In clinical research, a fundamental tension exists between two types of biological signals. Imagine a patient with a highly aggressive tumor. A "prognostic" biomarker tells us this patient has a poor outlook regardless of what we do. A "predictive" biomarker, however, tells us that despite the aggression, a specific drug will work for this individual.
The challenge is that these two signals are often entangled. As Shuhan Xiao and colleagues point out in Enhancing predictive imaging biomarker discovery through treatment effect analysis, general prognostic signals are often much stronger than treatment-specific ones. This "loudness" causes standard machine learning models to latch onto disease severity rather than therapeutic sensitivity, leading to a "reality check" where models fail to generalize to the clinic.
The Prognostic-Predictive Divide
To build intuition, we must distinguish between the baseline risk and the treatment interaction:
- Prognostic Biomarkers: These reflect the natural history of the disease. For example, Ulrich Hacker found that high levels of Angiopoietin-2 in gastric cancer patients strongly predict poor survival, but they do not predict whether the patient will benefit from the drug bevacizumab. The marker is a "sentinel" for aggressive disease, not a guide for therapy selection.
- Predictive Biomarkers: These identify specific subgroups likely to benefit from an intervention. Alexander Baras identified a two-protein signature (GDPD3 and SPRED1) that specifically forecasts resistance to platinum chemotherapy in bladder cancer. This allows clinicians to spare resistant patients from unnecessary toxicity.
As Shengri Tian notes in a review of prostate cancer, the field has many robust prognostic indicators (like PSA levels) but a critical lack of validated predictive markers that tell a clinician exactly which drug to use.
Why Models Conflate the Two
The primary reason models fail to separate these signals is the "prognostic noise" problem. In Overview and practical recommendations on using Shapley Values for identifying predictive biomarkers via CATE modeling, David Svensson and Ilya Lipkovich explain that standard models often misidentify a general survival signal as a causal treatment interaction.
This is exacerbated by study design. Xin Huang and Sven H. Loosen both highlight that without a control group (patients who did not receive the treatment), it is mathematically impossible to know if a biomarker is predicting the drug's success or just a less aggressive tumor phenotype.
Mechanisms for Isolation
To solve this, researchers are moving toward Conditional Average Treatment Effect (CATE) estimation. Rather than predicting an outcome ($Y$), these models predict the difference in outcomes between being treated and not being treated ($Y_{treated} - Y_{control}$).
- Meta-Learners: Alicia Curth and Mihaela van der Schaar developed frameworks like SNet that use neural networks to explicitly separate features into those that affect baseline risk and those that drive treatment response.
- Structured Penalties: In Structural modeling using overlapped group penalties for discovering predictive biomarkers for subgroup analysis, Wenxuan Deng uses "overlapped group penalties" to ensure that a predictive marker is only selected if its interaction with the treatment is statistically significant beyond its baseline effect.
- Consistent Metagenes: Qiyuan Li and Benjamin Haibe-Kains use unsupervised methods to find "consistent metagenes"—gene expression patterns that persist across multiple independent datasets—to filter out the noise of individual study cohorts and find robust biological drivers of drug response.
The Path to Clinical Utility
The goal of this research direction is to move beyond "how sick is the patient?" to "what is the best path for this patient?" This requires a shift in how we evaluate success. As Mattis Gottlow demonstrates with the SIDES algorithm, identifying a high-efficacy subgroup in one trial is not enough; the predictive signal must replicate across different cohorts to prove it isn't just a fluke of the trial's specific population.
By isolating the predictive signal from the prognostic background, researchers like Ruishan Liu are now able to map hundreds of specific genetic mutations to their actual effects on treatment response across tens of thousands of patients, providing the roadmap for truly personalized medicine.