Source-linked AI summary
Synergistic Drug Combination Prediction by Integrating Multi-omics Data in Deep Learning Models
Tianyu Zhang, Liwei Zhang, Philip R. O. Payne, Fuhai Li
TL;DR
Experimental drug-combination screening is difficult given the enormous number of possible pairs and varied synergy mechanisms. The paper introduces AuDNNsynergy, which integrates TCGA-trained omics encoders with drug features, and reports better performance than four state-of-art approaches across ten model measurements.
Problem
Experimental screening is challenging because the number of possible drug combinations is enormous and synergy mechanisms are heterogeneous.
Source: 1.Introduction · p. 2Method
AuDNNsynergy uses autoencoders trained on TCGA expression, mutation, and copy-number data, combining their cell-line representations with drug physicochemical features to predict synergy.
Source: 1.Introduction · p. 3Results
AuDNNsynergy outperformed four state-of-art approaches in ten model measurements.
Source: 1.Introduction · p. 3Takeaways & Limitations
The authors report that transfer learning from TCGA data can improve prediction accuracy.
Source: §6 · p. 10The authors note that few drug-combination datasets had been reported.
Source: §6 · p. 10
Abstract
from arXiv · showhide
Drug resistance is still a major challenge in cancer therapy. Drug combination is expected to overcome drug resistance. However, the number of possible drug combinations is enormous, and thus it is infeasible to experimentally screen all effective drug combinations considering the limited resources. Therefore, computational models to predict and prioritize effective drug combinations is important for combinatory therapy discovery in cancer. In this study, we proposed a novel deep learning model, AuDNNsynergy, to prediction drug combinations by integrating multi-omics data and chemical structure data. In specific, three autoencoders were trained using the gene expression, copy number and genetic mutation data of all tumor samples from The Cancer Genome Atlas. Then the physicochemical properties of drugs combined with the output of the three autoencoders, characterizing the individual cancer cell-lines, were used as the input of a deep neural network that predicts the synergy value of given pair-wise drug combinations against the specific cancer cell-lines. The comparison results showed the proposed AuDNNsynergy model outperforms four state-of-art approaches, namely DeepSynergy, Gradient Boosting Machines, Random Forests, and Elastic Nets. Moreover, we conducted the interpretation analysis of the deep learning model to investigate potential vital genetic predictors and the underlying mechanism of synergistic drug combinations on specific cancer cell-lines.
1.Introduction
The introduction frames experimental screening as challenging because the combination space is enormous and synergy mechanisms are heterogeneous. It presents AuDNNsynergy as an integrated predictor that outperformed existing approaches and was interpreted to investigate genetic predictors and synergy mechanisms.
1.Introduction
- Experimental screening is challenging because the number of possible combinations is enormous and their synergy mechanisms are heterogeneous.The introduction argues that computational prediction is important for identifying combinations for specific cancer subtypes or patients.Source: 1.Introduction · p. 2
- The introduction identifies overfitting, limited cell-line training data, poor generalization, and model opacity as challenges for deep-learning predictors.These issues motivate the authors’ focus on accuracy and model interpretation.Source: 1.Introduction · p. 3
- AuDNNsynergy combines gene-expression, mutation, and copy-number data with drug physicochemical properties to predict drug-combination synergy.Three autoencoders trained on TCGA tumor samples encode the omics data; the study also applies a cross-validation technique to select model hyperparameters.Source: 1.Introduction · p. 3
- AuDNNsynergy outperformed four state-of-art approaches in ten model measurements.The comparison is reported alongside integration of multi-omics and drug physicochemical features.Source: 1.Introduction · p. 3
- The authors interpret the model to investigate potential driver genetic features and underlying mechanisms of synergistic combinations.The stated aim is to identify candidate predictors and explore mechanisms, not to establish them as causal.Source: 1.Introduction · p. 3
2. Materials
The materials section describes the drug-combination screening data and the cancer-cell-line and TCGA omics inputs, then outlines their preprocessing into model features.
2.1 Data
- The screening dataset contains 23,062 measurements of 583 drug pairs across 39 cancer cell lines from seven cancer types.Cell viability after 48-hour treatment at four doses per drug was used to estimate synergy.Source: §2.1 · p. 3
- The study assembled gene-expression, mutation, and copy-number data for cancer cell lines and TCGA tumor samples.RNA-seq data covered 1,156 cell lines and 10,535 TCGA tumor samples; some unavailable cell-line profiles were substituted with data from related lines.Source: §2.1 · pp. 3–4
- Drug structures were represented using radius-6 extended-connectivity fingerprints, physicochemical properties, and toxicophore-substructure presence or absence.The properties were obtained using jCompoundMapper and Chemopy.Source: §2.1 · p. 4
2.2 Data pre-processing
- Preprocessing log-transformed gene expression, encoded non-silent mutations as binary variables, estimated copy number with GISTIC2, and removed zero-variance features.Synergy scores were based on the Loewe Additivity model; the passage also reports zero-mean, unit-variance normalization.Source: §2.2 · p. 4
- After preprocessing, cell lines had 50,196 expression, 17,643 mutation, and 20,318 copy-number features; each drug had 2,431 features.The resulting drug-combination/cell-line input is described as having 93,019 features.Source: §2.2 · p. 4
3. Model
The model uses TCGA-trained autoencoders to encode cell-line omics data for a deep neural network, with cross-validation and relevance propagation used for evaluation and interpretation. The section also describes the computing environment and the baseline-comparison setup.
3.1 Architecture of AuDNNsynergy
- Three autoencoders encode cell-line gene-expression, mutation, and copy-number data, and their outputs feed the prediction network.The autoencoders are trained using TCGA multi-omics data to transfer knowledge from the larger tumor dataset.Source: §3.1 · pp. 4–5
3.2 Cross-validation and model interpretation analysis
- Five-fold cross-validation estimates prediction performance, with each drug pair appearing in only one fold; layer-wise relevance propagation analyzes prediction features.Innvestigate is used to identify important drugs and encoded genomic features contributing to synergy scores.Source: §3.2 · pp. 5–6
3.3 Computation environment of AuDNNsynergy
- The model was deployed on Google Cloud Platform with an NVIDIA Tesla V100 GPU and the specified Ubuntu, CUDA, cuDNN, Keras, and TensorFlow setup.The listed system included four vCPUs and 32 GB of memory.Source: §3.3 · p. 6
3.4 Model comparison
- AuDNNsynergy was compared with DeepSynergy, Gradient Boosting Machines, Random Forests, and Elastic Nets on 8,846-feature data; the other models were also evaluated on 93,019-feature data.DeepSynergy could not be trained on the 93,019-feature dataset because the GPU memory was insufficient.Source: §3.4 · p. 6
4. Results
AuDNNsynergy outperformed the four comparison models, with strong drug-pair ranking results across cell lines and drugs. Its prediction performance varied across settings, and some cell-line predictions had large errors.
4.1 Synergistic drug combination prediction
- Drug-level rank correlations ranged from 0.55 to 0.86, with about 67% above 0.7 and about 67% of drugs having MSE below 200.Sunitinib and Doxorubicin were relatively easy to predict, with rank correlations above 0.7 and MSE below 100.Source: §4.1 · p. 7
- For over 74% of cell lines, the model correctly predicted the most synergistic pairs; its top predicted pair fell among the experimental top five for over 93%.Cell-line rank correlations ranged from 0.56 to 0.81, and over 55% of cell lines had correlations above 0.7.Source: §4.1 · p. 7
- The authors identified 51 poor predictions—2‰ of all data points—where predicted and experimental synergy values differed by over 100.The LNCAP, NCIH2122, and NCIH23 cell lines contributed 10, 9, and 12 such samples, respectively.Source: §4.1 · p. 7
4.2 Model comparison
- AuDNNsynergy outperformed four models, scoring 0.73 rank correlation and 241.12 MSE, and led on ROC AUC, PR AUC, ACC, PREC, and Kappa.The comparison covered Gradient Boosting Machines, Random Forests, Elastic Nets, and DeepSynergy.Source: §4.2 · pp. 7–8
- DeepSynergy was difficult to apply directly to the 93,019-feature dataset because optimization would involve more than 600 million parameters.The authors state it could not be run on the Google Cloud platform.Source: §4.2 · p. 8
5. Interpretation analysis
Interpretation analysis identified influential omic features and case-specific predictors associated with drug-combination responses. The authors also used AuDNNsynergy to predict combinations for TCGA samples beyond the screened cell lines.
5.1 Important omic features for the overall prediction
- More than 50% of the most important features were chemical structures; the selected omic features comprised 2 gene-expression, 25 mutation, and 14 copy-number genes.The number of important omic features changed little as more features were selected.Source: §5.1 · p. 8
5.2 Mechanism of synergy for specific drug combinations against given cell lines
- For Dasatinib–Carboplatin in ovarian cell line A2780, copy number of PTPRK was the most important gene predictor; PTPRK dephosphorylates SRC.For Sorafenib–Erlotinib, PAIP2 had a positive effect on response and miR-507 an adverse effect.Source: §5.2 · p. 9
5.3 Synergistic drug combination prediction for TCGA samples
- The authors applied AuDNNsynergy to predict synergy for 853 drug combinations in TCGA samples and report predictions for tumors beyond the 39 screened cell lines.They attribute robust predictions for new tumor samples to training on TCGA samples and transfer of tumor-sample knowledge.Source: §5.3 · p. 9
- For NSCLC with EGFR mutation, Dasatinib–Erlotinib ranked 7/583 and Dasatinib–Lapatinib ranked 5/583 in the predictions.The authors note that tumor and cell-line omics profiles might differ, although they consider some prediction results meaningful.Source: §5.3 · p. 9
- ABT-888/BEZ-235, MK-4827/BEZ-235, and MK4827/MK8669 were among the top-ranked predicted combinations, in the top 4%-7%.ABT-888 and MK-4827 target PARP.Source: §5.3 · p. 9
6. Discussion
AuDNNsynergy combines multi-type genomic representations with drug information and is reported to outperform existing drug-combination prediction models. The discussion also reports improved accuracy with TCGA transfer learning and interpretation findings, while noting dataset and measurement constraints.
6. Discussion
- AuDNNsynergy outperforms existing drug-combination prediction models.The model integrates multiple types of genomics data and drug information.Source: §6 · p. 10
- Transfer learning with TCGA data can improve prediction accuracy.The model uses autoencoders for gene expression, copy number, and mutation data.Source: §6 · p. 10
- Interpretation analysis identified omics features and potential synergy mechanisms for specific drug combinations in cancer cell lines.The discussion reports these as potential mechanisms, not established causal explanations.Source: §6 · p. 10
- Future extensions include larger screening datasets, additional genomics data types, and drug-structure graphs.Source: §6 · p. 10
- Drug-combination datasets are few and use heterogeneous synergy measurements, motivating merged screening data under a standard effectiveness measure.Source: §6 · p. 10