ProteinTalks

The "virtual cell" ambition — a model that predicts how a cell will respond to a perturbation before anyone runs the experiment — has mostly been pursued with transcriptomics, because that is where the large perturbation datasets live. Proteins are closer to the phenotype and closer to most drug targets, but generating perturbation proteomics at the scale a foundation model needs has only recently become tractable through data-independent acquisition mass spectrometry.

ProteinTalks, from Westlake University with collaborators at DP Technology, the AI for Science Institute, and Peking University, takes that route. The group first generated the data: 18 breast cancer cell lines treated with 63 FDA-approved drugs and their combinations, sampled at 6, 24, and 48 hours, yielding 16,311 DIA-MS runs and more than 38 million high-quality perturbed protein measurements. They then built a model whose inductive bias matches the data. Because a proteome under drug pressure is a trajectory rather than a snapshot, the network's core is parameterized as a neural ordinary differential equation, integrated forward to predict the proteome at each measured timepoint.

Pretraining on that trajectory-prediction objective is what gives ProteinTalks its transferable representation of protein network dynamics; downstream tasks — drug efficacy, synergy, xenograft response, patient prognosis — reuse it.

#Key Features

• Trajectory prediction, not endpoint prediction: A fourth-order Runge-Kutta neural ODE evolves the latent proteome through time, so the model learns how networks move rather than where they land.
• Drug structure as a second input stream: A downstream module combines the predicted perturbed proteome with 881-dimensional molecular fingerprints, 55-dimensional physicochemical descriptors, and target annotations.
• Synergy screening with wet-lab confirmation: Predicted synergistic combinations were tested in CCK-8 cell viability assays, including bosutinib-tucatinib and bosutinib-abemaciclib pairs.
• Generalizes beyond cell lines: The same pretrained model predicts drug response in patient-derived tumour xenografts and improves prognosis prediction when combined with clinical breast cancer data.
• Interpretable predictions: Interpretability analysis surfaces the proteins and pathways behind a predicted response, with gene interference experiments confirming roles for AKR1C3 and CMPK1 in specific drug responses.

#Technical Details

The first module ingests baseline proteomes covering 5,585 proteins concatenated with perturbation descriptors, lifts them through a linear layer to 32 dimensions and a convolutional block to 128, then parameterizes a two-layer Softplus network with dropout 0.1 as a neural ODE solved with an rk4 integrator; a decoder maps predictions back to proteomics space under an MSE loss against measured timepoints. The second module feeds those predicted proteomes plus drug structure into a multilayer perceptron. On drug responsiveness, ProteinTalks reports AUROC 0.960, AUPRC 0.854, and accuracy 0.910, ahead of bootstrap, random forest, logistic regression, SGD, KNN, and DeepSynergy baselines; across single and combination therapy efficacy the paper reports AUROC 5–27% above comparators. Under leave-one-cell-line-out cross-validation, 16 of 18 cell lines hold AUROC at or above 0.9. On a held-out set of 98 additional anti-cancer compounds and 5,138 dose-response measurements, accuracy is 0.619 and AUROC 0.671 overall, rising to 0.844 and 0.840 once compounds with mechanisms of action absent from training are excluded. The underlying ProteinTalks datasets are distributed through db.prottalks.com.

#Applications

The clearest use is preclinical oncology triage: given an untreated proteomic profile of a tumour model, rank candidate drugs and drug pairs before committing to a screen. The synergy results point at combination discovery specifically, where the search space is quadratic and exhaustive testing is impractical. The xenograft and clinical-cohort results extend the same representation toward treatment selection, and the interpretability analyses give a route to resistance-mechanism hypotheses rather than a bare prediction.

#Impact

ProteinTalks is an argument that perturbation proteomics is now large enough to support foundation-model pretraining, and that differential-equation inductive biases are a natural fit when the underlying data are time courses. The honest scope is narrower than "virtual cell" suggests: pretraining covers breast cancer cell lines and 63 drugs, and the drop to 0.671 AUROC on compounds with unseen mechanisms of action shows the representation extrapolates within familiar pharmacology better than beyond it. The work is a preprint under all-rights- reserved terms, with the datasets shared through a companion database. Nothing else has been released: the GitHub account the paper points to hosts no public repositories, so there is neither code nor a trained checkpoint, and applying ProteinTalks means rebuilding it from the methods section.