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

1 INTRODUCTION

• Strong overall-response agreement can coexist with weak prediction of reproducible drug-specific differences, while 89.6% of compounds occur in only one dataset.These findings motivate sharing chemical supervision and directly targeting matched-condition differences between compounds.Source: §1 · p. 2
• Bison connects native gene panels through shared representations and combines context-dependent reference responses with molecular deviations learned using matched drug-contrast supervision.Two discrete diffusion models learn the context-response and molecular-effect branches; a shared gene-query decoder restores predictions to native panels.Source: §1 · p. 2
• Jointly trained Bison achieves the highest mean overall-response and drug-contrast Pearson correlations across eight datasets, and raises mean drug-contrast correlation by 27.4% over dataset-specific training.The comparison includes 11 methods trained independently per dataset; joint-training gains occur across all eight datasets.Source: §1 · p. 2
• Case studies show that close analogs from other screens can improve predictions for unseen compounds.Bison composes molecular deviations with context-dependent responses while preserving native gene measurements.Source: §1 · p. 2

Chemical perturbation models

• Chemical perturbation models range from fingerprint-based ridge regression to conditional generative, disentanglement, flow, transport, and attention-based approaches.Recent frameworks LPM and State emphasize learning across heterogeneous experiments and cellular contexts.Source: §2 · pp. 2–3

Heterogeneous gene panels and perturbation benchmarks

• Gene-aware, set-based, and latent-code approaches address heterogeneous gene panels, while benchmarks and harmonized resources support shared evaluation across perturbation datasets.The paper combines native gene panels with evaluation of overall responses and drug-specific differences for globally unseen compounds.Source: §2 · p. 3

3.1 DATA AND PREDICTION PROTOCOL

• The benchmark retains 16,771 compounds across eight datasets and 131 cellular contexts, with 14 native gene-panel views.Six datasets provide Full and 2,000-gene HVG panels, while the two L1000 datasets provide 978 landmark genes each.Source: §3.1 · p. 3
• Compounds are globally split into TRAIN/VAL/TEST at approximately 70%/10%/20%, and 11 baseline families are fitted separately by dataset and view.Experiments use three independent global compound splits; treated records follow their compound’s assignment.Source: §3.1 · p. 3

3.2 EVALUATION METRICS

• Drug-contrast correlation centers observed and predicted responses separately across compounds, measuring agreement in drug differences while canceling shared additive components.The evaluation also reports repeatability by comparing model drug-contrast correlation with same-drug repeat agreement.Source: §3.2 · p. 4

3.3 OVERALL FIT CAN MISS DRUG DIFFERENCES

• Even the strongest baseline in each dataset falls below same-drug repeat agreement, and CMonge ranks fourth in overall response but last in drug contrast.Although 89.6% of compounds occur in one dataset, pooling TRAIN compounds increases close-analog coverage on OP3 and Tahoe; direct neighbor copying nevertheless lowers mean contrast from 0.046 to 0.040.Source: §3.3 · p. 4

4 METHODS

• Bison connects native gene panels through shared representations and composes a context-dependent reference response with a molecular deviation learned by two discrete diffusion models.The molecular deviation subtracts a control-coded molecular prediction from the query-drug prediction; the sum is calibrated afterward.Source: §4 · p. 4, Figure 4 · p. 5

4.1 SHARED ENCODING AND DECODING

• An attention-based encoder and finite scalar quantization create shared discrete codes, which gene-identity queries decode to each requested native panel.The encoder and decoder are pretrained by joint reconstruction on Full and HVG panels, then frozen; both diffusion models share stable coordinates while retaining view-specific input and output genes.Source: §4.1 · p. 5

4.2 MODELING THE SHARED RESPONSE

• The context-response DLM learns complete conditional profiles, using control-matched features and experimental conditions to retain shared response structure across query compounds.Masked-code reconstruction and a decoded-response correlation loss train it; at composition time its molecular input is fixed to a reference condition.Source: §4.2 · p. 5

4.3 LEARNING MOLECULAR DIFFERENCES

• The molecular-effect DLM uses context-modulated chemical tokens and matched contrasts to learn within-context drug differences and drug-by-context interactions.Within-context comparisons share cellular background, plate, dose and time; interaction supervision removes drug and context marginal means and applies only where matched grids exist, without requiring every drug in every dataset.Source: §4.3 · p. 6

4.4 COMPOSING THE PREDICTIONS

• Bison composes the context-model output with the molecular model’s query-drug deviation from a dataset-specific control reference.The globally fixed molecular-difference weight is α = 0.25; the reference response is restored once, and the untruncated composition proceeds to calibration.Source: §4.4 · p. 6

4.5 RESPONSE CALIBRATION

• For count-based datasets, Bison calibrates response magnitude using TRAIN condition-mean effects relative to matched plate controls, then clips predictions at zero.It also zeros sufficiently small predictions for genes detected in fewer than 10% of TRAIN profiles; L1000 retains the uncalibrated composition with nonnegative clipping.Source: §4.5 · pp. 6–7

5.1 COMPARISON WITH EXISTING PERTURBATION MODELS

• Bison reaches overall-response Pearson 0.4110 versus 0.3325 for State and drug-contrast Pearson 0.0478 versus 0.0342, leading on eleven and eight views, respectively.The dataset-level drug-contrast comparison is varied: competing methods score higher on SciPlex, Novartis, and Tahoe.Source: Table 1 · p. 7, §5.1 · p. 8
• On Cell-Eval, Bison leads mean Pearson Δ, perturbation-discrimination, and MAE across twelve count-based views, but State and Prophet lead two differential-expression recovery measures.Source: Table 2 · p. 8, §5.1 · p. 8

5.2 JOINT TRAINING IMPROVES UNSEEN-COMPOUND PREDICTION

• Joint training raises mean overall response 2.0% and mean drug contrast 27.4%; drug contrast improves in all eight datasets, while overall response improves in seven.With joint context-response training fixed, joint molecular-effect training raises drug contrast from 0.0397 to 0.0506; combining both gives the highest overall macro, 0.3991.Source: Table 3 · p. 8, §5.2 · pp. 8–9, Figure 5 · p. 8

5.3 INTERPRETING CROSS-DATASET RESPONSE PREDICTION

• Removing Novartis yields the lowest Panobinostat score among source removals in SciPlex, Tahoe, and L1000 P1, where Novartis TRAIN analogs have Tanimoto similarity 0.65–0.81.Two of the closest analogs retain Panobinostat’s hydroxamic acid.Source: Figure 6 · p. 9, §5.3 · p. 9
• The exploratory Hallmark analysis finds 41 pathways with significant molecule-level predictions in at least two datasets; twelve recur in at least six, spanning proliferation, stress and inflammatory signaling, and metabolism.Source: Figure 6 · p. 9, §5.3 · p. 9

5.4 ABLATIONS

• Ablations support separating shared responses from molecular differences and modeling cellular context; context modulation and interaction loss provide most drug-contrast gain, with a smaller increment from within-plate contrast supervision.Adding molecular correction improves overall prediction, and moderate scaling outperforms an unscaled deviation.Source: §5.4 · p. 9

• Bison preserves native gene panels while sharing chemical-response learning across datasets.Source: §6 · p. 10
• Joint training improves both overall response agreement and prediction of drug-specific differences, which matched contrasts make explicit in learning and evaluation.Source: §6 · p. 10
• The conclusion frames integration as aligning the prediction task as well as combining measurements.Source: §6 · p. 10

• The authors state that experiment-reproduction code will be provided in the supplementary files.Source: REPRODUCIBILITY STATEMENT · p. 11
• Section 3 and Appendix A describe data, panels, splits, control matching, and metrics; Section 4 and Appendix B document model and training details.Source: REPRODUCIBILITY STATEMENT · p. 11
• The main benchmark reports means and sample standard deviations over three independent global compound splits.Source: REPRODUCIBILITY STATEMENT · p. 11

• Generative AI tools were used for manuscript drafting and revision.Source: AI USE STATEMENT · p. 11
• Generative AI tools were also used for LaTeX formatting, experiment implementation, and analysis.Source: AI USE STATEMENT · p. 11
• The authors take responsibility for the final manuscript, experimental results, and conclusions.Source: AI USE STATEMENT · p. 11

A.1 DATASET COVERAGE AND GENE PANELS

• The eight datasets contribute six count-based Full and HVG views each and one landmark view from each L1000 dataset, forming a fourteen-view benchmark.The released HVG panels use all chemically eligible profiles, while model fitting and reference estimation use TRAIN data.Source: §A.1 · p. 13, Table 4 · p. 13

A.2 SCORING AND AGGREGATION

• Overall response is scored after subtracting context-, plate- and time-matched TRAIN controls, whereas drug contrast centers responses across at least three compounds within a matched group.Drug-contrast scoring requires exact TRAIN control matches; the overall scorer has specified fallback controls for L1000 Phase 1.Source: §A.2 · p. 13

A.3 REPEATABILITY REFERENCE

• Bison’s drug-contrast agreement is 45.0% of repeat-measurement agreement across datasets, averaged within views, datasets and splits.This compares predictions with another noisy measurement, not a hard prediction ceiling; VCPI1/2 use paired wells instead of cross-plate repeats.Source: §A.3 · p. 14, Table 5 · p. 14

B.1 BISON ARCHITECTURE AND TRAINING

• A shared gene representation connects native panels, while a context-response diffusion model learns complete profiles and a molecular-effect model learns deviations using matched drug-contrast supervision.The tokenizer is frozen for both models; the context branch is selected on overall response, and the molecular branch on drug contrast.Source: §B.1 · p. 15

B.2 PREDICTION AND CALIBRATION

• Inference composes reference-conditioned model calls, then applies dataset/view amplitude calibration and, for count-based datasets, detection-aware zeroing.L1000 retains the uncalibrated composition with nonnegative clipping; calibration scales the complete response around the exact control mean.Source: §B.2 · pp. 15–16

B.3 BENCHMARK EXECUTION

• Baseline models are fitted separately by dataset and native gene view, with TRAIN-fitted preprocessing and a common evaluator for control subtraction, replicate aggregation and scoring.Baselines retain their objectives and schedules, with checkpoint selection on validation compounds or terminal checkpoints for fixed-schedule runs.Source: §B.3 · p. 16

B.4 BASELINE ADAPTATIONS

• Baseline adaptations preserve each model’s core architecture or objective while supplying benchmark-compatible molecular features, controls and native-panel predictions.Adaptations include Morgan fingerprints for Ridge and PerturbNet, dataset-specific training instead of L1000 pretraining for chemCPA, and native-panel decoding for several models.Source: §B.4 · p. 17

B.5 RUN AND METRIC AVAILABILITY

• XPert’s larger Full-panel runs exceed the memory budget, while CellFlow’s broad-coverage L1000 runs exceed the runtime budget.Partial macros are marked and excluded from complete-coverage ranking.Source: §B.5 · p. 18
• Cell-Eval input-scale validation or condition exclusions affect thirteen method–view–split cells; Bison uses all conditions, while XPert covers only six HVG views and OP3 Full.The native benchmark uses the complete original prediction sets, and incomplete-coverage macros are excluded from complete-coverage ranking.Source: §B.5 · p. 18, §C.1 · p. 19

C.1 COMPLETE BENCHMARK RESULTS

• The benchmark reports overall-response and drug-contrast Pearson for native views across three independent global compound partitions, with split-and-training means and sample standard deviations.Bison holds its training initialization fixed but retrains with each split-specific tokenizer; the macros average the same available views within each split.Source: §C.1 · pp. 18–19, Table 6 · p. 19, Table 7 · p. 20

C.2 JOINT AND DATASET-SPECIFIC TRAINING

• 0.05063 versus 0.03974: joint training raises mean drug-contrast correlation by 27.4% over molecular SOLO training, with the context-response branch fixed to JOINT.The comparison averages views within datasets and weights all eight datasets equally.Source: §C.2 · p. 20

C.3 ABLATIONS

• 0.4046 to 0.4123: adding molecular deviation raises uncalibrated canonical overall Pearson, with a paired three-split increment of 0.0077 ± 0.0051.The fixed α = 0.25 gives the largest tested overall-Pearson gain, +0.0091; α = 1 reduces the shared response.Source: §C.3 · p. 21, Table 8 · p. 21
• Context modulation and interaction supervision raise three-split drug contrast from 0.0414 to 0.0477 and native overall Pearson from 0.4082 to 0.4097 before calibration.The factorial comparison attributes most aggregate gain to context components, with a smaller increment from within-plate contrast supervision.Source: §C.3 · p. 21, Table 8 · p. 21

C.4 CALIBRATION EFFECTS

• Amplitude calibration raises PDS-L1 from 0.7028 to 0.7778 and PDS-L2 from 0.6909 to 0.7647, while increasing MAE from 0.0537 to 0.0569.Detection-aware zeroing raises overlap@100 from 0.0137 to 0.1670 but lowers DE direction agreement from 0.7797 to 0.7006; overall Pearson changes little.Source: §C.4 · pp. 21–22, Table 9 · p. 22

C.5 CONTROL-REFERENCE SENSITIVITY

• Bison retains the highest complete-coverage mean under both control-reference conventions: 0.5637 with pooled controls and 0.4110 with plate matching.The same final predictions are scored with pooled context–time or matched context–plate–time TRAIN controls.Source: §C.5 · p. 22, Table 10 · p. 22

D.1 SOURCE CONTRIBUTIONS AND STRUCTURAL PROXIMITY

• Removing any dataset lowers its own drug contrast; Novartis contributes the largest mean gain to other datasets, while removing VCPI2 or either L1000 source raises OP3 by 0.06–0.09.The source-removal comparison uses the first split’s VAL compounds and a shared tokenizer and context-response branch.Source: §D.1 · p. 23, Figure 7 · p. 23
• Close Novartis analogs lose more on SciPlex, Tahoe and L1000 P1 after source removal, but weak rank associations and losses without close analogs show proximity does not account for all gains.The reported rank associations are Spearman −0.15 to 0.00.Source: §D.1 · p. 23

D.2 DIRECT RESPONSE COPYING FROM STRUCTURAL NEIGHBORS

• Expanding the direct-copying pool from target-dataset to all-dataset TRAIN compounds lowers mean drug contrast from 0.046 to 0.040, improving SciPlex and VCPI1 but reducing the other six datasets.The comparison uses five TRAIN neighbors per gene; broader structural coverage alone does not consistently improve direct response copying under this rule.Source: §D.2 · p. 24, Table 11 · p. 24

D.3 COMPLETE HALLMARK PATHWAY RESULTS

• 198 pathway–dataset pairs meet the pooled FDR and positive-correlation criterion, with 41 pathways qualifying in at least two datasets and 34 in at least four.The map covers all 44 pathways eligible in at least one dataset; twelve qualify in at least six datasets.Source: §D.3 · p. 24, Figure 8 · p. 25