Scholarlink

Background <p>There is a need for more accurate prognosis in multiple myeloma given the current heterogeneity in survival outcomes. In other cancers, the self-complementary nature of multi-omics is driving to the adoption of multi-omics data in personalized risk models. Simultaneously, deep learning is emerging as a top performing algorithm in risk prediction. These trends motivate us to assess the potential benefits of risk prediction with deep learning and multi-omics in the context of multiple myeloma.</p> Methods <p>We introduce MyeVAE (Myeloma VAE), a personalized multi-omics risk model for newly diagnosed multiple myeloma. MyeVAE is obtained by extending the traditional Variational Auto-Encoder (VAE) to jointly model an arbitrary combination of data modalities. MyeVAE incorporates a sub-task network (“survival network”) to generate risk predictions from omics-derived vector embeddings and directly observed clinical traits. We optimize MyeVAE using semi-supervised learning, combining the traditional evidence lower bound objective with the generalized version of Cox’s partial likelihood to guide the orientation of the latent space along risk-related axes.</p> Results <p>MyeVAE outperforms other algorithms at generalizing to microarray-based gene expression cohorts when trained on RNA-Seq gene expression. MyeVAE achieves a consistent validation C-index of 0.7 on overall survival when trained with multi-omics data. Gene expression emerged as the omics modality with highest predictive value across the four risk algorithms and six omics modalities we tested. Using SHAP, MyeVAE is found to recapitulate existing prognostic events, such as Gain(1q), t(4;14), and APOBEC activity. SHAP also hinted at the 26&#xa0;S proteasome regulatory subunit PSMD4 to be a key gene implicated in Gain(1q), which was corroborated by validation on four external cohorts.</p> Conclusions <p>We demonstrate the marginal benefits of including multi-omics data—especially RNA-Seq based gene expression—across MyeVAE and other risk algorithms. Our experiments with MyeVAE demonstrate the potential for deep learning-based approaches for risk prediction and biomarker discovery in cancers where omics data is available.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

MyeVAE: a multi-modal variational autoencoder for risk profiling of newly diagnosed multiple myeloma

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年8月28日

查看全文
  • Jia Geng Chang,
  • Jianbin Chen,
  • Guo-Liang Chew,
  • Wee Joo Chng

摘要

Background

There is a need for more accurate prognosis in multiple myeloma given the current heterogeneity in survival outcomes. In other cancers, the self-complementary nature of multi-omics is driving to the adoption of multi-omics data in personalized risk models. Simultaneously, deep learning is emerging as a top performing algorithm in risk prediction. These trends motivate us to assess the potential benefits of risk prediction with deep learning and multi-omics in the context of multiple myeloma.

Methods

We introduce MyeVAE (Myeloma VAE), a personalized multi-omics risk model for newly diagnosed multiple myeloma. MyeVAE is obtained by extending the traditional Variational Auto-Encoder (VAE) to jointly model an arbitrary combination of data modalities. MyeVAE incorporates a sub-task network (“survival network”) to generate risk predictions from omics-derived vector embeddings and directly observed clinical traits. We optimize MyeVAE using semi-supervised learning, combining the traditional evidence lower bound objective with the generalized version of Cox’s partial likelihood to guide the orientation of the latent space along risk-related axes.

Results

MyeVAE outperforms other algorithms at generalizing to microarray-based gene expression cohorts when trained on RNA-Seq gene expression. MyeVAE achieves a consistent validation C-index of 0.7 on overall survival when trained with multi-omics data. Gene expression emerged as the omics modality with highest predictive value across the four risk algorithms and six omics modalities we tested. Using SHAP, MyeVAE is found to recapitulate existing prognostic events, such as Gain(1q), t(4;14), and APOBEC activity. SHAP also hinted at the 26 S proteasome regulatory subunit PSMD4 to be a key gene implicated in Gain(1q), which was corroborated by validation on four external cohorts.

Conclusions

We demonstrate the marginal benefits of including multi-omics data—especially RNA-Seq based gene expression—across MyeVAE and other risk algorithms. Our experiments with MyeVAE demonstrate the potential for deep learning-based approaches for risk prediction and biomarker discovery in cancers where omics data is available.