PerturbRx: Learning Treatment-Conditioned Latent Transitions for Patient Drug Response Prediction
Authors: Yoshitaka Inoue, Minoh Jeong, Alfred Hero, Rui Kuang, Augustin Luna
Organizations: Department of Computer Science, University of Minnesota, Minneapolis, MN, USA · Computational Biology Branch, National Library of Medicine, Bethesda, MD, USA · Therapeutic Development Branch, National Library of Medicine, Bethesda, MD, USA · Department of Electrical and Electronic Engineering, Inha University, Incheon, South Korea · Department of Electrical Engineering and Computer Science, University of Michigan, Ann Arbor, MI, USA
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.
Figures & tables
Figure 1: Overview of population-level PerturbRx. (A) Source-stage learning from context-matched but unpaired control and treated single-cell populations. A drug- and dose-conditioned predictor maps control embeddings and intervention features to cell-wise transitions, which are added to the controls to form post-intervention embeddings. The predictor is trained using distributional, transition-direction, transition-displacement, and transition-magnitude objectives. (B) Transfer of the transition predictor to patient drug-response prediction. The predictor generates an intervention-conditioned transition Δ^p,u , and the response head predicts response from [zp;md;Δ^p,u] .
Figure 2: Source-stage perturbation prediction. A. Observed untreated-to-treated shift for an example held-out condition. B. Alignment between predicted and observed treated states for the same condition. Crosses denote population means. C. Performance across 800 randomly held-out conditions. Bold , underline , and † indicate the best, second-best, and third-best results, respectively.
Overall
Cisplatin
Fluorouracil ‡
Gemcitabine ‡
Sorafenib
Temozolomide
Methods
AUROC ↑
AUPRC ↑
AUROC ↑
AUPRC ↑
AUROC ↑
AUPRC ↑
AUROC ↑
AUPRC ↑
AUROC ↑
AUPRC ↑
AUROC ↑
AUPRC ↑
CODE-AE
0.534 † ± 0.053
0.564 ± 0.051
0.609 † ± 0.073
0.666 † ± 0.058
0.548 † ± 0.147
0.516 ± 0.134
0.360 ± 0.023
0.413 ± 0.023
0.631 † ± 0.015
0.689 † ± 0.040
0.581 † ± 0.028
0.587 † ± 0.031
DeepSADR
0.530 ± 0.011
0.519 ± 0.027
0.569 ± 0.103
0.596 ± 0.093
0.749 ± 0.113
0.768 ± 0.137
0.573 † ± 0.122
0.601 ± 0.121
0.720 ± 0.126
0.712 ± 0.148
0.508 ± 0.111
0.528 ± 0.106
TransDRP
0.494 ± 0.061
0.500 ± 0.068
0.484 ± 0.176
0.598 ± 0.104
0.354 ± 0.094
0.391 ± 0.044
0.593 ± 0.068
0.569 ± 0.070
0.473 ± 0.220
0.581 ± 0.156
0.496 ± 0.217
0.552 ± 0.175
WISER
0.580 ± 0.053
0.563 † ± 0.062
0.709 ± 0.151
0.756 ± 0.141
0.480 ± 0.168
0.526 † ± 0.175
0.578 ± 0.076
0.547 † ± 0.081
0.467 ± 0.105
0.574 ± 0.115
0.589 ± 0.124
0.621 ± 0.118
PerturbRx
0.626 ± 0.019
0.600 ± 0.029
0.684 ± 0.056
0.733 ± 0.093
0.674 ± 0.104
0.645 ± 0.134
0.551 ± 0.082
0.530 ± 0.055
0.753 ± 0.213
0.801 ± 0.174
0.653 ± 0.142
0.680 ± 0.107
Table 1: Patient-response prediction on the TCGA-186 benchmark.
Input representation / control
AUROC ↑
AUPRC ↑
Patient embedding
0.547 ± 0.040
0.518 ± 0.062
Patient + drug embedding
0.581 ± 0.019
0.556 ± 0.055
Patient + drug + shuffled transition
0.566 ± 0.029
0.551 ± 0.048
Patient + drug + random transition
0.589 † ± 0.021
0.574 † ± 0.039
Patient + drug + predicted post-state
0.609 ± 0.051
0.582 ± 0.062
PerturbRx (Patient + drug + transition)
0.626 ± 0.019
0.600 ± 0.029
Table 2: Ablation results on TCGA-186.
Overall
Cisplatin
Docetaxel ‡
Fluorouracil ‡
Gemcitabine ‡
Paclitaxel ‡
Methods
AUROC ↑
AUPRC ↑
AUROC ↑
AUPRC ↑
AUROC ↑
AUPRC ↑
AUROC ↑
AUPRC ↑
AUROC ↑
AUPRC ↑
AUROC ↑
AUPRC ↑
CODE-AE
0.525 † ± 0.061
0.649 ± 0.193
0.600 † ± 0.028
0.841 † ± 0.027
0.453 ± 0.056
0.557 ± 0.038
0.545 ± 0.024
0.683 ± 0.042
0.472 † ± 0.044
0.366 ± 0.043
0.556 ± 0.064
0.799 ± 0.040
DeepSADR
0.565 ± 0.147
0.735 ± 0.093
0.511 ± 0.071
0.807 ± 0.121
0.724 ± 0.244
0.812 ± 0.160
0.565 ± 0.103
0.741 ± 0.072
0.675 ± 0.055
0.582 ± 0.234
0.351 ± 0.093
0.732 † ± 0.258
TransDRP
0.508 ± 0.089
0.665 † ± 0.174
0.613 ± 0.164
0.885 ± 0.054
0.425 ± 0.241
0.588 ± 0.197
0.587 ± 0.192
0.710 † ± 0.178
0.422 ± 0.083
0.418 † ± 0.079
0.493 ± 0.253
0.726 ± 0.122
WISER
0.507 ± 0.016
0.650 ± 0.151
0.480 ± 0.040
0.829 ± 0.017
0.515 † ± 0.074
0.607 † ± 0.055
0.511 ± 0.032
0.667 ± 0.034
0.512 ± 0.032
0.422 ± 0.031
0.519 † ± 0.051
0.724 ± 0.033
PerturbRx
0.692 ± 0.055
0.787 ± 0.055
0.631 ± 0.168
0.864 ± 0.085
0.711 ± 0.141
0.802 ± 0.112
0.562 † ± 0.116
0.748 ± 0.098
0.472 † ± 0.095
0.395 ± 0.052
0.715 ± 0.135
0.842 ± 0.081
Table 3: Patient-response prediction on the TCGA-508 patient-grouped benchmark.
Figure 3: Predicted transition geometry in TCGA-508. PCA was fit jointly to pretreatment and predicted post-intervention embeddings for all 508 treatment episodes. Contours show the two state distributions within each response group, and black arrows connect their group centroids. Panel annotations report sample size, mean predicted transition magnitude ∥Δ^∥2 , and mean episode-level displacement in the two-dimensional PCA projection. The reported 2D shift is the mean paired episode-level displacement, not the centroid-arrow length.
Input representation / control
AUROC ↑
AUPRC ↑
Patient embedding
0.638 ± 0.038
0.756 ± 0.041
Patient + drug embedding
0.681 † ± 0.046
0.780 † ± 0.055
Patient + drug + shuffled transition
0.665 ± 0.030
0.770 ± 0.038
Patient + drug + random transition
0.676 ± 0.046
0.773 ± 0.046
Patient + drug + predicted post-state
0.685 ± 0.048
0.781 ± 0.054
PerturbRx (Patient + drug + transition)
0.692 ± 0.055
0.787 ± 0.055
Table 4: Ablation results on the TCGA-508.
Method
Overall
Alpelisib ‡
Buparlisib
Erlotinib ‡
Fluorouracil ‡
Gemcitabine ‡
Paclitaxel ‡
Ribociclib ‡
Ruxolitinib
Tamoxifen
Trametinib ‡
CODE-AE
0.432 ± 0.035
0.472 ± 0.193
0.474 ± 0.106
0.400 ± 0.253
0.445 ± 0.153
0.467 ± 0.317
0.589 ± 0.140
0.486 ± 0.047
0.439 ± 0.103
0.476 ± 0.268
0.407 ± 0.245
DeepSADR
0.637 ± 0.024
0.587 ± 0.079
0.595 ± 0.074
0.300 ± 0.217
0.304 ± 0.196
0.462 ± 0.349
0.511 ± 0.065
0.497 ± 0.082
0.762 ± 0.158
0.719 ± 0.206
0.430 ± 0.258
TransDRP
0.470 ± 0.026
0.418 ± 0.124
0.438 ± 0.107
0.367 † ± 0.183
0.508 † ± 0.373
0.463 ± 0.359
0.540 ± 0.232
0.525 † ± 0.047
0.342 ± 0.133
0.181 ± 0.262
0.640 ± 0.123
WISER
0.492 † ± 0.031
0.527 † ± 0.047
0.503 † ± 0.145
0.400 ± 0.384
0.616 ± 0.138
0.465 † ± 0.214
0.554 † ± 0.174
0.576 ± 0.084
0.575 † ± 0.209
0.595 † ± 0.337
0.537 † ± 0.322
PerturbRx
0.669 ± 0.025
0.631 ± 0.109
0.544 ± 0.054
0.767 ± 0.149
0.680 ± 0.172
0.473 ± 0.334
0.634 ± 0.264
0.607 ± 0.072
0.682 ± 0.221
0.657 ± 0.211
0.760 ± 0.189
Table 5: Drug-stratified performance on the shared 10-drug PDX benchmark. Values are mean ± standard deviation across five model-disjoint outer folds. Best, second-best, and third-best results are indicated by bold , underline , and † , respectively. Drug-specific AUPRC should be interpreted relative to the corresponding positive-class prevalence. ‡ indicates exact Tahoe-100M support by normalized drug name or canonicalized molecular structure.
Appendix figures & tables19 assets
Supplementary material from the paper’s appendix.
Appendix
Symbol
Meaning
c=(ℓ,b)
Source context defined by cell line ℓ and plate b
u=(d,q)
Intervention defined by drug d and dose q
Xc
DMSO control population in context c
Yc,u
Treated population in context c under intervention u
ZX,ZY
Frozen-encoder embeddings of Xc and Yc,u
αu
Learned intervention embedding for u
Appendix
Table 6: Notation used in the PerturbRx formulation.
Statistic
Value
Post-QC cells
100.6M
Plates
14
Cell lines
50
Treated drugs
379
Dose levels
0.05, 0.5, 5.0 μ M
Observed treated conditions
65,218
Appendix
Table 7: Statistics of the Tahoe-100M source-stage perturbation dataset. Observed treated conditions are defined as observed cell line, plate, drug, and dose combinations. Matched DMSO groups are defined by cell line and plate.
Drug
Samples
Responders
Non-responders
Cisplatin
40
20
20
Fluorouracil
28
12
16
Gemcitabine
46
23
23
Sorafenib
26
13
13
Temozolomide
46
23
23
Total
186
91
95
Appendix
Table 8: Drug-level composition of TCGA-186. Response labels are relapse-time proxies derived from the processed WISER-style benchmark.
Cancer type
Cis.
Flu.
Gem.
Sor.
Tem.
Total
Adrenocortical
0
0
0
1
0
1
Bladder
0
0
4
0
0
4
Brain
0
0
0
0
42
42
Cervical
5
0
0
0
0
5
Colorectal
0
8
0
0
0
8
Esophagus/Stomach
1
20
0
0
0
21
Appendix
Table 9: Cancer-type and drug composition of TCGA-186. Entries denote the number of unique patients associated with each cancer type and treatment. Cancer types were assigned using TCGA project metadata from the Genomic Data Commons.
Drug
Episodes
ORR+
ORR-
Prevalence
Tahoe match
Cisplatin
151
120
31
0.795
No
Docetaxel
52
29
23
0.558
Yes
Fluorouracil
106
69
37
0.651
Yes
Gemcitabine
112
42
70
0.375
Yes
Paclitaxel
87
65
22
0.747
Yes
Total
508
325
183
0.640
-
Appendix
Table 10: Drug-level composition of the TCGA-508 analytic cohort after expression matching. ORR-positive denotes complete or partial response, whereas ORR-negative denotes stable or progressive disease. An exact Tahoe match indicates agreement under harmonized drug-name and canonical-SMILES matching.
Tumor type
Cisplatin
Docetaxel
Fluorouracil
Gemcitabine
Paclitaxel
Total
BRCA
0
25
0
4
48
77
PAAD
1
0
10
66
0
77
CESC
62
0
4
1
2
69
STAD
5
2
51
0
1
59
HNSC
44
3
0
1
2
50
BLCA
10
3
0
17
1
31
Appendix
Table 11: Tumor-type composition of the TCGA-508 benchmark across the five evaluated drugs. Entries denote treatment-episode counts for each tumor type-drug combination.
Tumor type
Models
Episodes
Drugs
Response prevalence
BRCA
38
227
6
0.401
CM (SKCM)
32
64
2
0.375
CRC (COAD/READ)
43
168
7
0.470
NSCLC (LUAD/LUSC)
25
124
5
0.387
PDAC (PAAD)
39
208
6
0.486
Total
177
791
10
0.434
Appendix
Table 12: Tumor-type composition of the shared 10-drug PDX benchmark. Parenthetical labels denote the corresponding TCGA project codes. Response prevalence denotes the proportion of episodes labeled CR, PR, or SD.
Tumor type
Alpelisib
Buparlisib
Erlotinib
Fluorouracil
Gemcitabine
Paclitaxel
Ribociclib
Ruxolitinib
Tamoxifen
Trametinib
Total
BRCA
38
38
0
0
0
38
38
37
38
0
227
PDAC
36
35
0
0
34
0
35
33
0
35
208
CRC
42
40
0
41
1
0
42
1
0
1
168
NSCLC
25
25
25
0
0
24
25
0
0
0
124
CM
0
32
0
0
0
0
32
0
0
0
64
Total
141
170
25
41
35
62
172
71
38
36
791
Appendix
Table 13: Tumor-drug composition of the shared 10-drug PDX benchmark. Entries denote treatment-episode counts for each tumor type-drug combination.
Drug
Primary target / mechanism
Exact Tahoe match
Nearest Tahoe compound
ChemBERTa cosine
Alpelisib
PI3K α
Yes
Sonidegib
0.906
Buparlisib
Pan-PI3K
No
Bimiralisib
0.991
Cisplatin
DNA cross-linking
No
Sodium Salicylate
0.753
Docetaxel
Microtubule stabilization
Yes
Docetaxel (Trihydrate)
0.996
Erlotinib
EGFR
Yes
XRK3F2
0.823
Fluorouracil
TYMS / DNA–RNA incorporation
Yes
Methylthiouracil
0.882
Appendix
Table 14: Chemical coverage of target drugs in Tahoe-100M. Exact support indicates whether the target drug is present in Tahoe-100M by normalized drug name or canonicalized molecular structure. For each target drug, the nearest non-identical Tahoe compound is reported under ChemBERTa embedding cosine similarity and Morgan ECFP4 Tanimoto similarity. Exact target matches were excluded from nearest-neighbor searches. The two similarity measures are reported independently.
Component
Configuration
Source-stage transition learning
Cell encoder
Frozen scFoundation, d=3072
Drug encoder
Frozen ChemBERTa, d=768
Intervention encoder g
Input 768+1 ; hidden/output dim 512/512
Transition predictor Pθ
Input 3072+512 , hidden dim 1024 , output dim 3072
Dropout
0.1
Appendix
Table 15: Architecture and hyperparameter details for PerturbRx.
Figure 4: Source-target embedding distribution diagnostics. A-C. Balanced joint PCA of TCGA bulk-tumor embeddings, Tahoe DMSO single-cell embeddings, and Tahoe cell-line centroids using A raw embeddings, B sample-wise L2 -normalized embeddings, and C feature-wise standardized embeddings. D. Distribution of raw-embedding PC1 scores for TCGA patients and Tahoe DMSO cells. PCA was fit using equal numbers of TCGA patients and Tahoe cells.
AUROC
AUPRC
Drug
0.05
0.5
5.0
0.05
0.5
5.0
Cisplatin
0.684 ± 0.056
0.641 ± 0.086
0.647 ± 0.021
0.733 ± 0.093
0.705 ± 0.111
0.711 ± 0.046
Fluorouracil
0.674 ± 0.104
0.560 ± 0.214
0.600 ± 0.197
0.645 ± 0.134
0.526 ± 0.173
0.576 ± 0.197
Gemcitabine
0.551 ± 0.082
0.507 ± 0.100
0.491 ± 0.124
0.530 ± 0.055
0.511 ± 0.081
0.522 ± 0.106
Sorafenib
0.753 ± 0.213
0.667 ± 0.162
0.740 ± 0.201
0.801 ± 0.174
0.704 ± 0.160
0.773 ± 0.172
Temozolomide
0.653 ± 0.142
0.638 ± 0.133
0.647 ± 0.164
0.680 ± 0.107
0.655 ± 0.086
0.651 ± 0.115
Appendix
Table 16: Drug-wise dose sensitivity of PerturbRx on TCGA-186 under the fixed final response-model configuration. Values are mean ± standard deviation across five held-out-patient evaluations. Bold indicates the highest mean and underline the second-highest mean within each drug and metric. The inference dose is treated as a representation setting rather than a calibrated clinical exposure.
Figure 5: Transition-magnitude effect versus response-group size. Each point represents one cohort–drug stratum and shows the standardized difference in predicted transition magnitude between responders and non-responders. Negative values indicate smaller predicted transitions in responders. Effect estimates were more variable in sparsely represented strata, whereas larger strata tended to show negative or near-zero effects. This analysis is descriptive and is not used as primary evidence for a universal magnitude–response association.
Drug representation
Distribution loss
λcos
λmse
λnorm
Held-out condition cosine ↑
ChemBERTa-zinc-base-v1
MMD
0.1
3.0
10−3
0.2099
ChemBERTa-zinc-base-v1
Sinkhorn
0.3
1.0
0
0.1906
ChemBERTa-77M-MTR
MMD
1.0
1.0
10−3
0.2074
ChemBERTa-77M-MTR
Sinkhorn
0.3
3.0
10−3
0.1989
MoLFormer-XL
MMD
1.0
1.0
10−3
0.1856
MoLFormer-XL
Sinkhorn
1.0
3.0
10−3
0.1938
Appendix
Table 17: Best source-stage configuration within each molecular representation and distributional-loss combination. Performance is measured by transition cosine similarity on randomly held-out source conditions; higher is better.
AUROC
AUPRC
Drug
0.05
0.5
5.0
0.05
0.5
5.0
Overall
0.682 ± 0.054 †
0.692 ± 0.055
0.684 ± 0.062
0.783 ± 0.063 †
0.787 ± 0.055
0.785 ± 0.056
Cisplatin
0.686 ± 0.166
0.631 ± 0.168 †
0.638 ± 0.140
0.872 ± 0.105
0.864 ± 0.085 †
0.869 ± 0.075
Docetaxel
0.709 ± 0.215
0.711 ± 0.141
0.687 ± 0.173 †
0.816 ± 0.171
0.802 ± 0.112
0.772 ± 0.150 †
Fluorouracil
0.552 ± 0.143 †
0.562 ± 0.116
0.568 ± 0.076
0.714 ± 0.120 †
0.748 ± 0.098
0.750 ± 0.080
Gemcitabine
0.455 ± 0.095 †
0.472 ± 0.095
0.460 ± 0.108
0.387 ± 0.053
0.395 ± 0.052
0.385 ± 0.058 †
Appendix
Table 18: Drug-wise dose sensitivity of PerturbRx on TCGA-508 under the fixed final response-model setting. Values are mean ± standard deviation across five patient-grouped evaluation folds. Bold indicates the best mean within each drug and metric across the three inference doses, underline indicates the second-best mean, and † indicates the third-best mean. The inference dose is treated as a representation setting and is not calibrated to clinical treatment exposure.
AUROC
AUPRC
Drug
0.05
0.5
5.0
0.05
0.5
5.0
Fluorouracil
0.680 ± 0.172 †
0.705 ± 0.159
0.813 ± 0.153
0.749 ± 0.128 †
0.782 ± 0.120
0.831 ± 0.142
Buparlisib
0.544 ± 0.054
0.553 ± 0.044
0.537 ± 0.055 †
0.559 ± 0.056
0.538 ± 0.061
0.529 ± 0.051 †
Alpelisib
0.631 ± 0.109 †
0.658 ± 0.064
0.678 ± 0.082
0.590 ± 0.089 †
0.591 ± 0.107
0.678 ± 0.066
Ruxolitinib
0.682 ± 0.221
0.564 ± 0.144 †
0.738 ± 0.182
0.361 ± 0.268
0.312 ± 0.245 †
0.351 ± 0.190
Ribociclib
0.607 ± 0.072
0.559 ± 0.074 †
0.572 ± 0.069
0.535 ± 0.069
0.501 ± 0.097 †
0.508 ± 0.069
Appendix
Table 19: Drug-wise dose sensitivity of PerturbRx on the shared 10-drug PDX benchmark. Values are mean ± standard deviation across five model-disjoint outer folds. Bold indicates the best mean within each drug and metric across the three inference doses, underline indicates the second-best mean, and † indicates the third-best mean. The tested doses are representation settings inherited from Tahoe-100M and should not be interpreted as calibrated PDX exposures.
Benchmark
Drug
Patient+Drug
PerturbRx
Δ AUROC
TCGA-508
Cisplatin
0.648
0.631
-0.017
Docetaxel
0.636
0.711
+0.075
Fluorouracil
0.553
0.562
+0.009
Gemcitabine
0.462
0.472
+0.010
Paclitaxel
0.605
0.715
+0.110
TCGA-186
Cisplatin
0.669
0.684
+0.016
Appendix
Table 20: Drug-specific incremental utility of the transferred transition representation across target benchmarks. Values report mean AUROC across the corresponding outer folds or evaluation seeds. Δ AUROC is defined as AUROCPerturbRx−AUROCPatient+Drug . Positive values indicate improved discrimination after adding the predicted transition.
Benchmark
Factor
ρ
pexact
TCGA-508
Morgan ECFP4 Tanimoto
+0.900
0.083
ChemBERTa cosine
+0.900
0.083
TCGA-186
Morgan ECFP4 Tanimoto
+0.100
0.950
ChemBERTa cosine
+0.500
0.450
PDX
Morgan ECFP4 Tanimoto
-0.067
0.865
ChemBERTa cosine
-0.030
0.946
Appendix
Table 21: Rank associations between drug-specific transition utility, measured by Δ AUROC, and candidate explanatory factors. For chemical similarity, p -values are based on two-sided exact permutation tests. PDX target-domain support variables use the same exact permutation framework over the ten-drug PDX benchmark.
Figure 6: PDX responder prevalence and drug-specific transition utility. Each point represents one drug in the shared 10-drug PDX benchmark. Transition utility is measured as the change in drug-specific AUROC after augmenting the static patient–drug representation with the predicted transition. The observed responder fraction was positively rank-associated with transition utility ( ρ=0.673 , two-sided exact permutation p=0.039 ). The two largest decreases occurred for Tamoxifen and Ruxolitinib, which also had the lowest observed responder fractions.
Accurate prediction of patient-specific therapeutic response from pre-treatment transcriptomes is hindered by the scarcity of matched clinical response labels and post-treatment molecular profiles. Preclinical transfer-learning models can simulate drug-induced expression changes but are often hard to interpret and unstable, whereas knowledge-graph methods provide mechanistic context yet remain static and fail to capture drug-induced transcriptomic perturbation dynamics. We propose PREDIKTOR, a patient-centered multi-view framework that aligns a personalized network view with a transferable transcriptomic perturbation view to predict clinical drug response. For each patient, we construct an individualized gene regulatory network from tumor expression using DysRegNet and augment it with drug-target links from DrugBank; a graph neural encoder yields a drug-centric, mechanistically grounded embedding. In parallel, a frozen condition-specific gene-gene attention model pretrained on LINCS L1000 generates a simulated post-perturbation transcriptomic profile for the same patient-drug pair. We align the two views in a shared latent space via a CLIP-style contrastive objective with drug-context hard negatives, then concatenate the representations for end-to-end response classification. On TCGA, PREDIKTOR consistently outperforms state-of-the-art baselines under patient-, drug-, and tissue-split evaluations, and transfers zero-shot to the I-SPY2 trial, improving AUROC by 5.6% over competing methods. The aligned embeddings yield stable gene and pathway attributions that recover known mechanisms, supporting actionable and interpretable precision oncology.
Dongmin Bang, Sugyun An, Inyoung Sung +3
Interdisciplinary Program in Bioinformatics, Seoul National University, Seoul, Republic of Korea · AIGENDRUG Co., Ltd., Seoul, Republic of Korea · BK21 FOUR Intelligence Computing, Seoul National University, Seoul, Republic of Korea +2
Predicting drug response in patients from preclinical data remains a major challenge in precision oncology due to the substantial biological gap between in vitro cell lines and patient tumors. Rather than aiming to improve absolute in vitro prediction accuracy, this work examines whether explicitly separating representation learning from task supervision enables more sample-efficient adaptation of drug-response models to patient data under strong biological domain shift. We propose a staged transfer-learning framework in which cellular and drug representations are first learned independently from large collections of unlabeled pharmacogenomic data using autoencoder-based representation learning. These representations are then aligned with drug-response labels on cell-line data and subsequently adapted to patient tumors using few-shot supervision. Through a systematic evaluation spanning in-domain, cross-dataset, and patient-level settings, we show that unsupervised pretraining provides limited benefit when source and target domains overlap substantially, but yields clear gains when adapting to patient tumors with very limited labeled data. In particular, the proposed framework achieves faster performance improvements during few-shot patient-level adaptation while maintaining comparable accuracy to single-phase baselines on standard cell-line benchmarks. Overall, these results demonstrate that learning structured and transferable representations from unlabeled molecular profiles can substantially reduce the amount of clinical supervision required for effective drug-response prediction, offering a practical pathway toward data-efficient preclinical-to-clinical translation.
Camille Jimenez Cortes, Philippe Lalanda, German Vega
AI-based cellular perturbation prediction is essential to accelerate drug discovery and therapeutic identification. However, perturbation-specific signals are generally obscured by cellular complexity, making it difficult for existing methods to correctly capture perturbation effects. To address this problem, we propose PerturbEnergy, which introduces a conditional energy-based prior over perturbation representations. It assigns lower energy (higher probability) to representations compatible with the given perturbation and higher energy to others, thereby explicitly tilting the latent space toward perturbation-preferred regions. Besides, PerturbEnergy can be integrated into autoencoder architectures as we find its Evidence Lower Bound (ELBO) is equivalent to the ELBO of the complete autoencoder. We further prove that optimizing PerturbEnergy increases the lower bound on the conditional mutual information between representation and perturbation. Extensive experiments on four benchmark datasets demonstrate the effectiveness of our PerturbEnergy: it improves perturbation recognition especially in latent space with improvement of up to 83.2%; It also achieves state-of-the-art performance for predicting top-50/100/200 differentially expressed gene(DEG) response and remaining competitive on highly variable gene(HVG) prediction. The code is provided: https://github.com/BlueWhaleLab/PerturbEnergy
Yuchen Yan, Yatao Bian
Department of Computer Science, National University of Singapore