DeepCRE: Transforming Drug R&D via AI-Driven Cross-drug Response Evaluation

TL;DR

DeepCRE is introduced, a pioneering AI model designed to predict CRE effectively in the late stages of drug R&D, and demonstrates the capability of DeepCRE to systematically uncover a spectrum of drug candidates with enhanced therapeutic effects, highlighting its potential to transform drug R&D.

Abstract

The fields of therapeutic application and drug research and development (R&D) both face substantial challenges, i.e., the therapeutic domain calls for more treatment alternatives, while numerous promising pre-clinical drugs have failed in clinical trials. One of the reasons is the inadequacy of Cross-drug Response Evaluation (CRE) during the late stages of drug R&D. Although in-silico CRE models bring a promising solution, existing methodologies are restricted to early stages of drug R&D, such as target and cell-line levels, offering limited improvement to clinical success rates. Herein, we introduce DeepCRE, a pioneering AI model designed to predict CRE effectively in the late stages of drug R&D. DeepCRE outperforms the existing best models by achieving an average performance improvement of 17.7% in patient-level CRE, and a 5-fold increase in indication-level CRE, facilitating more accurate personalized treatment predictions and better pharmaceutical value assessment for indications, respectively. Furthermore, DeepCRE has identified a set of six drug candidates that show significantly greater effectiveness than a comparator set of two approved drugs in 5/8 colorectal cancer organoids. This demonstrates the capability of DeepCRE to systematically uncover a spectrum of drug candidates with enhanced therapeutic effects, highlighting its potential to transform drug R&D.

Figures (6)

• Figure 1: The objective and application of DeepCRE. a, Illustration of CRE distribution across different drug development stages, highlighting the potential benefits of in-silico CRE models in enhancing CRE availability during late-stage development. b, Comparative diagram showing the traditional vs. DeepCRE-assisted drug development processes. The DeepCRE-assisted method tends to increase success rates while reducing both time and costs. c, Conceptual framework of DeepCRE, aligning cell line and patient representations with the drug response space. This alignment enables the extrapolation of cell line drug responses to predict patient drug responses effectively. d, Heatmap displaying the Area Under the Curve (AUC) scores for eight patient-derived colorectal cancer (CRC) organoids treated with four different sets of drugs. e, Boxplot illustrating the AUC scores of the four drug sets, with Set A (identified by DeepCRE) demonstrating significantly greater efficacy than Set C in 5/8 CRC organoids.
• Figure 2: The construction of the DeepCRE models. a, Illustration of DeepCRE: the representations of cell lines and patients are aligned using a domain separation network (DSN) bousmalis2016domain architecture. b, Variants of the DeepCRE model. In the DSN setting, the private embeddings of the source domain and target domain can be kept similar via optimizing $L_{mmd}$ (DSN-mmd) or $L_{adv}$ (DSN-adv). Moreover, another variation is that the concatenation of private and shared embeddings is kept similar, which can be represented as DSRN-mmd and DSRN-adv. c, Two pretraining strategies lead to different alignments of cell lines and patients of one specific tumor type. d, Bar plots showing the percentage change in performance from all-data pretraining to tumor-type adaptive pretraining for eight DeepCRE models across 13 tumor types, with the numbers below representing the average percentage change among all tumor types. Three adv-loss-based DeepCRE models (the last column) generally improve significantly in 12 tumor types, except for GBM.
• Figure 3: DeepCRE outperforms other models in patient-level CRE across 13 tumor types. a, Heatmap comparing the AUROC performance of P-SDL, C-MDL, and tumor-type adaptive pretraining DeepCRE models. The listed models (from top to bottom) are P-SDL (dsn, dsna, Code-base, Code-mmd, Code-adv), C-MDL (DrugCell, Paccmann, three types of TGSA) and DeepCRE (AE, DSN, AE-mmd, DSRN-mmd, DSN-mmd, AE-adv, DSRN-adv, DSN-adv) models. DeepCRE models achieve SOTA (marked with *) performance in all tumor types. b,c, Heatmaps displaying alignment metrics for patients (brown) and cell lines (blue) across four pretraining methods: b, Maximum Mean Discrepancy gretton2012kernel (MMD) relative to original expression within each tumor type. The relative MMD generally decreases from AE, DSN, DSN-mmd to DSN-adv, indicating improved alignment. Tumor types with relative MMD of DSN-adv less than 0.3 are marked with * for further analysis in d. c, Kullback-Leibler (KL) divergence kullback1951information also decreases from AE, DSN, DSN-mmd to DSN-adv. The KL divergence of the original expression is not shown due to its nearly infinite value. d, t-SNE plots of gene expression encoded embeddings for four tumor types (marked with * in b) from original expression to four DeepCRE models. Notably, DSN-adv (the last row) exhibits significant improvement in the alignment of patients and cell lines compared with other models. e, Bar plots indicating the AUROC performance of P-SDL, C-MDL and the DeepCRE DSN-adv model. DSN-adv outperforms all P-SDL and C-MDL models across all tumor types. f, Line chart showing the percentage increase in performance of the DSN-adv model compared with the best P-SDL (brown) and C-MDL (blue) models across all tumor types. The maximum increase for the best P-SDL model is observed in SARC, while the maximum increase for the best C-MDL model is seen in READ. The READ-related tumor type COAD also exhibits a substantial increase. g, Scatter plot indicating the AUROC performance of C-MDL models, with different shapes representing the types of used data. Generally, mutation-data-based methods outperform expression-data-based methods for C-MDL models.
• Figure 4: DeepCRE outperforms other models in indication-level CRE. a, Workflow illustrating the validation process of drug candidates discovered by DeepCRE. Refer to the Methods section for a detailed description. DRP: drug response for patients. EDP: efficient drugs for patients. EDI: efficient drugs for indications. DEI: drug efficacy for indications. b, Heatmap displaying DRP score matrix, e.g., COAD, comprising 288 patients (rows) and 233 drug candidates (columns). c, Histogram plot of EDPs for COAD. The top 10 drug candidates with the highest efficiency in patients are highlighted in red. Three drug candidates are efficient for over 50% of COAD patients. d, Pie chart demonstrating 10 out of 17 drug candidates have records in Drugbank. e, Heatmap of the DEI table. Both predicted-efficient and Drugbank-recorded EDIs are colored blue and labeled “Proved”. Only predicted-efficient EDIs are colored purple and labeled “Promising”. Predicted-inefficient EDIs are colored grey and labeled “Inefficient”. f, Sankey plot depicting the distribution of in-clinical-test tumor types for the drug candidates. g, Heatmap of the in-clinical-test records, colored by the combination of predicted results and in-clinical-test records. Qualified drugs are labeled with a pentagram. h,i, Bar plots demonstrating the record numbers of the DSN-adv method and five C-MDL methods: h for EDIs, i for Qualified drugs.
• Figure 5: Validation of DeepCRE for efficient drug identification in the CRC21 patient. a, Workflow demonstrating the utilization of DeepCRE to identify four types of drug sets, which were subsequently evaluated using CRC organoids derived from a patient who relapsed after traditional XEOLX treatment, employing InSMAR-Chip drug testing methodology. DRC: drug response curve. b, Sankey plot illustrating four drug sets with diverse mechanisms of actions (MoAs) and targets. c, 2D XY-plot displaying the evaluation outcomes of DeepCRE as well as traditional methods. d, DRC colored by the drug set indicating the ranking of drug efficacy: A> B> C> D. e, DRC colored by the drug MoA indicating that certain drugs with specific MoAs exhibit relatively greater efficacy, such as PI3K/mTOR signaling and Chromatin histone acetylation.