PharmaFormer predicts clinical drug responses through transfer learning guided by patient derived organoid

Abstract

A major challenge in effective cancer treatment is the variability of drug responses among patients. Patient-derived organoids greatly preserve the genetic and histological characteristics even the drug sensitivities of primary tumor tissues, therefore provide a compelling approach to predict clinical outcome. However, the individual organoid culture and following drug response test are time and cost-consuming, which hinders the potential clinical application. Here, we developed PharmaFormer, a clinical drug response prediction model based on custom Transformer architecture and transfer learning. PharmaFormer was initially pre-trained with the abundant gene expression and drug sensitivity data of 2D cell lines, and was then finalized through a model further fine-tuned with the limited organoid pharmacogenomic data accumulated at the present stage. Our results demonstrate that PharmaFormer, integrating both pan-cancer cell lines and organoids of a specific type of tumor, provides a dramatically improved accurate prediction of clinical drug response. This study highlights that advanced AI models combined with biomimetic organoid models will accelerate precision medicine and future drug development.

Fig. 1. Schematic of the PharmaFormer model.

PharmaFormer leverages large-scale parallel drug testing on cell lines combined with gene expression data to build a pre-trained model that is further fine-tuned with organoid-specific data for clinical drug response predictions. A Workflow of PharmaFormer. B The architecture of PharmaFormer includes a feature extractor that integrates cellular and drug structural data into a transformer encoder tailored for drug response prediction. C Detailed composition of each algorithmic module, highlighting its role in prediction accuracy.

Fig. 2. Comparative analysis of PharmaFormer pre-trained model versus classical machine learning models on cell line dataset.

A Five-fold validation of average per-drug Pearson correlation coefficients comparing PharmaFormer to classical machine learning models. B Performance comparison of PharmaFormer and classical machine learning models across various tissue types, represented by average per-drug Pearson correlation coefficients. C Average per-drug Pearson correlation coefficients of PharmaFormer versus those of classical machine learning models across each TCGA tumor category. D Pearson correlation coefficients of PharmaFormer and classical machine learning models across FDA-approved drugs.

Fig. 3. Evaluation of PharmaFormer model in predicting drug response in colon cancer cohort and bladder cancer cohort.

A Kaplan–Meier survival curves for colon cancer patients receiving 5-fluorouracil and oxaliplatin treatment, comparing predictions from both the pre-trained and fine-tuned PharmaFormer models. B Kaplan–Meier survival curves for bladder cancer patients treated with gemcitabine and cisplatin, comparing predictions from PharmaFormer’s pre-trained and fine-tuned models. C Forest plot illustrating the effect of various training strategies on PharmaFormer’s predictive performance in colon cancer, including training with cell lines only, colon cancer organoids only, a combined dataset of cell lines and organoids, and a sequential training approach of cell lines followed by organoid fine-tuning. D Forest plot comparing PharmaFormer’s performance in bladder cancer under different training conditions: cell lines only, bladder cancer organoids only, combined cell lines and organoids, and sequential pre-training with cell lines followed by organoid fine-tuning. Statistical significance, hazard ratios, and confidence intervals were derived using the Cox proportional hazards model (CoxPHFitter) from the lifelines package.

Fig. 4. Performance evaluation of PharmaFormer with different fine-tuned dataset and canonical machine learning methods.

A, B Forest plots assessing the impact of the transfer learning approach on clinical drug response predictions, based on cell line and organoid data from matched tumor types (colon and bladder cancer). C, D Comparative clinical prediction performance of PharmaFormer against benchmark models for colon and bladder cancer. The analysis benchmarks PharmaFormer against classical machine learning algorithms, a biological network-based method, and a deep neural network (DNN). Forest plots display hazard ratios (HRs) and 95% confidence intervals, comparing models trained on cell lines only (blue) versus models pre-trained on cell lines and fine-tuned with organoid data (red). All metrics were computed using the Cox proportional hazards model.