Personalized medicine aims to match the right drug to the right patient based on measurable biological characteristics. Currently, most approaches rely on genetic mutations to predict which drugs will work - for example, testing whether a tumor carries a specific mutation before prescribing a targeted therapy.
However, genetic mutations alone are often unreliable predictors of drug response. For instance, more than 40% of AML patients with FLT3 mutations fail to respond to midostaurin, a drug approved specifically for FLT3-mutant leukemia. Similarly, about 65% of breast cancer patients with PIK3CA mutations did not respond to the PI3K inhibitor alpelisib in a major clinical trial. This suggests that mutations capture only part of the picture.
The reason is the extraordinary complexity of cancer biology: multiple pathways compensate for each other, mutations trigger cascading downstream effects, and cancers evolve under treatment pressure. A more complete view of cancer cell biology - beyond the DNA sequence - may be needed to accurately predict drug responses.
Proteomics and phosphoproteomics - the large-scale measurement of proteins and their active phosphorylated forms - provide a functional snapshot of what is actually happening inside a cell. These data types capture the outputs of genetic changes, not just the changes themselves, and may therefore be more predictive of how a cell will respond to drugs.
This study introduces DRUML (Drug Ranking Using Machine Learning), a computational platform that uses protein and phosphoprotein data from cancer cells to rank more than 400 anti-cancer drugs in order of their predicted effectiveness for a given cancer sample. The output is a ranked list - placing the most likely effective drugs at the top for any specific patient or cell line.
DRUML was trained using in-house proteomics and phosphoproteomics data from 48 cancer cell lines - including 26 AML (acute myeloid leukemia) lines, 10 esophageal cancer lines, and 12 liver cancer lines. This training data was generated by the researchers themselves using a highly sensitive mass spectrometry technique called LC-MS/MS.
A key innovation of DRUML is that it does not need to compare a patient sample to a reference or control sample. Instead, it uses internally normalized molecular markers that can be calculated from a single patient sample in isolation - a critical feature for clinical application where reference samples may not be available.
DRUML was then verified against drug response data from 53 cell lines obtained from 12 independent laboratories worldwide, confirming that its predictions hold up outside the environment in which it was trained.
The first step in DRUML is identifying Empirical Markers of Drug Responses (EMDRs). For each drug, the researchers identify proteins and phosphorylation sites (locations on proteins where a phosphate group is attached) whose levels are consistently higher in drug-sensitive cells versus drug-resistant cells, and vice versa. These become the markers that distinguish responders from non-responders.
Rather than feeding thousands of raw molecular measurements directly into the model - which would be noisy and hard to generalize - DRUML compresses this information into a single distance metric (D). This metric measures how strongly the sensitivity-associated markers outweigh the resistance-associated markers in a given sample. A high D value predicts sensitivity; a low value predicts resistance.
The predictive models themselves are ensembles of multiple machine learning algorithms, including random forests, support vector machines, deep learning, and neural networks. Using multiple algorithms and averaging their outputs improves accuracy and reduces the risk of overfitting - where a model learns the training data perfectly but fails on new data.
Models were trained and validated using 10-fold cross-validation: the 48 cell lines were repeatedly split into training sets (80%) and test sets (20%), ensuring the model performs well on samples it has never seen before. The best-performing algorithms produced predictions with a mean squared error below 0.1 and Spearman rank correlation above 0.7.
A critical test of any predictive model is whether it works on data collected by different laboratories with different equipment, protocols, and cell lines. DRUML was evaluated on label-free phosphoproteomics data from 8 colorectal cancer cell lines generated by an independent laboratory at a different institution. Without any retraining, DRUML predicted drug responses in these cells with a Spearman correlation of 0.70 - a strong result.
DRUML was then applied to a more ambitious dataset: proteomics data from 47 diverse cancer cell lines compiled from 11 separate published studies. Despite the heterogeneity in data sources and cancer types, DRUML achieved Spearman correlation values above 0.64 for all cell lines, with prediction errors below 0.1 AAC (area above curve) units for every cell line.
In practical terms, this means DRUML correctly ranked which drugs would work best within a specific cell type with high consistency. For the top 20 ranked drugs per cell line, 76% of predictions showed ranking differences of fewer than 20 positions compared to the actual measured responses - meaning the drugs DRUML predicted as most effective were genuinely close to the top of the actual efficacy list.
The model also revealed biologically meaningful patterns in the EMDRs: drugs with similar mechanisms of action (like different PI3K inhibitors) clustered together based on their molecular markers, confirming that the model is capturing real biology rather than random correlations in the training data.
Beyond cell line models, the researchers tested whether DRUML predictions have real clinical relevance in actual leukemia patients. They applied DRUML to phosphoproteomics data from 36 primary AML patient samples - bone marrow cells taken directly from patients before treatment.
All 36 patients received induction chemotherapy containing cytarabine, a standard AML drug. DRUML was used to predict how sensitive each patient's leukemia cells would be to cytarabine, based solely on the phosphoprotein profile of their cells.
The key finding: patients predicted by DRUML to be sensitive to cytarabine showed significantly longer overall survival than those predicted to be resistant (log-rank p value below 0.005). This is a strong result - it means the molecular signature captured by DRUML's cytarabine model genuinely tracked with whether patients benefited from the drug in a real clinical setting.
Furthermore, DRUML also predicted that patients resistant to cytarabine would respond better to alternative drugs. This kind of stratification - identifying which patients need different treatments - is exactly what precision medicine aims to achieve and represents the pathway from a laboratory prediction tool to a clinically actionable diagnostic.
The molecular data powering DRUML was generated by liquid chromatography coupled to tandem mass spectrometry (LC-MS/MS), a powerful analytical technique that can identify and quantify thousands of proteins and phosphorylation sites in a single experiment. The researchers performed 288 LC-MS/MS runs to profile 48 cell lines in triplicate.
This analysis identified 22,804 phosphopeptides (protein fragments carrying phosphate groups that indicate active signaling) and 6,455 proteins, generating nearly 4 million quantitative data points. The scale of this dataset is what enables the machine learning models to learn reliable markers of drug sensitivity across many different drugs and cancer contexts.
The researchers used label-free quantification - measuring protein abundance directly from the mass spectrometry signal rather than using chemical labels to compare samples. This approach is important for clinical applicability because it does not require comparing a patient sample to a reference, making it feasible to apply in real clinical settings where matched controls may not exist.
Drug response data (how much each drug suppresses cancer cell growth) was obtained from PharmacoDB, a public database of cancer cell line drug sensitivity measurements. By pairing this with the proteomics data, the researchers had the paired input-output data needed to train the machine learning models.
One of the central arguments of this paper is that protein data may be more useful than DNA data for predicting drug responses. Proteins are the actual machines of the cell - the direct targets of most drugs - while DNA only encodes the instructions. A mutation in DNA does not always change protein activity; conversely, some proteins become abnormally active without any detectable mutation.
The study suggests that phosphoproteomic data (measuring which proteins are activated by phosphorylation, a molecular on/off switch) is particularly informative because it captures the real-time signaling state of a cell - which pathways are active, which drugs their targets are sensitive to, and how the cell is likely to respond.
DRUML represents a step toward a new kind of companion diagnostic: instead of testing a single biomarker for a single drug, it simultaneously ranks hundreds of drugs using the full molecular fingerprint of a tumor. This could help oncologists identify the best available treatment option, rather than prescribing based on a single biomarker that may not capture the full complexity of the disease.
As mass spectrometry technology continues to improve in speed, cost, and sensitivity, proteomics-based tools like DRUML are becoming increasingly viable for clinical implementation - particularly for blood cancers like AML where tumor material is accessible from blood and bone marrow samples.
DRUML demonstrates that machine learning applied to proteomic and phosphoproteomic data can produce clinically meaningful predictions of anti-cancer drug efficacy across a wide range of cancer types and drug classes. The model achieves low prediction errors, high ranking accuracy, and generalizes well across independent datasets from different laboratories.
The clinical validation in AML patients is particularly significant: survival differences based on predicted cytarabine sensitivity confirm that DRUML is capturing biologically real and clinically relevant information, not just statistical noise from laboratory experiments.
The research team has made DRUML available as an open software package (DRUMLR), enabling other researchers and potentially clinicians to apply the same approach to new cancer types and patient datasets. Future work will focus on expanding the training datasets, incorporating additional molecular data types, and validating predictions in prospective clinical trials.
The broader vision is a future where a cancer patient's tumor cells are profiled by mass spectrometry, and the resulting data is fed into DRUML to generate a ranked list of drugs - effectively providing an individualized treatment priority list to guide oncologists in selecting the most likely effective therapies for that specific patient's cancer.