PerturbRx

A tumor is profiled once, before treatment starts. The molecular state after the drug takes effect is almost never measured, so the before-and-after pairs that would show what a therapy did to a patient's cells barely exist at clinical scale. Response predictors work around this by correlating a pretreatment profile and a drug descriptor with an outcome label — never representing the change the drug is supposed to cause, though response depends on the direction and size of that shift as much as on where the tumor starts.

PerturbRx makes that shift the feature. In its source stage, a drug- and dose-conditioned predictor learns latent transitions from a single-cell perturbation atlas, where control and treated populations from the same cell line and plate give population-level evidence of an intervention's effect even though no cell is observed both before and after. The predictor is then frozen and applied to a patient's pretreatment bulk RNA-seq profile and a candidate drug, producing a transition vector that a small supervised head combines with the static patient and drug representations to predict response. Learning how treatments alter molecular states is thereby separated from learning which alterations track clinical response, and the first half never touches patient data.

Released as an August 2026 preprint by the University of Minnesota, the National Library of Medicine at NIH, Inha University and the University of Michigan, it is not a from-scratch expression model: the new artifact is the transition predictor, composed over frozen scFoundation cell embeddings and frozen ChemBERTa drug embeddings.

#Key Features

• Transition as the representation: Each patient-drug pair is described by a predicted latent displacement, not a static profile; the stated objective is not cell-state reconstruction.
• Learning from unpaired populations: Training aligns predicted and observed treated distributions with maximum mean discrepancy, plus direction, displacement and magnitude terms.
• Frozen cross-domain transfer: The intervention encoder and transition predictor are fixed after source training and only a response head sees clinical labels, so perturbation data and patient outcomes never share trainable parameters.
• Dose conditioning: The intervention embedding takes a log-dose feature alongside the drug structure, so one compound yields different transitions at different exposures.
• Inductive: The predictor keys on representations, not fixed patient or drug identities.

#Technical Details

The source stage trains on Tahoe-100M: 100.6M post-QC cells across 50 cell lines, 14 plates and 379 drugs at three doses (0.05, 0.5 and 5.0 µM), yielding 65,218 treated conditions. Cells are encoded by a frozen scFoundation into 3,072 dimensions, drugs by a frozen ChemBERTa (seyonec/ChemBERTa-zinc-base-v1) into 768. A trainable intervention encoder maps the drug embedding plus log dose to a 512-dimensional vector; the transition predictor maps a cell embedding plus that vector through a 1,024-unit hidden layer to a 3,072-dimensional transition. Training runs 100 epochs of AdamW at learning rate 1e-4, in batches of 32 conditions. On 800 held-out source conditions PerturbRx reached MMD 0.0682, transition cosine 0.6715 and delta MSE 0.0096, ahead of a state-conditioned conditional autoencoder at 0.0758 / 0.6476 / 0.0108. Transferred frozen, it reached AUROC 0.626 ± 0.019 on the 186-record TCGA benchmark against 0.580 for WISER, 0.692 ± 0.055 across 508 episodes from 462 patients, and 0.669 ± 0.025 on 791 episodes from 177 patient-derived xenograft models under model-disjoint folds against 0.637 for DeepSADR. Ablations isolate the transition: the patient embedding alone gives 0.547 AUROC and adding the drug embedding 0.581, while shuffling transitions among patients (0.566) or substituting a random frozen predictor (0.589) does not reproduce the full gain.

#Applications

The intended use is prioritizing among candidate agents for a specific tumor when only a pretreatment transcriptome is in hand — the setting translational oncology groups face with archival cohorts and with xenograft panels used for preclinical triage. Because the transition is just a feature vector, it can also be appended to existing response models or used to compare treatment-induced geometry across cohorts: responders in the larger TCGA cohort had smaller predicted transition magnitudes than non-responders.

#Impact

PerturbRx is evidence that giga-scale perturbation atlases can pretrain for clinical prediction, not only for the cell-state reconstruction tasks that models like CPA and PerturbDiff are usually judged on. The caveats are substantial and the authors state them. The benchmarks are retrospective, the primary TCGA cohort uses relapse-time proxy labels rather than prospective RECIST endpoints, and aggregate gains do not hold uniformly per drug. The predicted transitions are never compared against measured post-treatment profiles, so they are not validated molecular trajectories. Held-out-drug evaluation degrades markedly, and inference dose is a representation setting rather than a calibrated clinical exposure. The source domain is cultured cell lines while the targets are bulk tumor RNA-seq, a gap the shared scFoundation space does not by itself close. Code is promised on publication via a public GitHub repository and an archived Zenodo release; neither is available yet.