PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Yoshitaka Inoue, Minoh Jeong, Alfred Hero, Rui Kuang, Augustin Luna

arXiv:2608.21349v1q-bio.QMcs.LG

TL;DR

Patient-level cancer treatment-response prediction is limited by scarce labels, tumor heterogeneity, and representations that omit treatment-induced molecular changes. PerturbRx learns transferable, treatment-conditioned latent transitions from unpaired single-cell perturbation data and combines them with patient-drug features, improving aggregate performance across TCGA and PDX benchmarks.

  • Problem

    Patient-level cancer treatment-response prediction lacks representations that explicitly capture treatment-induced molecular changes alongside pretreatment patient and drug features.

  • Method

    PerturbRx learns drug- and dose-conditioned latent transition predictors from unpaired single-cell populations, freezes them, and transfers transition features to patient-drug response prediction.

  • Results

    Across TCGA-186, TCGA-508, and PDX, pretrained transition features improved aggregate performance and outperformed representation and predictor controls.

  • Takeaways & Limitations

    Perturbation-pretrained latent transitions provide useful treatment-response representations beyond static patient and drug features, although gains remain heterogeneous across treatments.

  • Takeaways & Limitations

    Evaluation uses retrospective heterogeneous cohorts, including proxy relapse labels and bulk tumor data that differ from the source single-cell cultured-cell-line domain.

Abstract

from arXiv · show

Scarce data and tumor heterogeneity limit patient-level cancer treatment-response prediction. Existing approaches predict response from pretreatment molecular profiles and drug representations, without explicitly modeling the molecular changes expected under treatment. We propose PerturbRx, a treatment-conditioned representation learning framework that learns intervention-induced latent transitions and uses them as patient-drug response features. PerturbRx trains a drug- and dose-conditioned transition predictor from context-matched but cell-unpaired control and treated single-cell populations, then freezes and transfers the predictor to pretreatment patient profiles without requiring post-treatment measurements. The transition is combined with patient and drug representations to predict response. Across TCGA and patient-derived xenograft benchmarks, PerturbRx achieves the strongest aggregate predictive performance among the evaluated methods. These results support perturbation-pretrained latent transitions as useful representations for patient-level drug-response prediction.

1 INTRODUCTION

PerturbRx addresses patient-level cancer treatment-response prediction by modeling drug-induced molecular changes rather than relying only on static pretreatment and drug representations. It learns treatment-conditioned latent transitions from unpaired single-cell perturbation populations and transfers the frozen predictor to pretreatment patient profiles without post-treatment measurements.

  • Problem: Patient-level response prediction is limited by scarce clinical labels, tumor molecular heterogeneity, and differences between cancer cell lines and patient tumors.These limitations motivate transferring information from large-scale cell-line screens while recognizing their imperfect correspondence to patient tumors.
  • Motivation: Static patient, cell-line, and drug features do not explicitly represent the direction or magnitude of molecular change expected under treatment.The introduction frames treatment response as depending on both the pretreatment state and intervention-induced change.
  • Motivation: Single-cell perturbation atlases enable learning intervention-induced state changes from context-matched control and treated populations despite lacking paired individual cells.These atlases span many drugs, doses, and cellular contexts, providing population-level evidence for treatment effects.
  • Method: PerturbRx learns a drug- and dose-conditioned latent-transition predictor from unpaired single-cell populations, freezes it, and applies it to pretreatment patient profiles and candidate drugs.The resulting representation is patient- and drug-specific and does not require post-treatment profiles in the target domain.
  • Evaluation objective: The framework tests whether intervention-pretrained latent transitions add patient-response signal beyond static patient and drug representations.Its stated questions concern learning from unpaired populations, incremental response signal, and consistency across evaluation settings.
  • Evaluation: PerturbRx is instantiated with Tahoe-100M, scFoundation, and ChemBERTa representations and evaluated on TCGA and independent PDX treatment-response cohorts.The evaluation includes source-stage alignment, held-out-drug analyses, controlled representation and predictor comparisons, and patient-level analyses.

2 RELATED WORK

Prior work transfers information from cancer cell-line screens to patient drug-response prediction and learns intervention-induced changes from unpaired single-cell populations. PerturbRx extends these paradigms by transferring a drug- and context-conditioned latent transition model from large-scale single-cell perturbation data to patient-level prediction.

  • Patient Drug Response Prediction and Preclinical-to-Patient Transfer: Prior patient drug-response methods transfer information from cancer cell-line screens using domain adaptation, molecular-similarity-aware transfer, weak supervision, supervised transfer, or generative augmentation.Examples include CODE-AE, TransDRP, WISER, DeepSADR, and GANDALF.
  • Single-Cell Perturbation and Latent Transition Modeling: Single-cell perturbation models learn intervention-induced changes from unpaired control and treated cell populations.Representative models include scGen, CPA, CellOT, PerturbDiff, and Conditional Monge Gap.
  • Single-Cell Perturbation and Latent Transition Modeling: PerturbRx learns a drug- and context-conditioned transition model from a large-scale single-cell perturbation atlas spanning many drugs, doses, and cellular contexts, then transfers it to patient-level prediction.Unlike prior single-cell perturbation models primarily evaluated by reconstruction of post-perturbation cellular states or distributions, PerturbRx transfers the learned transition model to a patient-level target domain.

3 METHOD

PerturbRx is a two-stage framework that learns drug- and dose-conditioned latent transitions from context-matched, unpaired single-cell populations and transfers the frozen predictor to pretreatment patient profiles. It uses predicted transitions alongside patient and drug representations for response prediction without post-treatment measurements.

  • Source stage: The source stage learns drug- and dose-conditioned latent transition predictors from context-matched but unpaired control and treated single-cell populations.Matched populations share cell line and plate context, but individual cells are not paired before and after treatment.
  • Patient stage: In the patient stage, the frozen predictor generates intervention-conditioned transitions from pretreatment patient embeddings, and a response head combines these transitions with patient-drug features.Only the response head is trained on binary response labels; post-treatment patient molecular states are unavailable.
  • Source stage: For each condition, PerturbRx pairs a DMSO control population with a treated population defined by the same cell line and plate, drug, and dose.Only conditions with sufficient control and treated cell counts are retained.
  • Transition predictor: A frozen encoder maps cells into a 3072-dimensional latent space, while a trainable intervention encoder represents drug embeddings and log-transformed doses for transition prediction.The intervention embedding dimension is r = 512, and drug embeddings come from frozen ChemBERTa representations of SMILES strings.
  • Source objective: The source objective combines maximum mean discrepancy with constraints on transition direction, displacement, and magnitude to align predicted and observed treated populations.The loss weights are λcos = 0.1, λmse = 3.0, and λnorm = 10^-3, selected using source-stage validation only.

4 EXPERIMENTS

PerturbRx was evaluated through source-domain transition prediction and transferred patient-response benchmarks, with frozen intervention components and trained response heads. It achieved strongest aggregate performance across TCGA-186, TCGA-508, and PDX evaluations, while treatment-specific results varied.

  • Experimental setup: PerturbRx was pretrained on Tahoe-100M and transferred with its intervention encoder and transition predictor frozen to three target benchmarks.Only the response head was trained after transfer; evaluation used drug-stratified, patient-grouped, or model-disjoint held-out partitions according to benchmark.
  • Source-stage transition prediction: 0.0682 MMD, 0.6715 transition cosine, and 0.0096 delta MSE were achieved across 800 held-out source conditions, outperforming all evaluated baselines.Performance improved progressively from the global mean transition through linear regression, MLP, and a state-conditioned conditional autoencoder.
  • TCGA-186: 0.626 ± 0.019 AUROC and 0.600 ± 0.029 AUPRC gave PerturbRx the highest aggregate TCGA-186 performance, versus WISER’s 0.580 ± 0.053 AUROC and CODE-AE’s 0.564 ± 0.051 AUPRC.PerturbRx led mean AUROC and AUPRC for Sorafenib and Temozolomide and ranked second for Cisplatin and Fluorouracil.
  • TCGA-508: 0.692 ± 0.055 AUROC and 0.787 ± 0.055 AUPRC were the highest overall TCGA-508 scores, although other methods performed best for several individual drugs.Responders had smaller mean predicted transition magnitudes than non-responders, 12.83 versus 15.07, with an adjusted difference of βR−NR = −1.56, 95% CI [−2.64, −0.49], p = 0.004.
  • PDX: 0.669 ± 0.025 AUROC and 0.608 ± 0.037 AUPRC were the highest aggregate PDX scores, exceeding DeepSADR’s 0.637 ± 0.024 AUROC and 0.554 ± 0.018 AUPRC.The PDX evaluation used a shared 10-drug benchmark under model-disjoint evaluation, while drug-level leaders varied across treatments.

5 CONCLUSION AND DISCUSSION

PerturbRx learns treatment-conditioned latent transitions from unpaired single-cell control and treated populations and transfers them to patient-level drug-response prediction. Its performance gains coexist with heterogeneous drug utility, domain and benchmark limitations, incomplete source support, and a need for prospective biological validation.

  • Contribution: PerturbRx learns drug- and dose-conditioned latent transition predictors from context-matched but unpaired control and treated single-cell populations.The resulting displacements are transferred to treatment-response prediction.
  • Contribution: Across TCGA-186, TCGA-508, and PDX, pretrained transition features improved aggregate performance when added to patient-drug representations.The framework outperformed representation and predictor controls, although the supplied passage truncates the full comparison.
  • Limitations and future work: Retrospective heterogeneous benchmarks, domain differences, limited source perturbation support, uncalibrated inference dose, and absent matched post-treatment validation remain limitations.Future work should use prospective cohorts with matched pre- and post-treatment measurements, broader perturbation coverage, mechanistic interpretation, and dose mapping.
  • Limitations: Transition utility was heterogeneous across drugs and was not consistently explained by source-drug chemical similarity across target benchmarks.An exploratory positive association appeared in TCGA-508 but did not replicate in TCGA-186 or PDX.
  • Limitations: Exact source-drug matching was unnecessary for successful transfer, as Sorafenib lacked an exact source match but had Regorafenib as a close molecular neighbor.Source chemical coverage alone was insufficient to predict when transferred transition features would improve response prediction.

AI USE STATEMENT

Generative AI tools supported language editing, code debugging, and document formatting, while the authors reviewed and verified the paper’s scientific content and accepted full responsibility.

  • Generative AI tools assisted with language editing, code debugging, and document formatting.
  • The authors reviewed and verified all scientific claims, experimental design decisions, implementations, analyses, and reported results.
  • The authors take full responsibility for the paper’s content.

ETHICS STATEMENT … B.1 TAHOE-100M SOURCE DATASET

The study uses public, de-identified datasets for methodological research, without recruiting human subjects or accessing identifiable patient information. Tahoe-100M provides context-matched, unpaired single-cell control and treated populations indexed by experimental treatment conditions for source-stage perturbation modeling.

  • ETHICS STATEMENT: The study uses publicly available, de-identified molecular and treatment-response datasets, including TCGA and PDX.No new human subjects were recruited, and no identifiable patient information was accessed.
  • ETHICS STATEMENT: The analyses are intended for methodological research rather than direct clinical decision-making.Predicted latent transitions and treatment-response scores are not validated clinical recommendations.
  • A NOTATION: Table 6 summarizes notation for the source-stage transition model and its transfer to patient response prediction.The notation covers both source-stage modeling and downstream patient-response transfer.
  • B DATA AND COHORT CONSTRUCTION: Tahoe-100M is used as the source-stage single-cell perturbation dataset.The dataset supplies the perturbation populations used to learn the source-stage transition model.
  • B.1 TAHOE-100M SOURCE DATASET: Context-matched but unpaired control and treated populations pair DMSO controls with drug-treated cells from the same cell line and plate.The pairing matches experimental context without requiring cell-level correspondence.
  • B.1 TAHOE-100M SOURCE DATASET: Each source condition is indexed by cell line, plate, drug, and dose.Table 7 summarizes the resulting source dataset, with observed treated conditions defined by these combinations and matched DMSO groups by cell line and plate.

B.2 TCGA-186 PATIENT BENCHMARK … C.4 MODEL ARCHITECTURE AND HYPERPARAMETERS

The paper constructs TCGA and PDX patient-treatment benchmarks and evaluates PerturbRx through frozen source-stage intervention learning followed by target-stage response-head transfer. Source validation distinguishes interpolation across observed perturbation support from drug-disjoint generalization, while the architecture uses frozen scFoundation and ChemBERTa representations.

  • B.2 TCGA-186 PATIENT BENCHMARK: 186 patient-drug records from 186 unique patients form TCGA-186 across five chemotherapy agents, including 91 responders and 95 non-responders.Records use TCGA barcodes and relapse-time-based response labels.
  • B.3 TCGA-508 TREATMENT-RESPONSE COHORT: 508 treatment episodes from 462 unique patients across five drugs and 24 cancer types form TCGA-508, with 325 ORR-positive and 183 ORR-negative episodes.The final cohort has positive-class prevalence 0.640 after matching treatment episodes to pretreatment expression profiles.
  • B.4 PDX COHORT CONSTRUCTION: 791 treatment episodes from 177 PDX models across 10 shared drugs include 343 responders and 448 non-responders, with overall positive-response prevalence 0.434.Tumor-stratified response prevalence is descriptive because drug coverage is nonuniform and some tumor-drug combinations are sparse or absent.
  • B.5 CROSS-BENCHMARK DRUG COVERAGE IN TAHOE-100M: Cross-benchmark drug coverage is assessed by exact Tahoe-100M support and nearest non-identical compounds using ChemBERTa cosine and Morgan ECFP4 Tanimoto similarity.Cisplatin similarity estimates require caution because SMILES-based representations and Morgan fingerprints are not optimized for metal complexes.
  • C.1 SOURCE-STAGE TRAINING PROCEDURE: PerturbRx has two stages: source perturbation learning trains drug- and dose-conditioned transitions, then target transfer freezes them while training only a response head on patient response labels.This decouples intervention learning from patient response supervision and avoids requiring post-treatment patient measurements.
  • C.2 SOURCE-STAGE OBJECTIVE DETAILS: The source objective combines distributional alignment with cosine, displacement, and norm alignment terms to promote intervention-specific population transitions.The reported loss weights are λcos = 0.1, λmse = 3.0, and λnorm = 10−3, selected without target-domain response labels.
  • C.3 SOURCE-STAGE VALIDATION PROTOCOL: Source validation uses MMD, transition cosine, and delta MSE across random held-out and held-out-drug splits, which test interpolation and compound-level transfer respectively.The random split contains 6,691 held-out rows and is not drug-disjoint, whereas held-out-drug validation withholds complete compounds.
  • C.4 MODEL ARCHITECTURE AND HYPERPARAMETERS: Cells use frozen scFoundation embeddings of dimension d = 3072, while frozen ChemBERTa drug features combine with log-dose before trainable intervention encoding.Source-stage parameters are selected using Tahoe-100M validation data; target response-head and representation settings are reported separately.

C.5 BASELINE IMPLEMENTATIONS … D.2 DOSE SENSITIVITY ACROSS TARGET COHORTS

PerturbRx is compared with four cell-line-to-patient baselines under shared held-out evaluation boundaries, while cohort-specific protocols define patient- and PDX-disjoint testing. Additional diagnostics show strong bulk–single-cell embedding separation, and 0.05 µM is the best TCGA-186 intervention-representation setting across all evaluated drugs and metrics.

  • C.5 BASELINE IMPLEMENTATIONS: PerturbRx is compared with CODE-AE, TransDRP, DeepSADR, and WISER, spanning domain adaptation, transfer learning, and weak supervision.The baselines were developed for cell-line drug-response transfer and do not natively use large single-cell perturbation atlases such as Tahoe-100M.
  • C.5 BASELINE IMPLEMENTATIONS: Benchmark-specific changes preserve each baseline’s original architecture, objectives, supervision regime, and training procedure while harmonizing data formatting, identifiers, splits, and outputs.Held-out response labels are excluded from model selection and target-domain tuning.
  • C.6 TCGA-186 SEEN-DRUG, HELD-OUT-PATIENT EVALUATION: TCGA-186 evaluates seen-drug generalization to previously unseen patients using five patient-disjoint splits shared across methods.The cohort contains 186 patient-drug response records from 186 unique patients, and performance is summarized across five random seeds with aggregate and drug-stratified AUROC and AUPRC.
  • C.7 PDX MODEL-DISJOINT EVALUATION: PDX evaluation uses five outer folds grouped by PDX model identifier, preventing treatment episodes from the same model from crossing train and test partitions.Cross-method comparison uses 10 shared drugs, comprising 791 treatment episodes from 177 PDX models; method-specific adaptation remains restricted to non-held-out models.
  • C.7 PDX MODEL-DISJOINT EVALUATION: For PDX response prediction, PerturbRx concatenates pretreatment PDX, ChemBERTa drug, and frozen predicted-transition embeddings evaluated at q ∈ {0.05, 0.5, 5.0} µM.Drug-specific AUROC and AUPRC are summarized as mean ± standard deviation across the five model-disjoint folds.
  • D.1 SOURCE-TARGET EMBEDDING DISTRIBUTION DIAGNOSTICS: Balanced PCA diagnostics show TCGA and Tahoe embeddings remain clearly separated under raw, sample-wise L2-normalized, and feature-wise standardized representations.The analysis used all 186 TCGA patients and an equal-sized fixed-seed Tahoe DMSO sample before projecting remaining cells and centroids.
  • D.2 DOSE SENSITIVITY ACROSS TARGET COHORTS: The 0.05 µM inference setting achieved the highest mean AUROC and AUPRC for each of the five TCGA-186 drugs under the fixed final response-model configuration.Because clinical doses cannot be mapped to the Tahoe-100M scale, q ∈ {0.05, 0.5, 5.0} µM is treated as an intervention-representation setting rather than calibrated therapeutic exposure.

D.3 CROSS-COHORT HETEROGENEITY IN TRANSITION-MAGNITUDE ASSOCIATIONS

Across TCGA-508, TCGA-186, and PDX, transition-magnitude differences between responders and non-responders were heterogeneous and near zero overall. Restricting analyses to larger response strata produced negative pooled effects, but these post hoc findings remain sensitivity analyses rather than evidence of a universal signature.

  • Overall heterogeneity: Across 20 primary-dose cohort–drug strata, the pooled standardized effect was small and unresolved (g = −0.045, 95% CI [−0.269, 0.180]), with substantial heterogeneity (I2 = 63.0%).The most extreme positive estimates primarily occurred in smaller response strata, while better-represented strata showed negative or near-zero effects.
  • Sample-size sensitivity: With at least 25 observations in both response groups, the pooled effect was negative (g = −0.270, 95% CI [−0.474, −0.067]).Negative values indicate smaller predicted transitions in responders.
  • Sample-size sensitivity: Using a 30-observation threshold yielded a similar negative pooled effect (g = −0.283, 95% CI [−0.522, −0.043]).This threshold was examined post hoc as a sensitivity analysis.
  • Interpretation: Because the thresholds were examined post hoc, the negative sensitivity estimates are not evidence for a universal response-associated transition signature.The analysis was exploratory and descriptive rather than primary evidence.

E SOURCE-STAGE MODEL SELECTION AND HYPERPARAMETER TUNING · F TARGET-DOMAIN HYPERPARAMETER AND REPRESENTATION-SETTING SELECTION

Source-stage configurations were selected solely on held-out Tahoe-100M conditions using transition cosine similarity, while target-domain response-head and representation settings were selected separately on target development data with the intervention components frozen. Outer held-out partitions were reserved for final performance estimation, and inference dose was treated as a representation setting rather than a clinically calibrated exposure.

  • E SOURCE-STAGE MODEL SELECTION AND HYPERPARAMETER TUNING: Source-stage selection used only Tahoe-100M validation data, excluding TCGA and PDX response labels when choosing objectives, molecular representations, and loss weights.
  • E SOURCE-STAGE MODEL SELECTION AND HYPERPARAMETER TUNING: Transition cosine similarity was the primary source-stage criterion, with higher values indicating better agreement between predicted and empirical population-level transition directions.Cosine similarity was reported for validation, whereas the corresponding cosine loss was minimized during training.
  • E SOURCE-STAGE MODEL SELECTION AND HYPERPARAMETER TUNING: Source validation used randomly held-out cell-line-plate-drug-dose combinations from represented drugs, measuring generalization within observed drug support rather than to entirely unseen compounds.
  • E SOURCE-STAGE MODEL SELECTION AND HYPERPARAMETER TUNING: The loss-weight sweep covered λcos ∈{0, 0.1, 0.3, 1.0}, λmse ∈{0, 0.3, 1.0, 3.0}, and λnorm ∈{0, 10−3}, yielding 32 configurations trained for 300 optimization steps each.
  • E SOURCE-STAGE MODEL SELECTION AND HYPERPARAMETER TUNING: Source-stage comparisons tested MMD versus entropically regularized Sinkhorn optimal transport and compared ChemBERTa, ChemBERTa-77M-MTR, and MoLFormer on the same 376drug source subset.
  • F TARGET-DOMAIN HYPERPARAMETER AND REPRESENTATION-SETTING SELECTION: Target-domain selection froze the pretrained intervention encoder and transition predictor, selecting only response-head and representation settings from the corresponding target benchmark’s non-held-out development data.Selection followed the benchmark’s patient-level or model-level grouping constraints.
  • F TARGET-DOMAIN HYPERPARAMETER AND REPRESENTATION-SETTING SELECTION: For TCGA-186, the final response head used hidden dimension 128, dropout 0.7, learning rate 10−3, weight decay 10−3, and early stopping after 50 patience epochs within 500 maximum epochs.
  • F TARGET-DOMAIN HYPERPARAMETER AND REPRESENTATION-SETTING SELECTION: Inference dose was treated as a representation setting rather than calibrated clinical exposure; primary TCGA-186 results used 0.05 µM, whereas TCGA-508 used 0.5 µM.Dose sensitivity was evaluated separately for target benchmarks, and outer held-out partitions were reserved for final performance estimation.

G TCGA-508 DOSE SENSITIVITY UNDER THE FIXED FINAL RESPONSE MODEL · H PDX DOSE SENSITIVITY · I DRUG-SPECIFIC TRANSITION UTILITY AND RELIABILITY ANALYSIS

PerturbRx shows stable TCGA-508 performance across intervention-representation doses, with 0.5 µM having the highest mean AUROC and AUPRC. Drug-level analyses examine dose heterogeneity, transition utility relative to static baselines, and potential reliability boundaries without establishing causal determinants.

  • G TCGA-508 DOSE SENSITIVITY UNDER THE FIXED FINAL RESPONSE MODEL: Across the three inference-dose settings, TCGA-508 performance was stable, with 0.5 µM yielding the highest mean AUROC and AUPRC.All other modeling choices were held fixed while only the Tahoe-100M inference dose varied.
  • G TCGA-508 DOSE SENSITIVITY UNDER THE FIXED FINAL RESPONSE MODEL: Drug-specific TCGA-508 sensitivity was heterogeneous: Cisplatin favored 0.05 µM, Fluorouracil favored 5.0 µM, and remaining drugs generally favored 0.5 µM.The remaining drugs showed their strongest or near-strongest performance at 0.5 µM.
  • H PDX DOSE SENSITIVITY: PDX dose sensitivity was evaluated at q ∈{0.05, 0.5, 5.0} µM on the shared 10-drug subset used for cross-method comparison.Drug-specific AUROC and AUPRC were summarized as mean ± standard deviation across five model-disjoint outer folds.
  • H PDX DOSE SENSITIVITY: The PDX inference dose was treated only as a representation setting and was not calibrated to in vivo PDX exposure.The tested doses were representation settings inherited from Tahoe-100M.
  • I DRUG-SPECIFIC TRANSITION UTILITY AND RELIABILITY ANALYSIS: The drug-level analysis assessed whether transition-augmented response prediction added consistent signal across treatments and characterized heterogeneity in transferred-transition utility.It also examined potential explanations involving each target drug’s representation.
  • I DRUG-SPECIFIC TRANSITION UTILITY AND RELIABILITY ANALYSIS: For each drug d, transition utility was defined as the difference between transition-augmented response-model performance and the corresponding static patient–drug baseline.Comparisons used the same target-domain splits and response-model procedures as the corresponding primary benchmark.
  • I DRUG-SPECIFIC TRANSITION UTILITY AND RELIABILITY ANALYSIS: The post-hoc analyses characterize transfer behavior and potential reliability boundaries but do not establish causal determinants of transition utility.The analyses were exploratory and drug-level.

I.1 DRUG-SPECIFIC TRANSITION BENEFIT ACROSS TARGET BENCHMARKS · I.2 ASSOCIATION WITH SOURCE-DRUG CHEMICAL SIMILARITY · I.3 RESPONSE COMPOSITION AND LOW-RESPONSE REGIMES IN PDX

Transition benefit varied by drug and benchmark, with the largest gains on TCGA-186 and PDX concentrated in selected treatments. Chemical similarity was not consistently predictive, while PDX responder fraction showed a positive but potentially outlier-sensitive association with transition utility.

  • I.1 DRUG-SPECIFIC TRANSITION BENEFIT ACROSS TARGET BENCHMARKS: On TCGA-508, Paclitaxel (+0.110) and Docetaxel (+0.075) had the largest ∆AUROC gains, while Cisplatin decreased (−0.017).Gemcitabine (+0.010) and Fluorouracil (+0.009) showed small changes.
  • I.1 DRUG-SPECIFIC TRANSITION BENEFIT ACROSS TARGET BENCHMARKS: On TCGA-186, transition augmentation improved AUROC for all five drugs, led by Fluorouracil (+0.223) and Gemcitabine (+0.173).Gains were smaller for Sorafenib (+0.047), Cisplatin (+0.016), and Temozolomide (+0.002).
  • I.1 DRUG-SPECIFIC TRANSITION BENEFIT ACROSS TARGET BENCHMARKS: In PDX, Erlotinib (+0.200) and Gemcitabine (+0.142) improved AUROC, whereas Ruxolitinib (−0.243) and Tamoxifen (−0.214) had the largest decreases.Fluorouracil (+0.050) and Trametinib (+0.023) also improved AUROC, while several drugs changed little or decreased.
  • I.2 ASSOCIATION WITH SOURCE-DRUG CHEMICAL SIMILARITY: Nearest-source chemical similarity was assessed with Morgan ECFP4 Tanimoto and ChemBERTa cosine similarity against drug-specific ∆AUROC.The analysis tested whether source-atlas chemical representation explained heterogeneous transition utility.
  • I.2 ASSOCIATION WITH SOURCE-DRUG CHEMICAL SIMILARITY: On TCGA-508, both chemical-similarity measures correlated positively with ∆AUROC (ρ = 0.90), but exact permutation tests were nonsignificant (pexact = 0.083).The corresponding ∆AUPRC association was weaker (ρ = 0.70, pexact = 0.233).
  • I.2 ASSOCIATION WITH SOURCE-DRUG CHEMICAL SIMILARITY: Chemical similarity did not consistently explain transition utility: AUROC associations were weak on TCGA-186 and near zero in PDX.PDX associations were ρ = −0.067 for Morgan similarity and ρ = −0.030 for ChemBERTa similarity.
  • I.3 RESPONSE COMPOSITION AND LOW-RESPONSE REGIMES IN PDX: In PDX, total sample size, minority-class size, and minority-class fraction were not clearly associated with ∆AUROC, whereas responder fraction was positively associated (ρ = 0.673, pexact = 0.039).The analysis considered these four response-distribution characteristics as candidate explanations for drug-specific benefit.
  • I.3 RESPONSE COMPOSITION AND LOW-RESPONSE REGIMES IN PDX: The responder-fraction association was outlier-sensitive because Tamoxifen and Ruxolitinib had the lowest responder fractions and largest AUROC decreases; jointly excluding them reduced ρ to 0.381 (p = 0.352).Leave-one-drug-out analyses nevertheless preserved positive associations, with ρ ranging from 0.57 to 0.83.
Loading 2608.21349v1…