🧬 biology

PerturbRx: Learning Treatment-Conditioned Latent Transitions for Patient Drug Response Prediction

PerturbRx is a novel framework that predicts patient-level cancer drug responses by learning and transferring treatment-conditioned latent molecular transitions from single-cell data to pretreatment patient profiles, effectively overcoming data scarcity and tumor heterogeneity to achieve superior predictive performance.

Original authors: Yoshitaka Inoue, Minoh Jeong, Alfred Hero, Rui Kuang, Augustin Luna

Published 2026-08-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yoshitaka Inoue, Minoh Jeong, Alfred Hero, Rui Kuang, Augustin Luna

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Predicting how a specific cancer patient will respond to a specific drug is one of the most difficult challenges in modern medicine. Tumors are not uniform blocks of tissue; they are complex, shifting ecosystems of cells that vary wildly from person to person. Because of this biological diversity, a treatment that works wonders for one individual might fail completely for another. Currently, doctors often rely on broad statistical averages or preclinical tests done in petri dishes, but these methods struggle to capture the unique molecular reality of a patient's tumor. The core problem is that we rarely have the data to see what happens inside a patient's body after they start taking a medicine. We can see the tumor before treatment, and we can see the outcome later, but the crucial middle ground—the actual molecular changes the drug triggers inside the patient's cells—remains invisible.

To bridge this gap, researchers have turned to massive libraries of single-cell data, which act like a detailed map of how cells react to various chemical interventions in a controlled laboratory setting. These maps show how groups of cells shift their internal states when exposed to different drugs and doses. However, a single cell in a dish is not the same as a tumor in a human body. The new approach described in this research, called PerturbRx, attempts to use these laboratory maps to predict real-world outcomes without needing to wait for the patient to be treated first. The goal is to learn the "rules" of how drugs move cells from a healthy state to a treated state in the lab, and then apply those rules to a patient's pre-treatment profile to forecast whether the drug will work.

The researchers developed a two-step system to solve this puzzle. First, they trained a computer model on a massive dataset containing over one hundred million single-cell profiles from laboratory experiments. In these experiments, scientists had treated cells with various drugs at different strengths, but they did not track the same individual cell before and after treatment. Instead, they compared the average state of a group of untreated cells to the average state of a group of treated cells. The model learned to predict the specific shift, or transition, that a drug causes in a cell's internal chemistry. It learned not just what the drug looks like, but how that drug changes the cell's behavior based on the cell's starting condition and the drug's dosage.

Once the model learned these rules of cellular change, the researchers froze that knowledge and applied it to human patients. They took the pre-treatment genetic profiles of patients with cancer and asked the model to predict what would happen if those specific patients were given a specific drug. The model did not need to see the patient after treatment; it simply calculated the predicted molecular shift that the drug would cause in that specific patient's tumor. This predicted shift was then combined with the patient's original data and the drug's chemical signature to forecast the treatment outcome. The system essentially simulates the drug's effect in the patient's unique biological context, using the patterns it learned from millions of laboratory cells.

The team tested this method on three different sets of data involving hundreds of patients and dozens of drugs. In every case, adding the predicted drug-induced shift to the standard patient data improved the accuracy of the predictions. The system outperformed existing methods that rely only on static snapshots of the patient's tumor or simple comparisons of drug chemistry. For instance, in a large dataset of patient records, the new method correctly identified treatment responses more often than any other tested approach. The researchers found that the predicted changes were not just random noise; they carried specific information about whether a patient would respond to a therapy. In fact, the model learned that the direction and size of the predicted molecular change mattered more than simply guessing what the tumor would look like after treatment.

However, the study also revealed that this approach is not a universal cure-all. The benefit of using these predicted shifts varied significantly depending on the specific drug and the type of cancer. For some treatments, the model's prediction of cellular change provided a huge boost in accuracy, while for others, the improvement was small or even non-existent. The researchers found that the success of the method did not depend solely on how chemically similar the drug was to those used in the laboratory training data. Even drugs that were not exact matches in the training set could benefit from the system, provided the model could learn the general pattern of how that class of drugs affects cells. This suggests that the system is learning a deeper biological logic rather than just memorizing chemical structures.

The study concludes that learning how treatments alter molecular states is a powerful tool for predicting patient outcomes, even when we cannot measure those changes directly in the patient. By training on vast amounts of laboratory data and then transferring that knowledge to human patients, the researchers created a framework that can anticipate the effects of a drug before it is ever administered. While the method still faces challenges, particularly with drugs that have very few responders or limited data, it offers a promising new way to personalize cancer care. The work suggests that the future of drug prediction lies not just in looking at who the patient is or what the drug is, but in understanding the dynamic journey the drug takes the patient's cells through.

Technical Summary: PerturbRx

Problem Statement

Patient-level cancer treatment-response prediction is hindered by the scarcity of clinical response labels and the molecular heterogeneity of tumors. Existing methods typically transfer information from large-scale cancer cell-line screens to patient data, relying on static pretreatment molecular profiles and drug representations. A critical limitation of these approaches is that they do not explicitly model the molecular changes expected under treatment. While treatment response may depend on the direction and magnitude of intervention-induced changes, directly learning these changes from patients is difficult because matched pre- and post-treatment molecular profiles are rarely available at scale. Although single-cell perturbation atlases provide population-level evidence of state changes, individual cells are not paired before and after treatment, making it challenging to learn intervention-induced transitions for specific patient-drug pairs.

Methodology: PerturbRx

PerturbRx is a treatment-conditioned representation learning framework designed to learn intervention-induced latent transitions and utilize them as features for patient-drug response prediction. The framework operates in two distinct stages:

1. Source-Stage: Learning Transition Predictors

The model learns a drug- and dose-conditioned transition predictor from context-matched but unpaired control and treated single-cell populations (derived from the Tahoe-100M atlas).

Input: Control populations (DMSO) and treated populations (drug $d$ , dose $q$ ) from the same cell line and plate context.
Embedding: Cell profiles are embedded into a frozen latent space using scFoundation. Drugs are encoded using ChemBERTa (SMILES) combined with log-transformed dose.
Architecture: An intervention encoder $g$ maps drug and dose to an intervention embedding $\alpha_u$ . A transition predictor $P_\theta$ maps a control cell's latent state $z_X$ and the intervention embedding to a predicted latent transition vector $\hat{\Delta}$ .
Objective: The predictor is trained to match the distribution of the observed treated population while preserving the population-level transition. The loss function ( $L_{src}$ $L_{sr c}$ ) combines:
- Distributional Alignment: Maximum Mean Discrepancy (MMD) between predicted and observed treated populations.
- Transition Constraints: Cosine similarity (direction), Mean Squared Error (displacement), and norm constraints (magnitude) between the predicted population transition ( $\hat{\Delta}_{c,u}$ ) and the empirical transition ( $\Delta_{c,u}$ ).

2. Target-Stage: Transfer to Patient Response Prediction

The learned predictor is frozen and transferred to the patient domain to generate features for response prediction without requiring post-treatment measurements.

Input: A patient's pretreatment bulk RNA-seq profile ( $p$ ) and a candidate drug ( $d$ ).
Process: The frozen scFoundation encoder generates the patient embedding $z_p$ . The frozen transition predictor $P_\theta$ generates a patient-specific transition feature $\hat{\Delta}_{p,u}$ based on $z_p$ and the drug/dose intervention embedding.
Prediction: A response head $h_\psi$ predicts the binary response label using the concatenated features: $[z_p; m_d; \hat{\Delta}_{p,u}]$ , where $m_d$ is the drug embedding.
Training: Only the response head is trained on target-domain response labels; the transition predictor remains frozen.

Key Contributions

Formulation: The authors formulate patient drug-response prediction using predicted intervention-induced latent transitions, developing an inductive two-stage framework that generates transition representations for unseen patients and drugs without requiring post-treatment profiles in the target domain.
Performance on TCGA: On a controlled held-out-patient TCGA benchmark (TCGA-186), augmenting patient-drug features with pretrained transition features improved mean predictive performance over static representations, predicted post-intervention states, and various controls (shuffled transitions, random frozen predictors).
Cross-Domain Transfer: In an independent Patient-Derived Xenograft (PDX) cohort, the frozen transition predictor successfully transferred under model-disjoint evaluation, outperforming four established baselines (CODE-AE, TransDRP, DeepSADR, WISER) on a shared 10-drug benchmark.
Ablation Insights: The study demonstrates that the predictive gain is specific to the pretrained transition representation and the correspondence between the patient and the transition, rather than merely adding feature dimensionality or predicting a static post-treatment state.

Results

Source-Stage Alignment: PerturbRx achieved superior alignment (MMD = 0.0682) and transition prediction (cosine = 0.6715) compared to baselines including linear regression, MLPs, and conditional autoencoders on held-out single-cell conditions.
TCGA-186: PerturbRx achieved the highest aggregate AUROC (0.626) and AUPRC (0.600). Ablation studies showed that adding the transition feature increased AUROC from 0.581 (Patient + Drug) to 0.626.
TCGA-508: In a larger patient-grouped evaluation, PerturbRx achieved an AUROC of 0.692 and AUPRC of 0.787. Analysis of predicted transition geometry suggested responders exhibited smaller mean predicted transition magnitudes than non-responders, though this pattern showed heterogeneity across cohorts.
PDX Benchmark: PerturbRx achieved the highest aggregate AUROC (0.669) and AUPRC (0.608) on the shared 10-drug benchmark. Performance varied by drug; for instance, PerturbRx excelled on Erlotinib and Trametinib but underperformed on Ruxolitinib and Tamoxifen compared to baselines.
Heterogeneity: The utility of the transition feature was not uniform across all drugs or cohorts. In PDX, the benefit was not consistently explained by the chemical similarity between the target drug and the source perturbation atlas. However, a positive association was observed between transition utility and the observed responder fraction in the PDX cohort.

Significance and Claims

The paper claims that perturbation-pretrained latent transitions serve as useful representations for patient-level drug-response prediction, offering information beyond static baseline patient and drug features. By separating the learning of how treatments alter molecular states (in the source domain) from learning which alterations associate with clinical response (in the target domain), PerturbRx addresses the data scarcity of matched pre-post treatment profiles.

The authors emphasize that their objective is not the reconstruction of post-perturbation cell states, but rather the utility of the predicted displacement as a predictive feature. They acknowledge limitations, including the retrospective nature of target benchmarks, the domain shift between single-cell source data and bulk tumor target data, and the heterogeneity of transition utility across different drugs. They conclude that while source chemical coverage alone is insufficient to predict transfer success, the framework provides a robust mechanism for leveraging large-scale perturbation atlases to improve treatment-response prediction in the absence of matched post-treatment patient data.