal transduction pathways (pkc, tp53, S6rp, Src, Creb, Wnt), cytoskeleton and differentiation. As the association is directional, each pathway could be further characterized as either a ‘sender' or a ‘receiver' pathway based on whether it is acting more as the origin or the target of the link. The patient clusters from the ‘sender' pathways (e.g. Apoptosis, mTOR, Fli), could be easily translated to other pathways, while the patient clusters in ‘receiver' pathways (e.g. Hippo and Transcription), are highly predictable by patient clusters from multiple other pathways. We further constructed and compared the global pathway maps for patients in different cytogenetic groups. Comparison of pathway maps from patients with favorable, intermediate and unfavorable cytogenetics shows the power of this methodology to discern differences in the degree of correlation between protein functional groups. Favorable cytogenetics (T8;21) and inversion 16, because they are more similar have less patient to patient variation and thus have a more consistent and highly correlated pathway map with a higher number of connections.
Based on the RPPA data in AML patients, we built a global proteomic pathway map that captures the association between protein expression patterns in defined protein functional groups. We identified intensive interacting pathways as well as independent pathways, which indicate potential hubs and modulators of leukemic cell behavior. We further compared maps of different cytogenetic groups and revealed different correlation mappings. We are further refining the algorithms in order to study more focused changes within lower population subsets. Ultimately we believe that this will enable the matching of targeted agents to specific settings where the target is expressed and highly interactive based on proteomic data.
No relevant conflicts of interest to declare.
Asterisk with author names denotes non-ASH members.