Abstract
Extensive research has examined socioeconomic factors influencing prostate cancer (PCa) disparities. However, to what extent molecular and genetic mechanisms may also contribute to these inequalities still remains elusive. Although various in vitro, in vivo, and population studies have originated to address this issue, they are often very costly and time-consuming by nature. In this work, we attempt to explore this problem by a preliminary study, where a joint deep latent variable model (DLVM) is proposed to in silico quantify the personalized and race-specific effects that a genomic aberration may exert on the Gleason Score (GS) of each individual PCa patient. The core of the proposed model is a deep variational autoencoder (VAE) framework, which follows the causal structure of inference with proxies. Extensive experimental results on The Cancer Genome Atlas (TCGA) 270 European-American (EA) and 43 African-American (AA) PCa patients demonstrate that ERG fusions, somatic mutations in SPOP and ATM, and copy number alterations (CNAs) in ERG are the statistically significant genomic factors across all low-, intermediate-, and high-grade PCa that may explain the disparities between these two groups. Moreover, compared to a state-of-the-art deep inference method, our proposed method achieves much higher precision in causal effect inference in terms of the impact of a studied genomic aberration on GS. Further validation on an independent set and the assessment of the genomic-risk scores along with corresponding confidence intervals not only validate our results but also provide valuable insight to the observed racial disparity between these two groups regarding PCa metastasis. The pinpointed significant genomic factors may shed light on the molecular mechanism of cancer disparities in PCa and warrant further investigation.
Introduction
Prostate cancer (PCa) exhibits a stark racial disparity in the United States: African-American (AA) men face a 60% higher incidence rate and a 2.4 times higher mortality rate compared to European-American (EA) men. While socioeconomic factors like healthcare access and diet play a significant role, researchers suspect intrinsic molecular differences also drive these inequalities.
The challenge lies in quantifying these effects. Traditional in vivo or population studies are prohibitively expensive and time-consuming. Furthermore, genomic data is inherently complex; it is often difficult to determine if a mutation caused a high Gleason Score (a measure of cancer aggressiveness) or if both were influenced by unmeasured "latent" factors, such as complex gene-to-gene interactions. This study seeks to isolate the personalized, race-specific causal impact of genomic aberrations using a novel machine learning approach.
Methodology
The authors propose a Deep Latent Variable Model (DLVM) built on a Variational Autoencoder (VAE) framework. Standard statistical models often struggle with "confounders"—variables that influence both the treatment (a genetic mutation) and the outcome (Gleason Score).
The DLVM addresses this by:
- Modeling Latent Variables: Using the VAE to infer "hidden" influences (like gene crosstalk) that aren't explicitly recorded in the dataset.
- Counterfactual Inference: To find the causal effect, the model estimates what would have happened in the opposite scenario (e.g., predicting the Gleason Score of an AA patient if they hadn't had an ERG fusion).
- Architecture: The model uses two deep VAEs with three hidden layers each, utilizing Parametric Rectified Linear Units (PReLU) and the ADAM optimizer to process high-dimensional multi-omics data including mutation profiles, gene expressions, and clinical features.
Results and Discussion
The study utilized the TCGA "Prostate Adenocarcinoma" (PRAD) dataset, comprising 270 EA and 43 AA patients. The DLVM achieved significantly higher precision in causal inference than the previous state-of-the-art method (CEVAE), reducing the Root Mean Square Error (RMSE) by approximately 19.03% for AA patients and 22.98% for EA patients.
Key scientific insights include:
- The "Low-Grade" Danger: In low-grade PCa (GS ≤ 6), the causal effects of mutations like ERG fusions and SPOP were significantly higher in AA men than EA men. This suggests that AA men with "low-risk" scores may actually be at higher biological risk than their EA counterparts.
- ATM as an EA Driver: Conversely, somatic mutations in the ATM gene showed a statistically higher causal impact on Gleason Scores in European-American men across all grades.
- Identified Drivers: The study pinpointed ERG fusions, somatic mutations in SPOP and ATM, and copy number alterations (CNAs) in ERG as the most statistically significant genomic factors explaining the disparities.
Conclusion
This work demonstrates that deep latent variable models can reliably capture biological signals and quantify causal effects even with relatively small sample sizes for minority groups. The pinpointed genomic factors shed light on the molecular mechanisms of cancer disparities in PCa. Future work will aim to integrate socioeconomic and environmental factors as additional confounders and explore the combinatorial effects of multiple mutations interacting simultaneously.
Summary updated