Predictive Modeling for Personalized Treatment Response

This direction is split into other directions: treatment response heterogeneity. Its last version stays readable and citable here.

Version 1 of 198 papers99 conceptsfrozen 2026-10-04

In modern oncology, the most frustrating reality is the "responder gap." Whether it is immune checkpoint inhibitors (ICIs), neoadjuvant chemotherapy, or targeted radiotherapy, a therapy that is life-saving for one patient may be entirely ineffective—or even toxic—for another. As Reinhard Dummer noted in early reviews of cytokine therapy, identifying the molecular factors that predict this sensitivity is the primary hurdle to personalizing medicine.

Recent research has converged on a shared strategy: using machine learning to extract "latent" predictive signals from standard clinical data. By training models on medical imaging (radiomics), pathology slides (pathomics), and genetic signatures (multi-omics), researchers are moving away from "one-size-fits-all" guidelines toward a high-resolution prediction of individual patient outcomes.

The Macro View: Radiomics and Spatial Heterogeneity

The core intuition behind radiomics is that a tumor’s appearance on a CT, MRI, or PET scan is not just a shape, but a proxy for its underlying biology. Standard clinical assessments often look at tumor size, but radiomics extracts hundreds of quantitative features—texture, entropy, and "habitats"—that reflect the spatial heterogeneity of the tumor.

For instance, Huizhen Huang and colleagues demonstrated that DCE-MRI features can predict the expression of radioresistant factors like LRP-1 in rectal cancer. Similarly, Yuemin Wu introduced "habitat radiomics," which divides a tumor into distinct sub-regions based on imaging characteristics. This approach captures the complex environment of non-small cell lung cancer (NSCLC) more effectively than treating the tumor as a single mass, allowing for better prediction of both survival and adverse immune reactions.

The power of this method lies in its non-invasive nature. Zepang Sun describes this as a "digital biopsy," where CT scans are used to map the lymphoid and myeloid cell landscape in gastric cancer, identifying responders to immunotherapy without the need for repeated, painful tissue sampling.

The Micro View: Pathomics and the Cellular Landscape

While radiology looks at the whole organ, pathomics applies deep learning to standard H&E-stained biopsy slides. The intuition here is that the spatial arrangement of cells—the "architecture" of the disease—contains clues about the immune system's engagement with the tumor.

Daniela Allende developed the ABRS-P model, which uses AI to estimate a complex 10-gene expression signature directly from inexpensive biopsy images. This bypasses the need for costly genomic sequencing. Similarly, Fabien Lamaze showed that extracting texture features from cell density maps on pathology slides can predict PD-L1 expression and progression-free survival in lung cancer patients. These models work because they "see" patterns of immune infiltration and stromal architecture that are invisible to the human eye but critical for therapeutic success.

The Molecular View: Signatures of Resistance and Sensitivity

Beyond imaging, researchers are identifying specific molecular "gas pedals" and "brakes" that dictate response. Chaoqi Zhang developed a Costimulatory Molecule-based Signature (CMS) focusing on the B7-CD28 and TNF gene families, which effectively identifies high-risk lung cancer patients who are paradoxically the most likely to benefit from immune checkpoint inhibitors.

The search for these biomarkers often reveals surprising biological insights. For example, Pauline Hamon found that TREM2-expressing macrophages, usually considered immunosuppressive, are actually associated with better responses to PD-1 blockade in liver cancer. This suggests that the "hot" or "cold" status of a tumor microenvironment—driven by specific immune engagement archetypes—is more predictive than simple mutation counts, a concept formalized in the IMIREG 14-regulon signature across multiple cancer types.

Integration: From Data to Decision

The ultimate goal of this research is the synthesis of these disparate data points into clinical tools. Lizhi Shao pioneered a "radiopathomic" approach, merging macro-scale MRI data with micro-scale biopsy features to achieve nearly 88% accuracy in predicting responses to chemoradiotherapy.

This shift toward algorithmic decision-making is perhaps best exemplified by the Helicobacter pylori AI-Clinician, which uses reinforcement learning to tailor antibiotic combinations to individual patient profiles. By moving from static guidelines to dynamic, data-driven models, the field is beginning to solve the "responder gap," ensuring that patients receive the treatments they are most likely to benefit from while avoiding the toxicity of those they are not.

Papers in this version

Showing 60 of 98 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_predictive_modeling_for_personalized_tre_edb46ec6,
  title = {Predictive Modeling for Personalized Treatment Response},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/predictive-modeling-for-personalized-treatment-response/rdr_edb46ec67ee44461a0ea6f126dbb40d1},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_predictive_modeling_for_personalized_tre_edb46ec6_v1,
  title = {Predictive Modeling for Personalized Treatment Response},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/predictive-modeling-for-personalized-treatment-response/rdr_edb46ec67ee44461a0ea6f126dbb40d1/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}