Biologically-informed integration of drug representations for breast cancer treatment using deep learning

#### Project introduction Approximately 50% of breast cancer patients receiving neoadjuvant therapy do not achieve pathological complete response (pCR). Accurately predicting treatment response and selecting the optimal therapeutic strategy are unmet major challenges. To build a model with such functions, learning the interactions between tumor and treatments is necessary. Previous efforts to build response predictors have not incorporated this knowledge. Here, on the basis of data from 4,371 breast cancer patients undergoing neoadjuvant therapy across 31 datasets with transcriptomic and treatment information, we developed GDnet, an interpretable deep learning model integrating drug representations and tumor transcriptome profiles—functioning as a digital organoid—to predict the response to neoadjuvant therapy in breast cancer patients and aid in the selection of optimal treatment strategies. We demonstrate that GDnet (area under the curve (AUC) = 0.725) significantly outperforms transcriptome-only model (AUC = 0.683) in predicting treatment response. Then we conducted two series of in-silico simulated clinical trials based on I-SPY2 trial and external validation datasets respectively and showed that GDnet can optimize treatment selection and significantly increase the pCR rate in all trials. Moreover, the odds ratios (OR) for pCR increase from 1.6 to 2.5 linearly as optimization intensifies. Overall, biologically informed deep learning integrating drug representations and biological omics profiles can function as a digital organoid to optimize breast cancer neoadjuvant treatment selection and may have broader applications across various treatment settings and cancer types. #### Code description - "main.py" file: - This file mainly contain the codes to train GDnet in "3_model_construction_and_training.ipynb". Note: training of Gnetens and GNNetens are not performed give their high demand of cumputing resource. - Because the environment seems do not contain R packages, codes written in R are not included in the main script. - Other "*.ipynb" files: - Codes are written in jupyter notebooks with R and python kernel separately. - The notebook files are ordered by the prefixal number (start from 1, end with 9). #### Data description - "Cleaned_data": the patient transcriptome and clinical information - "GDSC_and_CTRP_data": GDSC and CTRP transcriptome and drug sensitivity data - "Reactome": original data downloaded from Reactome website - "Masks_with_prior_knowledge": generated masks containing the prior knowledge in Reactome - "Model_performance": the model performance saved for "4_model_performance_comparason.ipynb" - Trained_models: the trained models and the model predictions saved for "4_model_performance_comparason.ipynb" and "7_trial_group_aggignment_and_pCR_calculation_and_regimen_interpretation.ipynb" #### To train the model - run "main.py" script - trained models and model performance will be saved in "results" #### If you have any other questions, please feel free to let us know.