UniCure

Predicting how a tumor will respond to a drug is a problem where the training data and the target population disagree. Perturbation atlases such as LINCS, sci-Plex, and CMap contain millions of drug-treated transcriptomes, but they are almost entirely immortalized cell lines, which lack a tumor microenvironment, intratumoral heterogeneity, and patient-specific pharmacodynamics. Encoder-decoder models trained on those atlases — DeepCE, chemCPA, TransiGen, PRnet — inherit the mismatch and translate poorly to individual patients.

UniCure, from Luonan Chen's group at the Shanghai Institute of Biochemistry and Cell Biology with collaborators at Second Military Medical University, DP Technology, and several Chinese hospitals and universities, attacks the gap from both sides. Rather than learning cell and drug representations from scratch, it composes two existing pretrained foundation models: UCE for cell state and Uni-Mol for molecular structure. The pretrained pair is then adapted with Low-Rank Adaptation (LoRA) so that broad biological and chemical knowledge is preserved rather than overwritten. On the clinical side, the authors generated their own patient-derived data — three-dimensional tumor-like cell clusters (PTCs) that retain stromal and immune components — and used roughly 800 such profiles to fine-tune the model for individual patients.

The result is a transcriptome-level drug response predictor that operates on cell lines, single cells, and patient tissue within one framework. It was posted to bioRxiv in June 2025 and has not yet been peer reviewed.

#Key Features

• Composed foundation models: Cell states come from UCE (itself built on ESM2 protein embeddings and 36 million cells) and drugs from Uni-Mol (209 million molecular conformations), rather than task-specific encoders.
• FlexPert cross-attention: A sliding-window cross-attention module handles variable numbers of drugs per sample, which is what lets one model cover single agents, doses, and combinations.
• Unpaired training with MMD loss: A Maximum Mean Discrepancy objective learns the distributional shift induced by treatment without requiring matched pre- and post-perturbation samples, and generalizes better to unseen compounds than likelihood or optimal-transport losses.
• LoRA batch harmonization: Fine-tuning UCE with LoRA collapses platform-driven batch effects, so profiles of the same cell line from different sources cluster together instead of splitting.
• Patient-level adaptation: A further fine-tuning stage on PTC and real-world clinical profiles produces per-patient predictions and enables response-based patient stratification.

#Technical Details

UniCure learns a map from an unperturbed transcriptome and a set of (SMILES, dose) pairs to the perturbed transcriptome. Training is staged: the first stage stabilizes the cell representation using LoRA-adapted UCE, a query projection, and the UniCure decoder; the second adds key-value projections and the FlexPert decoder to predict post-treatment state under the MMD loss. Perturbation training used LINCS 2020 plus sci-Plex 3 and sci-Plex 4, totaling over 1.8 million profiles spanning more than 22,000 compounds, 166 cell types, and 24 tissues. Held-out evaluation gave Pearson and Spearman correlations above 0.9 with R² above 0.8 on LINCS 2020, and Pearson above 0.9 on both sci-Plex sets.

Against PRnet and TransiGen retrained from scratch on identical splits, UniCure was consistently better on LINCS 2020 and sci-Plex 3 across Pearson, Spearman, and R² (Wilcoxon rank-sum, P < 0.0001). For patient-level evaluation the authors used cross-cancer fine-tuning — adapting on other tumor types before testing on the held-out one — and UniCure outperformed the officially released PRnet and TransiGen on LUAD, BLCA, and TNBC PTC test sets. Patient-derived training data comprised 123 LUAD and 123 BLCA PTC perturbation profiles generated with a 40-compound library, 79 TNBC PTC samples, and 396 real-world profiles from hematologic, breast, and ovarian cancers.

#Applications

The intended uses run from discovery to the clinic: virtual screening of large compound libraries against a specific tumor transcriptome, dose and combination-effect prediction, mechanism-of-action grouping, and prioritization of therapies for an individual patient. Drug rankings produced for over 1,000 patients across pan-cancer cohorts correlated with overall survival and recovered known targeted agents within their indicated cancer types. The authors also nominated natural products against triple-negative breast, lung adenocarcinoma, and bladder cancer, and confirmed concentration-dependent inhibition of viability and colony formation in vitro.

#Impact

UniCure's contribution is architectural reuse: it shows that a cellular foundation model and a chemical foundation model can be joined by a comparatively small interaction module and adapted with LoRA, instead of retraining a bespoke encoder-decoder on perturbation data alone. Pairing this with patient-derived tumor-like clusters gives a route from public cell-line atlases to individualized prediction that does not require a large clinical dataset. The caveats are real: the work is a preprint; individualized predictions require an extra fine-tuning step on patient-derived material, which is not something most labs can generate; validation of nominated compounds is in vitro; and the preprint carries a CC-BY-NC-ND license, though the code and released weights on GitHub are Apache-2.0.