A research team has built an AI system that predicts a cancer patient's drug response using only pre-treatment tumor data.
The system, called PREDIKTOR, works by running two parallel analyses of the same patient-drug pair and then aligning them. The first builds a personalized gene regulatory network from the patient's tumor expression profile and layers in known drug-target relationships. The second uses a model pretrained on the LINCS L1000 dataset to simulate what the patient's gene expression would look like after drug exposure. A contrastive learning method — the same broad family of techniques behind image-text models like CLIP — pulls these two views into a shared space before making a classification. On The Cancer Genome Atlas benchmark, PREDIKTOR beat existing methods across patient, drug, and tissue evaluation splits. It also transferred to the I-SPY2 breast cancer trial without retraining, improving the area under the receiver operating curve by 5.6 percentage points over the next-best method.
The harder problem PREDIKTOR is attacking is the scarcity of matched clinical data: most patients have pre-treatment tumor profiles but no post-treatment molecular readout to train on. By combining a static knowledge-graph view with a dynamic perturbation simulation, the framework sidesteps needing that paired data at training time. The gene and pathway attributions it produces are also interpretable enough to recover known drug mechanisms — a requirement for clinical adoption that most black-box models skip.
Precision oncology has seen a wave of ML predictions that perform well on benchmarks and quietly disappear before a trial; zero-shot generalization to I-SPY2 is a more credible signal than TCGA accuracy alone, but prospective validation will still determine whether PREDIKTOR is a tool or a paper.