1197 Harnessing the cancer immunity cycle via machine learning models to generate novel strategies for personalized cancer therapy

Farnoosh Abbas Aghababazadeh, Luis Javier Alonso, Marcos López-De-Castro, Florent Guinot, Jonathan Desponds, Stephan Furger, Armen Karapetyan, Sascha Ott, Ewa A. Bergmann, Neil D. Jones, G. Vasudevan, Victor Sankarlingam, Shivapriya Chithambaram, Pascal Gautschi, G. Scott Chandler, Rajat Mohindra, Rachel Garonce-Hediger, Magnus Fontes, Vaios Karanikas, Marta Crespo, Cameron Ross MacPherson, Benjamin Haibe‐Kains, Mikel Hernáez, Rubén Armañanzas, Bernadett SzabadosRegular 2024 · DOI

TL;DR

This study leverages machine learning to dissect the complex cancer immunity cycle (CIC) by identifying key gene signatures associated with treatment response. By developing "Immune FITness" (IFIT) scores, they demonstrate a significant link between specific CIC stages and patient survival, paving the way for more personalized cancer therapies.

What Problem Does This Paper Address?

The cancer immunity cycle (CIC) is a conceptual framework that describes the multi-step process by which the immune system recognizes and eliminates cancer cells. While this model is crucial for understanding immunotherapy, precisely quantifying the contribution of each stage and how gene expression within these stages impacts patient outcomes has been challenging. This lack of precise quantification hinders the development of personalized treatment strategies that can effectively target specific points in the CIC. This paper aims to bridge this gap by developing data-driven models to identify and quantify CIC stage-specific gene expression patterns that predict treatment response and survival.

Methods

The researchers employed a two-pronged approach to model CIC response:

  1. Gene-Specific Effects (Approach 1): They started with a large pool of ~4000 genes associated with the CIC. Using a technique called Conformal Recursive Feature Elimination (CRFE), they refined this pool to identify 100 "RECIST-discriminating" marker genes. CRFE is a method for feature selection that aims to find a minimal set of features that can accurately predict a target variable (in this case, treatment response categories like Complete Response, Partial Response, Stable Disease, and Progressive Disease). This approach focuses on identifying individual genes that are strong indicators of response.
  2. Multi-Gene Effects (Approach 2): This approach focused on capturing the collective activity of genes within CIC stages. They inferred Immune FITness (IFIT) profiles, which are essentially activity scores for each CIC stage, derived from gene expression data. These IFIT scores were then used in Cox regression models to assess their association with overall survival (OS). This method aims to understand how the combined activity of genes within a CIC stage influences patient outcomes.

Additionally, they mapped the biological pathways associated with these CIC stages using validated GENCODE mappings to provide biological context.

Datasets

The study utilized anonymized data from four completed Roche atezolizumab clinical trials:

  • IMvigor210: Advanced urothelial carcinoma (UC) patients (N=1540).
  • IMpower130: Metastatic non-small cell lung cancer (NSCLC) patients (N=1926).

These datasets provided the necessary gene expression and clinical outcome data (overall survival, RECIST response) for training and validating their models. The paper does not introduce new benchmark datasets but rather leverages existing clinical trial data.

Key Insights and Ideas

  • CIC is Quantifiable: The study demonstrates that the complex cancer immunity cycle can be quantitatively analyzed using gene expression data, moving beyond a purely conceptual model.
  • Stage-Specific Signatures Exist: They successfully identified specific gene signatures that are differentially expressed across CIC stages and are predictive of treatment response. This suggests that targeting specific stages of the CIC might be a viable therapeutic strategy.
  • IFIT Scores as Prognostic Markers: The development of IFIT scores provides a novel way to measure the "immune fitness" within each CIC stage. The significant association of IFIT scores for stages 5C and 6 with overall survival is a key finding, highlighting the prognostic power of these scores.
  • Biological Pathways Inform CIC Stages: Linking CIC stages to specific biological pathways (e.g., antigen processing, PI3K-Akt signaling) provides a deeper understanding of the underlying mechanisms driving immune response and treatment outcomes.

Key Results

  • Reduced Gene Pool: The initial 4000 CIC-associated genes were reduced to 100 RECIST-discriminating marker genes using CRFE.
  • Predictive Accuracy: The gene-specific approach achieved:
  • UC: 73% accuracy and 60% AUC for predicting patient response.
  • NSCLC: 72% accuracy and 61% AUC for predicting patient response.
  • Survival Association: IFIT profiling showed a significant association (p<0.01) of CIC stages 5C and 6 with overall survival. These stages are linked to T-cell infiltration and intra-tumoral cancer recognition.
  • Pathway Enrichment: CIC stage 6 was associated with antigen processing and presentation, while CIC stage 5 was linked to PI3K-Akt signaling, cytokine-cytokine receptor interaction, chemokine signaling, Wnt signaling, and T-cell receptor signaling pathways.

Limitations and Future Work

  • Validation Ongoing: The authors state that validation is ongoing and they plan to expand the analysis to other cancer indications and earlier disease settings.
  • Complexity of CIC: While progress has been made, the authors acknowledge that untangling the full complexity of the CIC is essential for its advancement as an analytical tool.
  • Potential for Further Refinement: The study identified robust signatures, but further investigation into the identified stage-specific markers is warranted. The differences observed between the gene-specific and multi-gene approaches suggest that further exploration of shared biological processes across stages could yield deeper insights.

Summary updated

Abstract

<h3>Background</h3> The cancer-immunity cycle (CIC) provides a staged framework for understanding the interplay between immune response and disease (figure 1).<sup>1</sup> Precise quantification of gene expression in each stage and contribution to physiological effects remains elusive. This study aims to personalize CIC profiles and pinpoint CIC stages informing precision therapy and diagnostics. <h3>Methods</h3> Data was anonymized from prospective, randomized trials exploring atezolizumab in advanced urothelial (UC; NCT02951767, NCT02108652 &amp; NCT02302807, n = 1540) and non-small cell lung cancer (NSCLC; NCT02366143 &amp; NCT02367781, n = 1926). The objective was to model CI response outcomes according to CIC stage-specific gene expression profiles using two complementary approaches: Approach 1 assessed gene-specific effects. Genes were mapped to CIC stages after recursive RECIST classification<sup>2</sup> to identify marker genes as putative drivers (figure 2). Approach 2 assessed multi-gene effects. Immune FITness profiles (IFIT) were inferred as CIC stage-specific activity scores<sup>3</sup> followed by Cox regression against overall survival (OS; figure 3). The underlying biology of CIC stages were explained by mapping to validated GENCODE mappings (figure 4).<sup>4</sup> <h3>Results</h3> In the first approach we reduced 4000 CIC associated genes to 100 RECIST-discriminating markers including TRNT1, CD74, PTN, PHYH, ENO3, ITGA6, SERPINE1, and UNC5B differentially expressed in UC; and DHRS2, JAK3, IFNG, and ERBB3 in NSCLC (𝛂=.05; figure 2). These markers highlight stages 1, 2, 5, 5A/C, 6 and correctly classified UC patient response at 73% accuracy and 60% AUC; whereas, for NSCLC, 72% accuracy and 61% AUC was achieved. In the second approach IFIT profiling showed significant association of CIC-stages 5C and 6 to OS and describing T-cell infiltration via TLA and intra-tumoral cancer recognition, respectively (p&lt;0.01) (figure 3). GENCODE signatures highlighted key factors in antigen processing and presentation for CIC stage 6, as well as PI3K-Act, cytokine-cytokine receptor interaction, chemokine, Wnt, and T-cell receptor signaling pathways in CIC stage 5 (figure 4). <h3>Conclusions</h3> This study identifies robust gene signatures and demonstrates the potential role of the IFIT score for stage-specific risk stratification in cancer treatment. Gene-specific methods identified CIC stage-specific markers warranting further investigation and corroborate our multi-gene approach with differences likely arising due to shared biological processes across stages. Untangling this complexity will be essential to move the CIC framework forward as an analytic tool for precision therapy and diagnostics. Validation is ongoing with expansion to other indications and earlier disease settings. <h3>Acknowledgements</h3> This research was a collaborative effort of the imCORE Network, made possible through support and funding from F. Hoffmann-La Roche and Institut Roche, France. M. L-d-C. and R. A. were supported by the Gobierno de Navarra through projects ANDIA 2021 no. 0011-3947-2021-000023, and ERA PerMed JTC2022 PORTRAIT no. 0011-2750-2022-000000; M.H. and L.A. supported by RYC2021- 033127-I MCIN/AEI/10.13039/501100011033 y por la Unión Europea ‘NextGenerationEU’/PRTR and Gobierno de Navarra, Proyectos Estratégicos GRANATE. <h3>Trial Registration</h3> NCT02108652, NCT02302807, NCT02366143, NCT02367781. <h3>References</h3> Mellman I, Chen DS, Powles T, Turley SJ. The cancer-immunity cycle: indication, genotype, and immunotype. <i>Immunity</i>. 2023 Oct;<b>56</b>(10):2188–2205. López-De-Castro M, García-Galindo A, Armañanzas R. Conformal recursive feature elimination.<i> arXiv</i>. 2024;2405.19429:1–35. Available from: https://arxiv.org/abs/2405.19429 Ruiz-Arenas C, Marín-Goñi I, Wang L, Ochoa I, Pérez-Jurado LA, Hernaez M. NetActivity enhances transcriptional signals by combining gene expression into robust gene set activity scores through interpretable autoencoders. <i>Nucleic Acids Res</i>. 2024 May 22;<b>52</b>(9):e44. Available from: https://doi.org/10.1093/nar/gkae197. Bareche Y, Kelly D, Abbas-Aghababazadeh F, Nakano M, Esfahani PN, Tkachuk D, Stagg J. Leveraging big data of immune checkpoint blockade response identifies novel potential targets. <i>Ann Oncol</i>. 2022;<b>33</b>(12):1304–1317. <h3>Ethics Approval</h3> Patients included in this analysis were enrolled in Roche sponsored studies with atezolizumab and the informed consent obtained for study participation had provision for secondary use of data.
BibTeX
@inproceedings{aghababazadeh2024harnessing,
  author = {Farnoosh Abbas Aghababazadeh and Luis Javier Alonso and Marcos López-De-Castro and Florent Guinot and Jonathan Desponds and Stephan Furger and Armen Karapetyan and Sascha Ott and Ewa A. Bergmann and Neil D. Jones and G. Vasudevan and Victor Sankarlingam and Shivapriya Chithambaram and Pascal Gautschi and G. Scott Chandler and Rajat Mohindra and Rachel Garonce-Hediger and Magnus Fontes and Vaios Karanikas and Marta Crespo and Cameron Ross MacPherson and Benjamin Haibe‐Kains and Mikel Hernáez and Rubén Armañanzas and Bernadett Szabados},
  title = {{1197 Harnessing the cancer immunity cycle via machine learning models to generate novel strategies for personalized cancer therapy}},
  booktitle = {Regular},
  year = {2024},
  doi = {10.1136/jitc-2024-sitc2024.1197},
}
Figure