Bison: Cross-Dataset Learning for Unseen-Compound Perturbation Prediction

Yunfan Liu; Kasra Ghorbani; Yufei Huang; Zicheng Liu; Jiangbin Zheng; Jingbo Zhou; Shaorong Chen; Chang Yu

cs.LGSep 26, 2026

Bison: Cross-Dataset Learning for Unseen-Compound Perturbation Prediction

Authors: Yunfan Liu, Kasra Ghorbani, Yufei Huang, Zicheng Liu, Jiangbin Zheng, Jingbo Zhou, Shaorong Chen, Chang Yu,

Organizations: Zhejiang University · Westlake University · Beihang University · Nanjing University

Abstract

Predicting transcriptional responses to unseen compounds is limited by fragmented chemical coverage and heterogeneous experimental platforms and gene panels. To assess molecular generalization across these settings, we build on Chem-PerturBridge to benchmark eight datasets with 16,771 compounds, withholding test compounds from every training dataset. This comparison reveals that high overall response agreement can coexist with weak prediction of drug-specific differences, despite reproducible signals across repeated measurements. To exploit complementary chemical supervision while targeting these differences, we introduce Bison: a shared gene representation connects native panels, while two discrete diffusion models compose context-dependent responses with molecular deviations learned through matched drug-contrast supervision. A single Bison model jointly trained across all eight datasets achieves the highest mean overall-response and drug-contrast Pearson correlations on the full benchmark in comparison with 11 methods trained independently per dataset. Compared with dataset-specific training of the same architecture, joint training increases mean drug-contrast correlation by 27.4%, with gains across all eight datasets and improvements in overall response prediction. These results demonstrate how matched drug contrasts turn complementary screens into shared molecular supervision for unseen-drug response prediction while preserving native gene measurements.

Figures & tables

Figure 1: The task and the evaluation challenge. (a) Joint training combines observed drug-response profiles across datasets with different native gene panels. Test drugs are absent from every training dataset; crosses denote unmeasured profiles. (b) The shared background dominates overall responses and is easier to fit than drug-specific variation. Fitting this common component pulls predictions for different compounds toward the same profile, suppressing drug-specific differences.

Figure 2: Chemical space across datasets. Shared Morgan-fingerprint t-SNE of 16,771 compounds. Colors highlight each dataset against the gray background; $n$ denotes compound counts.

Figure 3: Overall agreement and drug-specific prediction. (a) Overall response and drug contrast for ten complete-coverage baselines: three-split means and sample SDs over 14 native views. (b) Drug contrast of the best available baseline and a same-drug repeat, averaged over views and splits on matched repeat-eligible samples. Novartis uses HVG; VCPI repeats are within-plate wells.

Figure 4: Bison architecture. (a) Shared encoding and gene-query decoding connect native panels. (b) The molecular DLM uses cellular context to modulate chemical tokens and condition context memory and learns both within-context drug differences and drug–context interactions. The shared prediction includes the control reference; the drug deviation subtracts a control-coded molecular prediction. Their sum is calibrated afterward.

L

and

K

denote layer and chemical-token counts; training losses are defined in Section 4.3 .

Overall response
Method	OP3	SciPlex	VCPI1	VCPI2	Novartis	Tahoe	P1	P2	Macro
Global Mean	7.1 (1.1)	17.8 (0.7)	18.1 (0.3)	15.5 (0.1)	29.8 (0.8)	20.3 (1.1)	8.7 (0.0)	4.1 (0.1)	16.4 (0.2)
Linear	28.9 (2.9)	44.6 (1.4)	12.7 (0.1)	5.9 (0.5)	25.4 (0.9)	36.2 (2.7)	15.6 (0.3)	2.9 (0.3)	23.3 (0.3)
Ridge	39.0 (1.6)	49.8 (0.6)	18.5 (0.4)	15.7 (0.2)	30.7 (0.8)	42.0 (1.0)	15.6 (0.3)	6.3 (0.1)	29.5 (0.3)
chemCPA	33.1 (0.7)	44.1 (0.7)	12.7 (0.1)	11.0 (0.2)	23.4 (1.6)	36.8 (1.7)	14.3 (0.2)	6.0 (0.6)	24.5 (0.2)
PRnet	26.4 (1.5)	44.3 (0.6)	12.5 (3.2)	13.4 (0.1)	28.5 (0.3)	33.1 (1.5)	15.5 (0.2)	5.1 (0.2)	24.1 (0.6)

Table 1: Benchmark results on globally unseen compounds. Pearson

\times 100

, mean (sample SD) over three compound splits. Dataset columns average Full/HVG (L1000: landmark genes); Macro weights fourteen native views equally. Top: plate-matched overall response; bottom: drug contrast. Bold/underline: best/second-best complete-coverage mean. ∗ Partial coverage, not ranked; dashes: unavailable; NA: undefined. Per-view results and resource limits are in Appendices B – C .

Method	Views	Pearson $\Delta\uparrow$	MAE $\Delta\downarrow$	PDS L1 $\uparrow$	PDS L2 $\uparrow$	PDS cosine $\uparrow$	DE direction $\uparrow$	Overlap @100 $\uparrow$
Global Mean	12	18.11 (0.21)	13.16 (0.02)	69.33 (0.15)	67.44 (0.06)	70.43 (0.23)	83.49 (1.77)	0.79 (0.02)
Linear †	12	25.67 (0.43)	9.51 (0.08)	69.05 (0.43)	67.01 (0.25)	72.72 (0.32)	74.36 (0.39)	1.33 (0.05)
Ridge	12	32.67 (0.43)	6.48 (0.08)	72.47 (0.14)	70.58 (0.16)	77.76 (0.06)	82.32 (0.81)	0.98 (0.06)
chemCPA †	12	26.85 (0.26)	6.99 (0.05)	71.12 (0.41)	69.76 (0.46)	76.56 (0.45)	53.39 (0.84)	15.08 (0.17)
PRnet	12	26.38 (0.73)	7.21 (0.09)	71.37 (0.69)	69.31 (0.80)	75.05 (0.57)	53.74 (0.94)	19.29 (0.19)
biolord	12	31.23 (1.21)	6.80 (0.18)	72.19 (0.29)	70.15 (0.37)	76.43 (0.26)	68.04 (0.32)	0.98 (0.03)

Table 2: Official Cell-Eval results. Mean (sample SD) across three compound splits; values

\times 100

. Bison includes amplitude calibration and detection-aware zeroing. Twelve Full/HVG views of six count-based datasets; DE direction uses ten views. Bold: best complete-view mean. ∗ Partial view coverage, unranked. † Some evaluations use valid-condition subsets; Bison uses all conditions.

Figure 5: JOINT–SOLO drug contrast. Native-view TEST means; dashed line: equal scores. P1/P2: L1000.

	Overall response				Drug contrast
Dataset	S/S	S/J	J/S	J/J	S	J
OP3	39.66	39.41	40.44	40.10	2.50	7.16
SciPlex	45.36	46.25	46.37	47.18	9.46	9.58
VCPI 1	33.36	33.33	33.27	33.50	2.34	3.37
VCPI 2	33.46	33.31	33.52	33.61	2.24	2.32
Novartis	49.71	50.48	52.11	52.60	3.91	4.82

Table 3: Joint training by component. TEST Pearson (

\times 100

), averaged over views. S/J: separate/joint training. Overall: context/molecular; drug: molecular (joint context). Macro: equal dataset weights.

Figure 6: Molecular evidence and pathway signals. (a) Panobinostat and Novartis TRAIN analogs (

T

: Tanimoto similarity). (b) Panobinostat VAL drug contrast; whiskers span other source removals. (c) Twelve Hallmark pathways predicted in at least six datasets. Color: truth-SD-scaled HVG TEST Pearson; dots: pooled VAL/TEST FDR

<0.10

and positive correlation in both splits. Gray cells indicate unavailable values; P1/P2 denote L1000 phases.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Profiles	Molecules	Contexts	Full genes	HVG/landmark
OP3	1,813	138	4	5,198	2,000
SciPlex	4,974	186	3	20,251	2,000
VCPI1	27,517	2,272	1	20,629	2,000
VCPI2	18,139	1,488	1	20,629	2,000
Novartis	46,748	3,770	1	20,038	2,000
Tahoe	67,018	376	50	20,573	2,000

Appendix

Table 4: Benchmark coverage. Profiles include treated and control bulk/pseudobulk records. Full counts measured genes in the benchmark vocabulary; the two L1000 datasets each contribute one 978-gene landmark view. Molecules and cellular contexts are deduplicated in the total row.

Dataset	Eligible units	Repeat $r$	Bison $r$	Bison / repeat (%)
OP3	268–353	0.228	0.091	40.2
SciPlex	870–944	0.147	0.078	53.1
VCPI1	2,498–2,692	0.030	0.024	80.2
VCPI2	1,610–1,723	0.030	0.016	53.0
Novartis	9,276–9,430	0.284	0.039	13.9
Tahoe	3,010–3,492	0.298	0.056	18.5

Appendix

Table 5: Drug-contrast agreement relative to repeated measurements. Correlations and ratios are means across three splits. Unit counts span native views and splits; VCPI1/2 use paired wells, and other datasets use cross-plate repeats.

Method	View	OP3	SciPlex	VCPI1	VCPI2	Novartis	Tahoe	P1	P2	Macro
Global Mean	Full	6.9 (1.0)	21.2 (0.6)	20.5 (0.3)	18.0 (0.1)	32.2 (0.7)	20.2 (0.9)	8.7 (0.0)	4.1 (0.1)	16.4 (0.2)
Global Mean	HVG	7.3 (1.1)	14.3 (0.8)	15.7 (0.3)	13.1 (0.1)	27.5 (0.8)	20.4 (1.2)	8.7 (0.0)	4.1 (0.1)	16.4 (0.2)
Linear	Full	27.8 (2.3)	45.9 (1.5)	14.3 (0.1)	6.8 (0.4)	26.9 (0.8)	35.6 (2.7)	15.6 (0.3)	2.9 (0.3)	23.3 (0.3)
Linear	HVG	30.0 (3.5)	43.3 (1.4)	11.0 (0.1)	5.1 (0.6)	23.8 (0.9)	36.8 (2.8)	15.6 (0.3)	2.9 (0.3)	23.3 (0.3)
Ridge	Full	37.5 (1.3)	50.9 (0.5)	20.8 (0.4)	17.9 (0.3)	33.0 (0.8)	41.6 (1.1)	15.6 (0.3)	6.3 (0.1)	29.5 (0.3)
Ridge	HVG	40.5 (2.0)	48.7 (0.6)	16.3 (0.4)	13.5 (0.2)	28.5 (0.9)	42.5 (0.9)	15.6 (0.3)	6.3 (0.1)	29.5 (0.3)

Appendix

Table 6: Overall response on globally unseen compounds. Native plate-matched Pearson

\Delta\times 100

, mean (sample SD) across three compound splits. Full/HVG rows are blue/orange and ranked separately; P1/P2 denote L1000 Phases 1/2, each with one landmark view. Bold and underline mark the best and second-best displayed means, respectively.

Method	View	OP3	SciPlex	VCPI1	VCPI2	Novartis	Tahoe	P1	P2	Macro
Global Mean	Full	NA	NA	NA	NA	NA	NA	NA	NA	NA
Global Mean	HVG	NA	NA	NA	NA	NA	NA	NA	NA	NA
Linear	Full	0.9 (1.5)	4.5 (0.7)	0.3 (0.3)	0.1 (0.4)	1.8 (0.3)	1.9 (1.5)	2.0 (0.4)	0.5 (0.4)	1.8 (0.1)
Linear	HVG	0.9 (1.5)	6.5 (1.2)	0.4 (0.3)	0.1 (0.5)	2.1 (0.3)	2.5 (1.9)	2.0 (0.4)	0.5 (0.4)	1.8 (0.1)
Ridge	Full	3.7 (4.1)	6.6 (1.7)	1.3 (0.4)	1.3 (0.3)	3.6 (0.3)	2.0 (0.9)	2.2 (0.4)	1.9 (0.2)	3.3 (0.3)
Ridge	HVG	4.6 (4.6)	8.9 (2.2)	1.7 (0.5)	1.6 (0.3)	4.3 (0.4)	2.3 (1.0)	2.2 (0.4)	1.9 (0.2)	3.3 (0.3)

Appendix

Table 7: Drug-contrast prediction on globally unseen compounds. Within-plate drug-contrast Pearson

\times 100

, mean (sample SD) across three TEST splits. Full/HVG rows are blue/orange and ranked separately; P1/P2 denote L1000 Phases 1/2. Bold and underline mark the best and second-best displayed means, respectively.

Predictor	Overall Pearson $\Delta$	Increment	Drug contrast
Context-response DLM	0.4046 (0.0033)	–	–
Bison (two DLMs)	0.4123 (0.0055)	$+0.0077$ (0.0051)	0.0477 (0.0041)

Appendix

Table 8: Ablations of Bison. (a) Uncalibrated canonical scores: mean (sample SD) and paired increments. (b) Overall-Pearson gain over

\alpha=0

with both DLMs fixed;

\dagger

: fixed coefficient. (c) Drug contrast: Context combines context modulation and interaction loss; Contrast is the within-plate loss. Panels (b,c) weight datasets equally after averaging native views.

Metric	Views	Uncalibrated	+ amplitude	+ detection zeros
Pearson $\Delta\uparrow$	12	0.4321 (0.0055)	0.4336 (0.0055)	0.4337 (0.0055)
MAE $\Delta\downarrow$	12	0.0537 (0.0006)	0.0569 (0.0008)	0.0568 (0.0008)
MSE $\Delta\downarrow$	12	0.0112 (0.0005)	0.0120 (0.0005)	0.0120 (0.0005)
PDS-L1 $\uparrow$	12	0.7028 (0.0030)	0.7778 (0.0049)	0.7782 (0.0048)
PDS-L2 $\uparrow$	12	0.6909 (0.0026)	0.7647 (0.0052)	0.7647 (0.0052)
PDS-cosine $\uparrow$	12	0.8430 (0.0040)	0.8458 (0.0038)	0.8458 (0.0038)

Appendix

Table 9: Effects and tradeoffs of postprocessing. Three-split means and sample SD on fixed common Cell-Eval views. Values are on their original scales.

Method	Context–time	Plate matched	Views
Ridge	0.2320 (0.0049)	0.2954 (0.0034)	14/14
chemCPA	0.1492 (0.0034)	0.2445 (0.0019)	14/14
PRnet	0.1603 (0.0025)	0.2408 (0.0062)	14/14
biolord	0.2029 (0.0113)	0.2856 (0.0103)	14/14
PerturbNet	0.1536 (0.0127)	0.2400 (0.0051)	14/14
CellFlow	0.2493 ∗ (0.0199)	0.3288 ∗ (0.0110)	12/14

Appendix

Table 10: Sensitivity to the control reference. Native Pearson

\Delta

, mean (sample SD) across three TEST splits. Identical stored predictions are scored against pooled context–time and matched context–plate–time controls. Each split averages a fixed common set of official views. Partial means (*) are gray and excluded from full-coverage ranking; bold/underline mark the best/second-best mean. Bison uses the same corrected molecular reference as Table 1 .

Figure 7: Dataset-level source-removal effects. Change in VAL drug contrast ( $\times 100$ ) relative to joint training at original schedule step 20,000. Rows identify the removed source and columns the evaluated dataset; outlined diagonal cells remove the target dataset itself. The right column averages each row over the other seven datasets.

Dataset	Own-dataset pool	Eight-dataset pool	Change
OP3	0.121	0.079	$-0.042$
SciPlex	0.031	0.047	$+0.016$
VCPI1	0.018	0.023	$+0.005$
VCPI2	0.021	0.015	$-0.005$
Novartis	0.049	0.041	$-0.007$
Tahoe	0.047	0.036	$-0.011$

Appendix

Table 11: Direct copying from structural neighbors. Drug-contrast Pearson on the first split's VAL compounds, using five TRAIN neighbors per gene. Full/HVG routes are averaged within each dataset; both L1000 routes use the landmark panel.

Figure 8: Complete Hallmark pathway map. All 44 pathways with eligible gene coverage in at least one dataset. Colors show TEST molecule-level correlation of truth-SD-scaled pathway effects; dots mark pooled VAL/TEST FDR $<0.10$ and positive correlations in both splits. Gray denotes ineligible gene coverage. The right column counts marked datasets, and the separator identifies the twelve pathways displayed in Figure 6 c. Rows are ordered by this count and then alphabetically.

Cross-Dataset Generalization

Cross-dataset generalization studies whether AI models retain accuracy, calibration, and usefulness on data from sources beyond their training set.

See all 135 papers

Chem-PerturBridge: a harmonized compendium of small molecule perturbation transcriptomic effects
May 29, 2026Artur Szałata, Olga Novitskaia, Maiia Shulman +3Cross-Dataset Generalization Molecular Representation Learning
Institute of Computational Biology, Helmholtz Center Munich · TUM School of Life Sciences Weihenstephan, Technical University of Munich · Institut Curie, INSERM U1331, Computational Oncology +1
Chem2Gen-Bench: Benchmarking Chemical-to-Genetic Translation in Perturbation Response Space
Jun 19, 2026Yuxiang Lin, Ying Chen
National Institute for Data Science in Health and Medicine, Xiamen University · School of Informatics, Xiamen University
LLM-Guided Retrieval for Prediction of Molecular Perturbation Responses
Aug 3, 2026Betty Xiong, Jan-Christian Huetter, Gabriele Scalia +2Single-Cell Perturbation Modeling Drug Discovery
Department of Biomedical Data Science Stanford University Stanford, CA 94305, USA · Biology Research & AI Development, Genentech DNA Way, South San Francisco, CA 94080, USA

May 29, 2026J/K move · Enter open · S save

Chem-PerturBridge: a harmonized compendium of small molecule perturbation transcriptomic effects

Artur Szałata, Olga Novitskaia, Maiia Shulman, Matthew Mella+2

Institute of Computational Biology, Helmholtz Center Munich · TUM School of Life Sciences Weihenstephan, Technical University of Munich · Institut Curie, INSERM U1331, Computational Oncology+1

Cross-Dataset Generalization Molecular Representation Learning

Open paper AskPDF arXiv

Large perturbation models require training data encompassing chemical, cellular, and assay diversity. Current transcriptomic resources for small-molecule modeling, however, are fragmented across technologies, metadata conventions, controls, doses, and preprocessing pipelines. We introduce Chem-PerturBridge, a harmonized multi-dataset resource comprising over 37k compounds, 136 cellular contexts, and 1.25M transcriptomic samples across eight assay types, with standardized identifiers, metadata, and replicate-aware condition-level effects. We use the resource to evaluate matched-condition agreement across datasets and replicate agreement within datasets. Matched same-compound conditions generally show weak agreement in fine-grained logFC rankings and magnitudes across most dataset pairs, often falling below same-context different-compound baselines. In contrast, logFC direction agreement is substantially more stable and usually exceeds these baselines. We further evaluate Chem-PerturBridge as a pretraining resource for compound representation learning. Under a compound-held-out OP3 evaluation split, embeddings pretrained on Chem-PerturBridge improve over L1000-only embeddings, Morgan fingerprints, and the descriptor-free OP3 baseline across metrics. An extensive molecule-holdout evaluation across 11 datasets further shows that models trained on Chem-PerturBridge outperform or match those that are not. Chem-PerturBridge therefore supports both diagnostic evaluation of cross-dataset signature agreement and model-oriented reuse of heterogeneous perturbation transcriptomic data.

Abstract

Cross-Dataset Generalization

Explore similar work

Chem-PerturBridge: a harmonized compendium of small molecule perturbation transcriptomic effects

Chem2Gen-Bench: Benchmarking Chemical-to-Genetic Translation in Perturbation Response Space

LLM-Guided Retrieval for Prediction of Molecular Perturbation Responses