The research, led by Xulin Pan from The Pennsylvania State University, introduces a leakage-controlled benchmark specifically designed to test how well synergy classification models perform when they are forced to generalize to entirely new entities. The study utilizes the DrugComb v1.5 dataset, a comprehensive repository containing 739,964 experiments across 26 studies, 17 tissues, 288 cell lines, and 4,268 compounds. By strictly excluding any measured dose-response readouts and relying solely on pre-treatment identity and context descriptors such as compound identities, cell line types, tissue origins, and clinical development phases, the researchers created a controlled environment to isolate the predictive power of metadata alone.

The core of the investigation focuses on the Zero Interaction Potency (ZIP) score, a standard metric used to quantify drug synergy. The researchers derived a three-class endpoint from this score using conventional thresholds of plus or minus 10, categorizing interactions as synergistic, additive, or antagonistic. Two primary models were evaluated under this framework: a logistic regression baseline and a class-weighted gradient-boosting model using the LightGBM algorithm. The evaluation was conducted under both predefined grouped splits and specific cold-start splits, which simulate scenarios where the model encounters drugs, cell lines, or entire studies it has never seen during training.

On the standard test set, the LightGBM model demonstrated strong performance, achieving an accuracy of 0.782 and a balanced accuracy of 0.805. It significantly outperformed the logistic regression baseline, a result confirmed by paired McNemar testing with a p-value less than 0.001. The model also achieved a macro-F1 score of 0.703 and a macro one-vs-rest area under the curve of 0.922. Precision-recall analysis indicated that the model provided roughly six-fold enrichment for the rare antagonistic and synergistic classes over their base rates, suggesting that it could effectively triage potential combinations for further experimental screening.

However, the study’s most critical findings emerge when the models are subjected to distribution shifts that mimic real-world deployment challenges. When the model was tested on unseen drugs, its balanced accuracy dropped to 0.694. The decline was more severe for unseen cell lines, where balanced accuracy fell to 0.482. Most strikingly, when the model was evaluated on entirely unseen studies, its performance collapsed to the random floor, indicating that it could not generalize beyond the specific experimental contexts it had learned. An ablation study confirmed that this failure was not driven by the study identifier itself, but rather by the lack of generalizable features across different experimental setups.

Permutation importance analysis identified cell-line identity as the dominant predictor in the model’s decision-making process. This finding led the researchers to test a simple baseline: a lookup table that retrieved base rates for each specific cell line and drug pair. Surprisingly, this non-parametric approach recovered a balanced accuracy of 0.778, which was within just 0.027 of the full LightGBM model’s performance. This proximity suggests that the majority of the model’s apparent success on standard splits is attributable to conditional memorization of entity-specific base rates rather than the learning of complex pharmacological relationships.

The implications of these results are profound for the field of computational pharmacology. The study argues that standard random or grouped splits can materially overstate the performance of synergy prediction models when they are deployed in cold-start scenarios. By providing a leakage-controlled lower bound, the benchmark offers a rigorous standard against which richer molecular and omics-based models can be measured. The authors emphasize that future evaluations should routinely include cold-start splits alongside conventional metrics to provide a more accurate picture of a model’s true generalization capabilities.

This work highlights a broader issue in machine learning for biology: the gap between in-distribution performance and out-of-distribution robustness. While high accuracy scores on standard benchmarks are often cited as evidence of a model’s utility, they may mask fundamental limitations in its ability to predict outcomes for novel combinations. As the pharmaceutical industry increasingly turns to AI to reduce the cost and time of drug discovery, the need for rigorous, leakage-controlled benchmarks becomes essential to avoid investing in models that cannot reliably predict the behavior of new drug pairs.

The study does not dismiss the value of machine learning in drug synergy prediction but rather refines the expectations for what these models can achieve. It demonstrates that useful synergy triage is achievable for known entities without the need for complex molecular features, but it warns that this capability is heavily dependent on the specific context of the training data. For new drugs or cell lines, the predictive power diminishes sharply, underscoring the need for models that can integrate deeper mechanistic insights or molecular descriptors that are invariant across different experimental conditions.

By establishing this benchmark, the researchers provide a critical tool for the scientific community to assess the true generalization potential of synergy prediction algorithms. The findings call for a shift in how models are evaluated, moving beyond simple accuracy metrics to include rigorous cold-start tests that reflect the realities of drug development. This approach will help ensure that computational tools are not only accurate in retrospective analyses but also reliable in prospective applications, ultimately contributing to more efficient and effective drug discovery processes.

Subject of Research: Evaluation of machine learning generalization in drug-combination synergy prediction

Article Title: A leakage-controlled cold-start benchmark for drug-combination synergy classification

Article References: Pan, X. (2026). A leakage-controlled cold-start benchmark for drug-combination synergy classification. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06665-z

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06665-z

Keywords: drug synergy, machine learning, benchmark, cold-start, pharmacology, LightGBM, data leakage, drug discovery, computational biology, ZIP score, leakage-controlled, drug-combination

The drug targets the phosphatidylinositol-4,5-bisphosphate 3-kinase, AKT, mechanistic target of rapamycin signaling cascade, a central growth and survival pathway that is frequently dysregulated in both canine and human cancers. When this pathway runs amok, it drives neoplastic cells to proliferate relentlessly, evade apoptosis, crank up metabolism, and, in the case of lymphoid malignancies, stimulate the production of vascular endothelial growth factor, a molecule that supports tumor neovascularization. Because single-node inhibitors of the pathway often deliver only brief or incomplete responses, likely due to well-known feedback loops within the signaling network, the researchers turned to VDC597, a compound that blocks PI3K along with both mTOR complexes, mTORC1 and mTORC2, thereby hitting the pathway at multiple points simultaneously.

In their experiments, the team tested VDC597 against two canine cell lines: 1771, derived from a spontaneously occurring B-cell lymphoma, and CLL1390, derived from chronic lymphocytic leukemia. Both lines had previously been documented to carry mutations in or constitutively express activated proteins of the PI3K-AKT-mTOR pathway, making them appropriate models for pathway-targeted therapy. Western blot analysis showed that a 24-hour treatment with the drug produced a clear, dose-dependent reduction in phosphorylated AKT at the serine 473 residue, the molecular signature of pathway activation. The leukemia line proved more sensitive, showing significant suppression at concentrations as low as 0.2 micromolar, while the lymphoma line required a full micromolar dose to achieve statistically significant inhibition.

Timing experiments revealed that the drug’s effects on signal transduction are both rapid and durable. Following a single 1 micromolar dose, phosphorylation of AKT dropped quickly and remained at roughly half of control levels for a full 24 hours. Downstream phosphorylation of 4EBP1, a translation-regulating target of mTORC1, fell even more sharply, with statistically significant suppression measurable from half an hour after treatment through the two-hour mark. The researchers verified these biochemical findings by immunohistochemistry on fixed cell pellets, observing reduced perinuclear and cytoplasmic staining for both phosphoproteins after drug exposure, an encouraging confirmation that the compound shuts down signaling across the entire pathway rather than just a single branch.

The functional consequences were equally striking. After 72 hours of exposure, VDC597 halved the viable cell population at concentrations within or near what is pharmacologically achievable, with mean half-maximal inhibitory concentrations of 1.41 micromolar for CLL1390 and 2.56 micromolar for 1771, values informed by an earlier phase I trial of the drug in tumor-bearing dogs. Live-cell imaging, tracking fluorescently labeled nuclei and dead-cell stains over 48 hours, showed that the drug did not merely slow proliferation: it actively killed the lymphoma cells, with cell death increasing in a dose-dependent manner and becoming significantly elevated at concentrations of 0.25 micromolar and above. Enzyme-linked immunosorbent assays further demonstrated that VDC597 suppressed vascular endothelial growth factor production in both cell lines, again in a dose-dependent fashion, suggesting the drug may undermine the angiogenic support system that lymphoid tumors rely upon.

Because combination therapy, not single agents, ultimately determines how well targeted drugs translate into clinical benefit, the team paired VDC597 with the three chemotherapy drugs most commonly used in canine lymphoma protocols: dexamethasone, doxorubicin, and vincristine. The results were revealing. Combining VDC597 with dexamethasone produced genuinely synergistic growth inhibition, quantified by combination indices below 1 using the Chou-Talalay method, with the effect strongest at lower drug concentrations. Pairing the drug with doxorubicin yielded additive to mildly synergistic effects at lower doses, though the highest doxorubicin concentration produced antagonism, possibly because PI3K inhibition dampens the reactive oxygen species generation through which doxorubicin promotes apoptosis. Vincristine, whose mechanism overlaps with the pathway’s role in microtubule function during mitosis, showed only additive to mildly antagonistic interactions.

The strong synergy with dexamethasone is mechanistically compelling. Glucocorticoids kill lymphoid cells through receptors whose nuclear translocation and downstream signaling are potentiated by PI3K-AKT-mTOR inhibition, a finding previously documented in human B-cell acute lymphoblastic leukemia. The Colorado State results thus support the hypothesis that adding a dual PI3K/mTOR inhibitor to standard chemotherapy could deepen responses or sensitize resistant tumors, particularly in lymphoma, where the clinical reservoir of effective options is thin.

Importantly, the study went beyond cell culture to examine pathway activation in actual patients. Using tissue microarrays built from 118 lymph node samples collected from 84 dogs with spontaneous B-cell lymphoma, the researchers scored phosphorylated AKT and phosphorylated 4EBP1 by immunohistochemistry, employing both traditional manual Allred scoring and artificial intelligence-assisted whole-slide analysis with Visiopharm software to generate finer-grained H-scores. The two approaches correlated well, validating the use of machine learning tools for quantitative pathology in veterinary specimens. The phosphorylation signals proved highly variable from tumor to tumor, underscoring that not every canine B-cell lymphoma is driven by an equally active PI3K-AKT-mTOR pathway.

Two clinically meaningful correlations emerged. First, tumors with lower expression of major histocompatibility complex class II, a known adverse prognostic marker measured by flow cytometry, showed significantly higher levels of phosphorylated AKT, linking aggressive immune phenotype to stronger pathway activity. Second, samples taken after relapse following CHOP chemotherapy showed higher phosphorylated 4EBP1 than matched pre-treatment samples, a result that retained significance in paired manual scoring. However, no correlation was found between either phosphoprotein and progression-free interval, overall survival, WHO stage or substage, or flow cytometric immunophenotypic markers, and phosphorylated AKT did not rise in parallel with phosphorylated 4EBP1 in the relapse samples. The authors caution against overinterpreting these limited associations and suggest several possible explanations, including AKT phosphorylation at sites not detected by the assay, alternative 4EBP1 kinases such as PIM2 or RAS/MAPK signaling, or mutations downstream of AKT.

The study has clear limitations that its authors acknowledge. Only one cell line each was tested for lymphoma and leukemia, so the full heterogeneity of canine lymphoid malignancies, including T-cell lymphoma and acute lymphoblastic leukemia, remains unexplored. No washout experiments were performed to determine how long signaling stays suppressed after the drug is withdrawn, information that would be essential for designing oral dosing schedules. And the finding of elevated phosphorylated 4EBP1 without corresponding phosphorylated AKT in relapsed tumors awaits mechanistic confirmation. Nevertheless, the core message is clear: VDC597 collapses PI3K-AKT-mTOR signaling, kills lymphoid cancer cells in vitro at achievable concentrations, synergizes with dexamethasone, and may sensitize tumors to standard chemotherapy. As dogs with spontaneous diffuse large B-cell lymphoma are increasingly recognized as valuable models for human non-Hodgkin lymphoma, these results position dual PI3K/mTOR inhibition as a strategy worth advancing toward larger cell line panels, animal studies, and ultimately clinical trials in both species.

Subject of Research: Evaluation of the dual PI3K/mTOR inhibitor VDC597 against canine B-cell lymphoma and chronic lymphocytic leukemia cells in vitro, with immunohistochemical assessment of PI3K-AKT-mTOR pathway activation in spontaneous canine B-cell lymphoma.

Article Title: In vitro effects of the PI3K/mTOR dual-inhibitor, VDC597, in canine lymphoma and leukemia

Article References: Meuten, T., Brill, S. A., Brady, R. V., Farrell, K. B., Rose, B. J., Schlein, L. J., & Thamm, D. H. (2026). In vitro effects of the PI3K/mTOR dual-inhibitor, VDC597, in canine lymphoma and leukemia. Veterinary Oncology, 3(1), Article 13. https://doi.org/10.1186/s44356-026-00065-4

Image Credits: AI Generated

DOI: 10.1186/s44356-026-00065-4

Keywords: canine lymphoma, canine leukemia, VDC597, PI3K, mTOR, AKT, 4EBP1, signal transduction, chemotherapy, immunohistochemistry, veterinary oncology, drug combination

The rationale behind the trial rests on the distinctive metabolic biology of pancreatic tumors. Pancreatic ductal adenocarcinoma is characterized by a dense desmoplastic stroma and a poorly vascularized tumor microenvironment, conditions that limit oxygen and nutrient delivery and force cancer cells to rely heavily on adaptive metabolic pathways. Among these, glutamine metabolism occupies a central position. Glutamine serves as a key nitrogen donor for nucleotide synthesis, feeds the tricarboxylic acid cycle through glutaminolysis, and supports glutathione production, which helps tumor cells buffer oxidative stress. In the hypoxic, nutrient-poor setting of a pancreatic tumor, this dependence on glutamine becomes particularly pronounced, which is precisely why investigators have long been interested in manipulating glutamine availability as a therapeutic strategy.

Paradoxically, the GlutaPanc approach involves supplementing patients with l-glutamine rather than depriving tumors of it. The underlying concept draws on pharmacological modulation of glutamine handling in ways that may potentiate chemotherapy. Gemcitabine, a nucleoside analog, competes for cellular transport and activation pathways that intersect with nucleotide metabolism, and intracellular pools influenced by glutamine-dependent de novo synthesis can affect how effectively the drug is incorporated into DNA. By altering the metabolic state of tumor cells, exogenous glutamine may shift the balance of gemcitabine activation and catabolism, potentially increasing the cytotoxic payload delivered to malignant cells while leaving normal tissues comparatively unaffected. Similar metabolic priming strategies have been explored in other malignancies, but pancreatic cancer, with its extreme metabolic stress, offers a particularly compelling testing ground.

The trial design reflected the careful staging typical of early-phase oncology studies. As an open-label, single-arm phase 1 study, GlutaPanc enrolled participants with advanced pancreatic ductal adenocarcinoma who were candidates for first-line treatment with gemcitabine and nab-paclitaxel. Rather than adding a fourth cytotoxic agent, the investigators layered oral l-glutamine supplementation onto the established doublet, escalating the dose of the amino acid to determine how much could be given safely before dose-limiting toxicities emerged. This design allowed the team to characterize the tolerability profile of the triplet in a controlled manner and to define the dose that would be carried forward into larger efficacy studies.

Safety and feasibility were the primary endpoints, and the trial’s central conclusion is that the combination was deliverable in this patient population. Establishing feasibility matters enormously in pancreatic cancer, where patients frequently present with poor performance status, weight loss, and compromised nutritional reserves. Cachexia and malnutrition are near-universal features of advanced disease, and any regimen that adds burden to an already fragile patient population risks being unusable in practice. The finding that l-glutamine could be incorporated without compromising the administration of gemcitabine and nab-paclitaxel therefore addresses a genuine unmet need, because it opens the door to metabolic interventions that do not come at the cost of treatment intensity.

The determination of a recommended phase 2 dose is the practical output that will shape the next stage of clinical development. Phase 1 trials in oncology traditionally escalate a cytotoxic agent until toxicity becomes unacceptable, but studies of metabolic supplements require a more nuanced approach, balancing pharmacological plausibility against tolerability and adherence. By formally defining the dose of l-glutamine to be used in combination with the chemotherapy doublet, the GlutaPanc investigators have created a standardized protocol that future trials can follow, reducing heterogeneity and enabling meaningful comparison of results across studies. This kind of dose-finding groundwork is unglamorous but essential; without it, subsequent efficacy trials risk being uninterpretable.

The broader significance of the trial lies in its position within a growing movement to integrate metabolic therapeutics into mainstream cancer care. For decades, the Warburg effect and its emphasis on glucose consumption dominated thinking about tumor metabolism, but the past fifteen years have seen glutamine emerge as an equally important nutrient axis. Pancreatic cancer cells in particular have been shown in preclinical models to scavenge glutamine and route it into pathways that support redox balance and biomass production. Translating those laboratory observations into clinical benefit has proven difficult, with several glutamine-targeting strategies faltering in trials. GlutaPanc represents a different tack: rather than blocking glutamine utilization with an enzyme or transporter inhibitor, it modulates the metabolic environment pharmacologically in a way that is compatible with existing chemotherapy.

Nab-paclitaxel, the albumin-bound formulation of paclitaxel used in the trial, deserves mention in its own right. When combined with gemcitabine, nab-paclitaxel improved survival in metastatic pancreatic cancer and became a standard first-line option for patients who can tolerate the regimen. The doublet works in part by depleting the tumor stroma and improving drug delivery, effects that complement gemcitabine’s DNA-damaging mechanism. Adding a metabolic modulator to this regimen is conceptually coherent, because the stroma-modulating activity of nab-paclitaxel may partially relieve the nutrient deprivation that drives glutamine dependence in the first place. Understanding how these three components interact at the level of tumor physiology will be an important question for the phase 2 program.

Cautious interpretation remains essential at this stage. Phase 1 trials are designed to answer questions of safety and dosing, not to demonstrate that a new combination prolongs survival, and the GlutaPanc results should be understood as a green light for further study rather than a treatment advance in themselves. Patients and clinicians will need to await randomized phase 2 and ultimately phase 3 data before drawing conclusions about whether l-glutamine supplementation genuinely improves outcomes when added to gemcitabine and nab-paclitaxel. Nonetheless, the trial addresses a disease with desperately limited options, where five-year survival rates remain in the single digits and where even incremental improvements in first-line therapy can translate into meaningful gains for thousands of patients worldwide.

The GlutaPanc trial also highlights the value of rigorously testing biologically motivated ideas in the clinic. Metabolic interventions are often dismissed as nutritional support rather than true therapeutics, yet the systematic dose-finding approach applied here treats l-glutamine with the same methodological seriousness applied to any investigational drug. As the recommended phase 2 dose now moves forward, the oncology community will be watching to see whether manipulating one of cancer’s favorite nutrients can genuinely bend the curve in pancreatic ductal adenocarcinoma, a disease that has stubbornly resisted nearly every therapeutic innovation thrown at it over the past half-century.

Subject of Research: A phase 1 trial evaluating l-glutamine combined with gemcitabine and nab-paclitaxel in advanced pancreatic ductal adenocarcinoma

Article Title: l-Glutamine in combination with first-line gemcitabine and nab-paclitaxel in advanced pancreatic ductal adenocarcinoma: an open-label, single-arm, phase 1 GlutaPanc trial

Article References: Gong, J., Muranaka, H., Choi, S. Y., Tighiouart, M., Bhute, S., Aja, E. R., Jacobs, J. P., Stotland, A., Van Eyk, J., Elmadbouh, O. H. M., Edderkaoui, M., Tanaka, S., Furuya, H., Osipov, A., Lorber, J., Billet, S., Morris, A., ten Hoeve-Scott, J., Graeber, T., … Bhowmick, N. A. (2026). l-Glutamine in combination with first-line gemcitabine and nab-paclitaxel in advanced pancreatic ductal adenocarcinoma: an open-label, single-arm, phase 1 GlutaPanc trial. Nature Cancer. https://doi.org/10.1038/s43018-026-01225-z

Image Credits: AI Generated

DOI: 10.1038/s43018-026-01225-z

Keywords: pancreatic cancer, l-glutamine, gemcitabine, nab-paclitaxel, phase 1 trial, GlutaPanc, glutamine metabolism, tumor metabolism, clinical trial, oncology, drug combination, pancreatic ductal adenocarcinoma