Scholarlink

<p>Tackling cancer with effective therapies is a major challenge our society faces with significant costs and complexities of designing and validating treatments. In this article we show that representation learning based on real-world data can suggest indications for which a chosen mechanism of action is likely to be effective. We trained a machine learning model to rank histology-based malignant indications against expected biological relevance of anti-PD-1 treatment, leveraging reference indications and an embedding generating approach for features based on events in the patient’s medical journey. We call our approach INSPIRE. When restricting the model to data before broad establishment of PD-1 inhibitors in the clinic, the method successfully prioritizes 70% of subsequent approvals. This indicates that INSPIRE could accelerate the process of selecting, testing and eventually treating patients with effective drugs, therefore reducing costs and improving care for patients.</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

Finding the most promising indications for novel treatments in oncology

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月25日

查看全文
  • Maren Eckhoff,
  • Stefan Klingelschmitt,
  • Lisbeth Van Ruijssevelt,
  • Sebastian Grossmann,
  • Thomas Zichner,
  • Paul Stümpges,
  • David Roschewitz,
  • Bernhard Mlecnik,
  • Michael Merger,
  • Patrick Graen,
  • Jeevan Kumar,
  • Alex Devereson,
  • Stephan Wurzer,
  • Joachim Bleys,
  • Björn Albrecht,
  • Daniel Hach,
  • Christian Haslinger

摘要

Tackling cancer with effective therapies is a major challenge our society faces with significant costs and complexities of designing and validating treatments. In this article we show that representation learning based on real-world data can suggest indications for which a chosen mechanism of action is likely to be effective. We trained a machine learning model to rank histology-based malignant indications against expected biological relevance of anti-PD-1 treatment, leveraging reference indications and an embedding generating approach for features based on events in the patient’s medical journey. We call our approach INSPIRE. When restricting the model to data before broad establishment of PD-1 inhibitors in the clinic, the method successfully prioritizes 70% of subsequent approvals. This indicates that INSPIRE could accelerate the process of selecting, testing and eventually treating patients with effective drugs, therefore reducing costs and improving care for patients.