A new study published in Applied Intelligence by Jiaming Liu and Qi Tan of South China Normal University and Pei Yang of South China University of Technology tackles exactly this challenge. The team proposes a framework called Condition-Controlled Optimal Transport, or CCOT, which models the transformation between unperturbed and perturbed cellular distributions as a mathematical transport problem. Rather than trying to pair individual cells across conditions, which is impossible with the available data, the method learns a mapping between entire probability distributions of gene expression states. The work, published as volume 56, article 470 of the journal, was supported by the National Natural Science Foundation of China and the Guangdong Basic and Applied Basic Research Foundation.

Optimal transport is a branch of mathematics with roots stretching back to Gaspard Monge in the eighteenth century, concerned with finding the most efficient way to move mass from one distribution to another. In recent years it has found fertile ground in machine learning, and in single-cell biology specifically, where researchers have used it to align datasets across conditions, modalities, and time points. The intuition is appealing: if a drug pushes a population of cells from one transcriptional state distribution to another, the most plausible transformation between those distributions can be framed as an optimal transport map. Learning that map from data then allows one to generate predictions of how a new population of cells will respond.

The central innovation of CCOT lies in its condition-controlled mechanism, which the authors describe as bridging conditional and unconditional models within the optimal transport framework. The idea echoes classifier-free guidance, a technique popularized in the image generation community, where a model is trained to operate both with and without conditioning information and the two modes are interpolated at inference time. In CCOT, the conditioning signal is the drug perturbation itself. By learning transport maps that can function with or without the drug condition, the framework gains a knob that controls how strongly the perturbation shapes the predicted cellular response, allowing the model to balance faithful in-distribution prediction against robust generalization.

Under the hood, CCOT leverages partially input convex neural networks to parameterize the transport maps. Input convex neural networks, introduced in the machine learning literature in 2017, are architectures whose outputs are convex functions of their inputs, a property that makes them natural candidates for computing optimal transport maps via the dual formulation of the transport problem. Prior work, including the supervised training of conditional Monge maps and neural optimal transport approaches for single-cell perturbation responses, established this connection. CCOT extends it by making the condition control explicit and by aligning the learned transport with biological priors, one of three key innovations the authors list alongside the condition-controlled mechanism and cross-drug generalization.

Cross-drug generalization is arguably the property with the greatest practical stakes. Drug discovery pipelines routinely need to prioritize candidate compounds, and experiments are expensive; a model that can only describe drugs it has already seen is of limited use. CCOT is designed to generalize to novel therapeutic compounds, predicting responses out-of-distribution rather than merely interpolating within the training data. This is a notoriously difficult setting, because a new molecule may act through mechanisms only partially represented in the training set, and models can easily fail by defaulting to memorized responses or by producing biologically implausible expression profiles.

The evaluation was carried out on SciPlex3, a benchmark dataset drawn from the massively multiplex chemical transcriptomics work published in Science in 2020, which profiles gene expression at single-cell resolution across many cell types, drugs, and doses. The authors report that CCOT significantly outperforms state-of-the-art methods in capturing complex and nonlinear cellular responses, with its advantage most evident in capturing higher-order statistical properties and gene-specific responses for differentially expressed genes. Because no ground-truth paired control and perturbed cells exist, the evaluation relies on distributional metrics, including maximum mean discrepancy, energy distance, perturbation signatures, the coefficient of determination, and the Fréchet Inception Distance, each probing a different aspect of how closely the predicted distribution matches the observed one.

The appendix results add useful nuance. Across the full 977-gene space, not just the top 50 differentially expressed genes, CCOT retains its advantage on the distribution-level metrics, improving out-of-distribution maximum mean discrepancy by 66 percent over the next best model, while a competing method, ChemCPA, holds a marginal and statistically insignificant lead on some moment-based metrics. Ablation studies underscore the architecture’s dependencies: removing the drug encoder proves catastrophic, driving the coefficient of determination to a dismal negative value, while removing the condition-control mechanism or the cell type alignment degrades every metric. The authors also document a sensitivity analysis of the inner iteration count, finding stable performance for values of five or greater but a performance collapse at fifteen due to the known instability of adversarial min-max optimization, and they settled on ten inner iterations to match baseline configurations exactly.

Training dynamics receive careful treatment as well. Because the dual potential losses in the min-max optimization are adversarial and oscillate rather than decrease monotonically, the team monitors maximum mean discrepancy on the in-distribution test set as the true indicator of distributional alignment, evaluating every fifty batches and selecting the checkpoint with the minimum value as an implicit early stopping mechanism. Over a fixed budget of 100,000 outer iterations, matching the defaults of the CELLOT and CondOT baselines for a strictly controlled comparison, the losses converge from large negative values toward zero with narrowing oscillations, and the best checkpoint was selected at iteration 98,400 within a stable plateau rather than during an unconverged descent.

The broader significance of this work sits at the intersection of generative machine learning and pharmacology. If models like CCOT can reliably predict how cells respond to compounds they have never seen, the implications range from cheaper pre-screening of drug candidates to mechanistic insight into how perturbations propagate through cellular signaling networks. The authors position the framework as a tool for advancing drug discovery and understanding cellular mechanisms, and they have made the complete source code, including training scripts, preprocessing steps, and evaluation code, publicly available on GitHub, with the datasets themselves drawn from publicly accessible repositories. As with any computational prediction, validation against new laboratory experiments remains the ultimate test, but the study offers a principled mathematical route through one of single-cell biology’s most persistent obstacles: predicting the future state of a cell that must be destroyed in order to be measured.

Subject of Research: Machine learning-based prediction of single-cell drug perturbation responses using optimal transport

Article Title: Condition-Controlled optimal transport for cellular perturbation response prediction

Article References: Liu, J., Tan, Q., & Yang, P. (2026). Condition-Controlled optimal transport for cellular perturbation response prediction. Applied Intelligence, 56(15), Article 470. https://doi.org/10.1007/s10489-026-07521-6

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07521-6

Keywords: single-cell sequencing, optimal transport, perturbation response prediction, drug discovery, input convex neural networks, classifier-free guidance, SciPlex3, gene expression, machine learning, computational biology, out-of-distribution generalization, ChemCPA

The study, led by Mohammed Al-Ani of Hamad Bin Khalifa University in Qatar together with Siddhi P. Jani, Halima Bensmail, and Raghvendra Mall, builds on one of the most valuable datasets in leukemia research: the BeatAML cohort. That resource contains specimens from 805 patients with acute myeloid leukemia, each profiled for drug sensitivity across 165 different compounds in laboratory assays, alongside rich clinical annotations and both DNA and RNA sequencing data. Rather than treating this trove as a static catalog, the researchers turned it into training material for a model that learns how the molecular identity of a drug interacts with the cellular traits of a patient’s cancer.

What makes PREDICT-AML technically distinctive is the way it represents both sides of the drug-patient equation. On the drug side, the framework employs four complementary representations: classical physicochemical descriptors that capture properties such as molecular weight and lipophilicity; a novel set of knowledge-distilled embeddings, abbreviated KD-Embed, which compress what the model has learned about drug behavior into compact numerical vectors; and embeddings generated by two chemical language models, ChemBERTa and MolFormer, which are transformer architectures trained to read molecular structures the way language models read text. These representations allow the model to reason about drugs it has effectively seen in training data even when their chemical details differ substantially.

On the patient side, the model integrates several novel features that go beyond a simple list of mutations. The researchers engineered patient-specific oncogenic pathway proximity scores, which quantify how close a patient’s tumor expression profile sits to known cancer-driving signaling pathways. They also computed cell-state enrichment scores, which describe whether the leukemia cells resemble particular normal cell types, such as monocyte-like states, and they incorporated the patient’s genomic profile of somatic mutations. This combination of features lets the model capture the biological context in which a drug must act, rather than treating each patient as a bag of independent genetic alterations.

For the predictive engine itself, the team systematically searched for the best architecture using Optuna, a hyperparameter optimization framework, and identified TabPFN, a tabular foundation model, as the optimal choice. TabPFN represents a relatively new class of models designed specifically for the kind of structured, tabular data that dominates biomedical research, and its selection here underscores how foundation-model approaches are migrating from images and text into clinical prediction. When paired with the knowledge-distilled drug embeddings, TabPFN achieved a Pearson correlation of 0.690 between predicted and observed drug responses on a held-out test set, with a mean absolute error of 37.712.

One of the study’s most important methodological findings concerns how such models should be evaluated. The researchers found that random cross-validation, a common practice in machine learning studies, overestimates performance by approximately 13 percent relative to patient-stratified cross-validation, in which the model is tested on patients it has never encountered during training. Because the clinical goal is to predict responses for new patients rather than to memorize existing ones, the authors argue that patient-stratified evaluation is the appropriate benchmark. The distinction matters enormously for translational research: models that look impressive under random splits may fail when deployed on the next patient who walks through the clinic door.

The team did not stop at internal validation. They tested PREDICT-AML on two completely independent external datasets, LeeAML and FIMM-AML, generated at other institutions. The model achieved a Pearson correlation of 0.597 on LeeAML and an absolute correlation of 0.557 on FIMM-AML, confirming that its predictions generalize across patient populations, assay platforms, and clinical settings. In head-to-head comparisons, PREDICT-AML clearly outperformed established baselines: ElasticNet, a regularized regression method, managed only a correlation of 0.360, while MDREAM reached a Spearman correlation of 0.680. These results suggest that the combination of rich drug representations, multi-omics patient features, and the TabPFN architecture captures signal that simpler approaches miss.

Ablation experiments, in which the researchers systematically removed categories of input features, yielded biological insights of their own. Gene expression emerged as the single dominant predictor of drug response, reflecting the fact that a cell’s transcriptional state largely determines its vulnerability to any given compound. Perhaps more striking, the combination of somatic mutations with cell-state enrichment alone matched the performance of the full multi-omics feature set, hinting that a leaner panel of measurements could someday deliver comparable predictive power at lower cost. The knowledge-distilled embeddings proved to be the primary drug-identity signal, confirming the value of learning compact drug representations from large sensitivity datasets.

Interpretability analysis using SHAP values, a technique that quantifies each feature’s contribution to individual predictions, revealed recurring biological determinants of drug sensitivity across multiple compounds. Monocyte-like cell-state enrichment, expression of the BCL6 gene, and activity of the immunogenic cell death pathway all appeared repeatedly as markers of sensitivity. These findings are more than statistical curiosities; they point toward biological mechanisms that could be exploited therapeutically and suggest that the model is learning genuine biology rather than spurious correlations. For clinicians, such interpretability is essential before any predictive tool can inform treatment decisions.

The authors position PREDICT-AML as an interpretable, externally validated tool for individualized therapeutic prioritization in acute myeloid leukemia, a disease with limited therapeutic options and pressing clinical need. While the model’s predictions would still require prospective clinical validation before guiding real-world prescribing, the framework demonstrates that integrating chemical language models, tabular foundation models, and carefully engineered biological features can produce drug response predictions that hold up across institutions. As datasets like BeatAML continue to grow and similar resources accumulate for other cancers, the approach pioneered here may mark a step toward a future where a leukemia patient’s treatment is selected by an algorithm that has already seen, in molecular detail, how thousands of tumors responded to hundreds of drugs.

Subject of Research: Machine learning prediction of personalized drug response in acute myeloid leukemia using multi-omics data

Article Title: PREDICT-AML: personalized response via drug interaction with cellular traits for acute myeloid leukemia

Article References: Al-Ani, M., Jani, S. P., Bensmail, H., & Mall, R. (2026). PREDICT-AML: personalized response via drug interaction with cellular traits for acute myeloid leukemia. Journal of Translational Medicine. https://doi.org/10.1186/s12967-026-08817-4

Image Credits: AI Generated

DOI: 10.1186/s12967-026-08817-4

Keywords: acute myeloid leukemia, PREDICT-AML, machine learning, TabPFN, drug response prediction, multi-omics, BeatAML, chemical language models, precision oncology, SHAP interpretability, external validation, transcriptomics

The core insight behind DrugSAGE is that no single tumor profile exists in a vacuum. Cell line repositories such as the Cancer Cell Line Encyclopedia (CCLE) and drug sensitivity resources such as the Genomics of Drug Sensitivity in Cancer (GDSC) database contain rich measurements of how hundreds of cancer cell lines respond to hundreds of compounds, paired with transcriptomic profiles of those lines. The challenge has always been transferring that knowledge to patient tumors, whose expression patterns differ systematically from cell lines grown in culture. DrugSAGE addresses the transfer problem by constructing a graph in which each sample, whether a cell line or a patient tumor, is connected to its nearest neighbors in transcriptomic space. A graph neural network then learns to aggregate features across these connections, so that a patient tumor’s predicted response to a drug is informed not only by its own gene expression signature but also by the responses of the cell lines that most closely resemble it.

Technically, DrugSAGE builds on the GraphSAGE architecture, a framework originally designed to generate embeddings for nodes in large graphs by sampling and aggregating features from local neighborhoods. In the drug response setting, the nodes are biological samples and the node features are genome-wide gene expression values, typically derived from RNA-sequencing data. The aggregation mechanism allows the model to smooth and enrich each sample’s representation with information from similar samples, which is particularly valuable when the target sample’s own data are noisy or incomplete. This is precisely the situation in clinical imputation tasks: a patient tumor has a measured transcriptome but no measured drug response, and the model must infer that response from the patterns learned across the cell line graph.

One of the most distinctive features of the method is its commitment to biological interpretability, an attribute often sacrificed in deep learning approaches. Rather than relying solely on opaque layers of artificial neurons, DrugSAGE incorporates a customized linear layer that encodes gene-pathway annotations directly into the model’s architecture. Genes are grouped according to known pathway memberships, and the model’s first stage of processing respects this biological organization. As a result, the learned weights can be traced back to specific pathways, giving researchers a window into why the model made a particular prediction. The authors complement this architectural interpretability with SHAP analysis, a game-theoretic technique for attributing a model’s output to its input features, which allows them to identify the key genes and pathways driving each drug response prediction.

The benchmarking strategy was deliberately broad. The authors evaluated DrugSAGE across independent bulk RNA-seq datasets and single-cell datasets, testing whether the model’s predictions were associated with known drug targets and with patient groups stratified by treatment outcome. In the bulk setting, the model was trained on cell line data from CCLE and GDSC and then applied to patient cohorts from The Cancer Genome Atlas (TCGA) and datasets deposited in the Gene Expression Omnibus (GEO), spanning cancer types that include breast cancer, lung adenocarcinoma, colon adenocarcinoma, head and neck squamous cell carcinoma, skin cutaneous melanoma, stomach adenocarcinoma, and thyroid carcinoma. The reported results showed significant associations between predicted responses and established pharmacology, suggesting that the model captures biologically meaningful signal rather than dataset-specific artifacts.

Perhaps the most forward-looking aspect of the study is the extension to single-cell drug response prediction. Bulk transcriptomics averages gene expression across millions of cells in a tumor, masking the subclonal diversity that increasingly appears to govern therapeutic resistance. Single-cell RNA sequencing resolves that diversity cell by cell, but drug response measurements at single-cell resolution are scarce and technically demanding to generate. DrugSAGE’s graph-based design makes it naturally suited to this sparse-data regime, because each single cell’s predicted response can be aggregated from the responses of transcriptionally similar cells and samples. The authors report that DrugSAGE effectively predicts single-cell drug responses, opening the possibility of dissecting how different subpopulations within the same tumor might respond differently to the same therapy.

The clinical validation extended to treatment-stratified patient groups, where the model’s predicted drug responses were compared against actual pathological outcomes. In breast cancer cohorts, for example, the framework was applied to distinguish patients who achieved pathological complete response from those with residual invasive disease after neoadjuvant treatment. The ability of predicted response scores to separate these groups provides evidence that the computational imputation reflects clinically relevant biology, not merely statistical correlation. Similar analyses across other cancer types reinforced the pattern: samples predicted to be sensitive to a drug were enriched among patients known to benefit from it, while predicted resistance aligned with poor outcomes such as shortened progression-free survival.

The interpretability layer yielded biologically coherent findings as well. When the authors interrogated which genes and pathways drove predictions for specific drugs, the model recovered known mechanisms of action and resistance. For targeted agents, key driver genes such as EGFR in lung cancers and VEGFR-related signaling in angiogenesis-directed therapies emerged as influential features, consistent with the established pharmacology. Pathway-level analysis highlighted processes such as epithelial-mesenchymal transition, a well-characterized program associated with invasion and drug resistance, as contributors to predicted response patterns. This convergence between model attribution and prior biological knowledge is a critical benchmark for any machine learning tool intended for clinical translation, because it suggests the model has learned mechanisms rather than memorizing confounds.

The broader context makes the contribution timely. Drug response prediction has been pursued with a range of computational strategies, including elastic net regression, deep neural networks, and variational autoencoder-based approaches such as VAEN and multi-omic synthetic augmentation methods. Each approach trades off accuracy, interpretability, and data requirements in different ways. DrugSAGE’s benchmarking showed superior or comparable performance relative to existing methods, but its distinguishing advantages are the graph-based aggregation of neighbor information and the pathway-informed architecture that makes predictions traceable. In a field where black-box models have struggled to earn clinical trust, a framework that can explain itself in the language of genes and pathways may prove more actionable for oncologists and translational researchers alike.

Limitations and next steps remain, as with any computational method trained primarily on cell line data. Cell lines capture only a fraction of tumor biology, and the fidelity of any imputation ultimately depends on how well the nearest-neighbor structure of the graph bridges the gap between culture dish and clinic. Nevertheless, the open-access publication of DrugSAGE, with its reliance on publicly available resources from GDSC, TCGA, and GEO, means that the broader research community can immediately test, extend, and refine the approach. As single-cell drug screening technologies mature and more treatment-stratified patient cohorts become available, graph-based transcriptome aggregation of the kind embodied in DrugSAGE could become a standard layer in the pipeline that translates tumor gene expression into concrete treatment decisions, moving precision oncology one step closer to predicting, rather than observing, each tumor’s response to therapy.

Subject of Research: Graph neural network-based prediction of cancer drug response from transcriptomic data

Article Title: DrugSAGE: a transcriptome aggregation approach using cell lines for drug response imputation

Article References: Jia, P., & Zhao, Z. (2026). DrugSAGE: a transcriptome aggregation approach using cell lines for drug response imputation. Genome Medicine. https://doi.org/10.1186/s13073-026-01781-0

Image Credits: AI Generated

DOI: 10.1186/s13073-026-01781-0

Keywords: DrugSAGE, graph neural network, drug response prediction, transcriptomics, precision oncology, cancer cell lines, GDSC, CCLE, single-cell RNA-seq, gene-pathway interpretability, pharmacogenomics, machine learning