A contextual bandit is an algorithm that repeatedly picks one action for each new case, such as one drug for one tumor, sees how well it worked, and updates its choices. Its job is to find, for each tumor, the drug that beats the alternatives. A prognostic gene signature is a small panel of genes, typically 3 to 6, whose weighted expression predicts how long a patient survives. The obvious way to combine them is to feed the signature to the bandit as its description of the patient. This proposal argues that this is the wrong place to put it. A prognostic signature predicts how well a tumor does regardless of treatment. A bandit's choice depends only on differences between drugs for that tumor. The signature therefore says little about which drug to pick. It says a lot about how noisy each observed drug response will be.
We propose to place the signature in a shared baseline term of the bandit's reward model. That term absorbs the "this tumor is resistant to everything" component of each response, so the drug-specific terms learn from cleaner signals. This is the same logic as covariate adjustment in A/B testing. We test the claim on the in vitro drug-screening benchmark where contextual bandits have already been studied, and we make a quantitative prediction. If the signature explains a fraction $R^2$ of the between-tumor variation in overall drug sensitivity, early-round regret should shrink by roughly $\sqrt{1-R^2}$. If that prediction fails, the idea fails.
Core Claim
The closest existing work is A Biologically Plausible Benchmark for Contextual Bandit Algorithms in Precision Oncology Using in vitro Data. It runs Thompson sampling bandits on about 900 cancer cell lines from the Genomics of Drug Sensitivity in Cancer screen (GDSC), where every line was tested against every drug. Thompson sampling is a bandit method that picks actions by sampling from its current beliefs about each one. The benchmark compresses more than 18,000 molecular features into a 20-dimensional UMAP embedding, an unsupervised non-linear projection, and finds that neural bandits beat linear ones and beat clinical guideline rules. That paper supplies our testbed and our strongest competitor.
The research file frames the gap as "use interpretable signatures instead of UMAP as context." We think that version is likely to underperform, for a reason the gene-signature literature itself raises. Signatures such as the ccRCC six-gene panel of Identification of a Six-Gene Prognostic Signature Characterized by Tumor Microenvironment Immune Profiles in Clear Cell Renal Cell Carcinoma estimate a survival curve. They conflate natural disease course with treatment-specific response, so their information is mostly about main effects rather than drug-by-tumor interactions. The benchmark's own "information saturation" result points the same way: adding guideline features to raw genomics did not help a flexible bandit. That result constrains us. Adding the signature as one more input is not expected to matter.
Our claim is structural. Write the reward for tumor $x$ under drug $a$ as
$$ r(x,a) = b\big(s(x)\big) + g_a(x) + \varepsilon , $$
where $s(x)$ is the signature's risk score (the weighted sum of its genes' expression), $b$ is a low-dimensional baseline shared by all drugs, $g_a$ is the drug-specific part, and $\varepsilon$ is noise. Decisions depend only on $g_a$. Two things change when $b$ is made explicit. First, every observation updates $b$, whichever drug was chosen, so the baseline is learned from all pulls rather than separately within each arm. Second, the residual noise seen by the $g_a$ terms falls from $\mathrm{Var}(b+\varepsilon)$ to roughly $\mathrm{Var}(\varepsilon)+(1-R^2)\mathrm{Var}(b)$. For linear-Gaussian Thompson sampling, regret scales with the noise standard deviation. That scaling is the source of the $\sqrt{1-R^2}$ prediction, and we expect it to hold mainly in early rounds, before a learned baseline catches up.
Research Program
The first step is offline and decides whether the project continues. We compute each published signature's risk score on GDSC cell-line expression and measure how much of each line's mean response across all drugs it explains. The general-sensitivity factor is computable offline only because GDSC has every drug tested on every line. We use the published gene lists and, where the original papers give them, the published coefficients; otherwise we refit coefficients on TCGA patient cohorts using the stated LASSO-Cox pipeline. The signatures are:
- the five-gene ferroptosis LUAD signature from Establishment of a ferroptosis-related gene signature for prognosis in lung adenocarcinoma patients;
- the six-gene ccRCC panel cited above;
- the four-gene OSCC panel from A Risk Score Model Based on Drug-Sensitivity-Related Genes Has the Potential to Predict Oral Squamous Cell Carcinoma Prognosis.;
- the three-gene breast chemoresistance panel from Identification of a prognostic chemoresistance-related gene signature associated with immune microenvironment in breast cancer.
The breast panel includes PP14571, an uncharacterized transcript that may be missing from cell-line platforms. We report that panel with and without it.
We run each signature both on its matched lineage (lung, kidney, head-and-neck, and breast lines) and pan-cancer, because transfer from patients to plastic-grown cell lines is itself uncertain. If the signatures explain essentially none of the general-sensitivity variance, we report that negative result and stop. A negative here would be informative in its own right: it would show that patient prognostic biology does not reach in vitro pharmacology.
The second step builds the bandit. We implement the two-part model in two versions. One is a linear Thompson sampler with a shared Bayesian baseline in $s(x)$ and per-drug linear terms over a molecular embedding. The other is a neural version that mirrors the benchmark's bootstrapped and dropout networks, with a separate scalar baseline head fed only by the signature score. Both run on the benchmark's three reward definitions, which are absolute, ordinal, and percentile transforms of the IC50 (the drug concentration that halves cell viability). Percentile rewards normalize each drug's scale but not each tumor's overall sensitivity. Per-tumor normalization is impossible online, because it would require seeing the tumor's response to drugs that were never given.
The third step is a secondary diagnostic for the signature literature. Comparing how much of each signature gene's contribution lands in $b$ versus in the $g_a$ terms gives an empirical prognostic-versus-predictive split for published panels. This addresses the conflation problem that static Cox models cannot resolve.
Baselines
The naive reference points are uniform random drug assignment and the rule-based clinical guideline agent from the GDSC benchmark paper. The guideline agent is reported to cover only "roughly 4.9%" of cell lines, and contextual bandits beat it. We reimplement both on the same lineage subsets and reward definitions.
The non-contextual Thompson sampler is the adaptive-trial alternative. It isolates whether any tumor information helps, and Bandit Algorithms for Precision Medicine, a survey of bandit methods for medicine, reviews its known regret scaling.
The strongest existing method is the benchmark's neural Thompson sampler on the 20-dimensional UMAP context. The paper reports that it significantly outperformed linear and Gaussian-process bandits, but the regret values appear only in figures. No reported numeric result was found on Lacuna for its cumulative regret, so we reproduce it under identical cell-line orderings, seeds, network size, and hyperparameter search budget.
Sparse LinUCB on the full expression matrix represents the theoretically motivated linear alternative from the same survey. LinUCB is a linear bandit that picks the action with the highest optimistic estimate, and the sparse version penalizes the number of genes used.
The "obvious idea" baseline feeds the signature genes as context to the same bandits, with no separate baseline term.
Two ablations isolate the key ingredient. The first keeps the two-part structure but learns the baseline from the UMAP features rather than from the signature. This tests whether the signature, as opposed to the shared baseline alone, matters. The second fills the baseline slot with random gene panels of matched size and matched expression level. This tests whether any small gene set works, which is plausible if the signal is generic proliferation.
Beating the strongest baseline means two things, judged over at least 20 random orderings. Early-horizon cumulative regret must be lower with confidence intervals that exclude zero, over the first few hundred rounds. Late-horizon regret must be no worse.
Why not just…?
Why not just use the signature as the bandit's context? Prognostic genes carry main-effect information. A bandit discards main effects when it compares arms, and the benchmark's information-saturation finding suggests extra inputs add little to a flexible model. We include this exact configuration as a baseline. If it matches the two-part model, our structural claim is wrong and we will say so.
Why not just let a neural bandit learn the baseline itself from all genes? This is the real threat. With enough data, a learned baseline should match a signature-based one. Our claim is therefore explicitly about sample efficiency: a 3-to-6-gene parametric baseline should be learnable within tens of rounds rather than hundreds. The learned-baseline ablation measures where, or whether, the curves cross. If they cross almost immediately, the signature adds nothing beyond the architectural change, and the contribution shrinks to "use a shared baseline head."
Why not just select genes by supervised LASSO on GDSC itself? In-sample selection uses the full counterfactual drug matrix. That would not exist when the bandit runs on real patients, where only one treatment is ever observed per patient. Selection on the bandit's own logged data is the fair version, and we run it as an additional learned-baseline variant. The patient-derived signature has the distinct advantage of being fixed before any drugs are tried.
Why not just train an offline predictor like DRPreter and pick its top drug? DRPreter is a pathway-structured graph network that predicts IC50 offline with lower error than prior graph models; no numeric value was found in the file. It requires a fully labeled training matrix and has no exploration mechanism. It answers a different question from ours, which concerns online learning from one observed treatment per tumor. We therefore treat it as an upper-bound oracle rather than a competitor.
Proposed Experiments
The decisive checks run in order, and each can end the project.
First, the offline $R^2$ of each signature on general drug sensitivity, compared with matched random panels. A signature must clearly beat random panels to proceed.
Second, a calibration test of the $\sqrt{1-R^2}$ prediction. We plot the observed ratio of early regret, two-part versus signature-free, against the ratio predicted from each signature's measured $R^2$, across signatures and lineages. Agreement supports the variance-reduction mechanism. Gains that do not track $R^2$ would point to some other cause.
Third, robustness across the three reward definitions. Under ordinal and percentile rewards, the tumor-level baseline should matter less, so the gain should shrink accordingly. Gains that stay constant there would contradict our account.
Fourth, the prognostic-versus-predictive decomposition of signature genes, reported per panel. For the ccRCC immune panel, this tests whether the immune-checkpoint-linked genes behave as pure prognosis in cell lines that lack any immune compartment. We expect they will, because the immune signal comes from the tumor's surroundings, which cell lines lack.
All results are in vitro. Translating to patients would require survival rewards and safety constraints that this study does not address.
References
- A Biologically Plausible Benchmark for Contextual Bandit Algorithms in Precision Oncology Using in vitro Data
- Identification of a Six-Gene Prognostic Signature Characterized by Tumor Microenvironment Immune Profiles in Clear Cell Renal Cell Carcinoma
- Establishment of a ferroptosis-related gene signature for prognosis in lung adenocarcinoma patients
- A Risk Score Model Based on Drug-Sensitivity-Related Genes Has the Potential to Predict Oral Squamous Cell Carcinoma Prognosis.
- Identification of a prognostic chemoresistance-related gene signature associated with immune microenvironment in breast cancer
- Bandit Algorithms for Precision Medicine
- DRPreter: Interpretable Anticancer Drug Response Prediction Using Knowledge-Guided Graph Neural Networks and Transformer