Network-based machine learning in colorectal and bladder organoid models predicts anti-cancer drug efficacy in patients.

Nat Commun 2020 AI 7 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Pages 1-2
The Challenge of Predicting Drug Response

Most Biomarkers Fail to Translate. Identifying which patients will respond to a given chemotherapy drug is critical for improving outcomes and avoiding unnecessary toxicity. While large-scale drug screening in cell lines and genomic studies have identified many potential biomarkers, these frequently fail to predict responses in actual patients with cancer.

Machine learning models trained on preclinical data often struggle because of a fundamental mismatch: there are far more genes measured (thousands of expression features) than drug response data points (hundreds of organoid or cell line samples). This disproportion, combined with the biological differences between simple cell cultures and human tumors, leads to unstable and non-generalizable models.

Three-dimensional organoid cultures represent a promising middle ground. These tumor-derived mini-organs grown in the lab closely resemble the molecular profiles and drug response patterns of actual human tumors, making them a more faithful preclinical model than traditional flat cell cultures for biomarker discovery.

TL;DR: Standard machine learning approaches for predicting chemotherapy response fail to generalize from lab models to patients, motivating the need for a more biologically informed framework.
Pages 1-2
Using Protein Networks to Guide Biomarker Discovery

Network Proximity as a Feature Filter. The key innovation of this study is using a protein-protein interaction (PPI) network to filter candidate biomarker pathways before training any machine learning model. Instead of searching among all 674 Reactome biological pathways, only pathways located close to the drug's target in the PPI network are considered as candidates.

The logic is grounded in network medicine: genes involved in similar biological functions tend to cluster in the same network neighborhood. Therefore, pathways relevant to a drug's mechanism of action should be near its molecular targets in the interaction network. Using the STRING PPI network with 13,824 proteins and over 323,000 interactions, the team calculated network distances between each drug's known targets and every biological pathway.

Pathways closer to drug targets than expected by chance were selected as input features. The expression levels of these pre-selected pathways in organoid models were then used to train a ridge regression machine learning model against organoid drug sensitivity (IC50 values). Pathways with the highest predictive performance in organoids were selected as biomarkers and applied to classify real cancer patients.

TL;DR: The method filters candidate biomarker pathways using their proximity to drug targets in a protein interaction network before machine learning training, reducing noise and overfitting.
Pages 2-4
Organoid Pharmacogenomics and Patient Validation

Two Cancer Types, Two Drugs. The framework was tested on two distinct cancer-drug combinations. For colorectal cancer treated with 5-fluorouracil (5FU), data from 19 colorectal cancer organoid samples were used. For bladder cancer treated with cisplatin, 9 bladder cancer organoid samples were used. Drug sensitivity was measured as IC50 values from these organoid drug screens.

Predictive pathway biomarkers identified in organoids were then validated in clinical patient populations. For 5FU, 114 colorectal cancer patients from TCGA with known treatment history and survival outcomes were used. For cisplatin in bladder cancer, 77 patients were classified using the biomarker pathway expression levels from their tumor gene expression data.

Patients were divided at the median predicted drug resistance score into predicted responders and non-responders. Kaplan-Meier survival analysis and log-rank tests were used to assess whether the biomarker-based classification predicted meaningful differences in overall survival. Negative control cohorts of untreated patients were used to confirm that observed survival differences were drug-specific rather than simply prognostic of baseline patient health.

The biomarkers were additionally validated in isogenic drug-sensitive and drug-resistant cancer cell lines, tested against 10,000 iterations of random pathway selection (bootstrapping), and compared against known mutation-based biomarkers from independent clinical datasets to comprehensively establish robustness.

TL;DR: Biomarkers discovered in organoid screens were validated in 114 colorectal and 77 bladder cancer patients, untreated control cohorts, isogenic cell lines, and random bootstrapping tests.
Pages 3-4
Biomarkers Discovered and Validated in Patients

Activation of BH3-Only Proteins for 5FU in Colorectal Cancer. Among the 37 pathways proximal to the 5FU drug target in the PPI network, the pathway controlling activation of BH3-only proteins showed the highest predictive performance in colorectal cancer organoids. BH3-only proteins are the essential initiators of apoptosis (programmed cell death) in response to DNA damage, which is how 5FU works.

In 114 5FU-treated colorectal cancer patients, those classified as predicted responders based on BH3-only protein pathway expression had significantly better overall survival (P = 0.014). Notably, no deaths were observed among predicted responders during the study period, while predicted non-responders had only about 50% 5-year survival. In 298 untreated colorectal patients, the same biomarker showed no survival difference (P = 0.16), confirming the finding was drug-specific.

Amino Acid Synthesis for Cisplatin in Bladder Cancer. Among 30 pathways proximal to cisplatin targets in the network, the amino acid synthesis and interconversion pathway showed the highest predictive performance in bladder cancer organoids. In 77 cisplatin-treated bladder cancer patients, predicted responders had significantly better 5-year survival (approximately 75%) compared to non-responders (approximately 50%), with a log-rank P value of 0.01.

Cross-validation confirmed that the biomarkers were cancer-type and drug-specific. Swapping the biomarkers (using the colorectal cancer biomarker for bladder cancer patients and vice versa) produced statistically insignificant survival predictions, confirming these are not generic prognostic markers but genuinely reflect drug response biology.

TL;DR: The BH3-only protein pathway predicted 5FU response in colorectal cancer (P=0.014) and the amino acid synthesis pathway predicted cisplatin response in bladder cancer (P=0.01), both confirmed drug-specific and not merely prognostic.
Pages 5-7
Why Other Machine Learning Approaches Failed

Network Proximity is the Critical Difference. The study systematically compared the network-proximity approach against five alternative methods to understand why it worked while others did not. Using whole-genome transcriptomics as features failed (P = 0.79 to 0.82 for both cancers). Using all-pathway transcriptomics as features also failed (P = 0.82 to 0.82).

Selecting features based on network centrality (how connected a pathway is in the network, using degree, betweenness, or closeness measures) also failed to predict patient survival. Similarly, selecting only the direct first, second, or third-degree network neighbors of drug targets without the pathway-level abstraction failed to classify patient outcomes with statistical significance.

A published feature selection method by Bolis et al. based on expression correlations with drug sensitivity produced P values of 0.15 and 0.96, well above significance. A state-of-the-art deep learning model by Sharifi-Noghabi et al. trained on the same data produced P values of 0.91 and 0.82. Every competing approach failed where the network-proximity pathway method succeeded.

The results highlight that the critical insight is not network connectivity per se, but network proximity specifically at the pathway level. Drug-relevant biological signals are captured when pathways that collectively cluster near drug targets are treated as features, rather than individual genes or network-central pathways that may be biologically irrelevant to the drug's mechanism.

TL;DR: Five alternative machine learning approaches all failed to predict patient survival for both cancer types, while only the network proximity-guided pathway selection succeeded, demonstrating that this specific feature selection strategy is critical.
Pages 7-9
Cisplatin Biomarker Connects to Known Clinical Biology

ERCC2 Mutations and Amino Acid Metabolism Align. A key validation was whether the computationally discovered biomarkers correlated with independently known clinical biomarkers based on somatic mutations rather than expression data. For bladder cancer and cisplatin, ERCC2 mutations in the DNA repair gene are a clinically established predictor of cisplatin sensitivity.

Bladder cancer patients with ERCC2 mutations had significantly lower predicted drug resistance scores based on the amino acid synthesis pathway biomarker (P = 0.002), meaning ERCC2-mutated patients were correctly classified as more likely to respond to cisplatin. This concordance between a transcriptomic biomarker and an independent mutation-based biomarker provides strong cross-validation that the pathway captures real biology.

The mechanism linking amino acid metabolism to cisplatin resistance was independently supported by prior research showing that cisplatin-resistant bladder cancer cells have hypermethylated gene promoters in amino acid and polyamine metabolism pathways, reducing expression of these genes. The biomarker discovered computationally from organoid data aligns with experimental evidence of mechanistic resistance.

Clinical trials are already exploring amino acid depletion therapy in cancer (such as ADI-PEG 20 treatment). The finding that reduced amino acid synthesis pathway expression predicts cisplatin resistance in bladder cancer raises the possibility that combining amino acid depletion approaches with cisplatin could restore or enhance cisplatin sensitivity in resistant tumors.

TL;DR: The cisplatin response biomarker correlated with ERCC2 mutation status (P=0.002), a clinically established predictor, and aligns with experimental evidence that amino acid metabolism drives cisplatin resistance in bladder cancer.
Pages 9-10
Organoids as a Bridge to Clinical Prediction

Organoids Outperform Standard Cell Lines. Using large-scale pharmacogenomic data from conventional cancer cell lines (the GDSC database) instead of organoid data failed to strongly predict patient survival for either cancer type. This reinforces that the molecular fidelity of patient-derived organoids to primary tumors is essential for successful biomarker translation.

The bootstrapping analysis confirmed the findings were not due to chance. In 10,000 iterations of random pathway selection, the actual network-proximity selected pathways ranked higher than random selections in predicting both patient survival and drug sensitivity in isogenic cell lines, with empirical P values of 0.0012 for colorectal cancer and 0.014 for bladder cancer.

The framework provides interpretable results, a key advantage over black-box deep learning approaches. Because the biomarkers are defined by specific biological pathways near drug targets, the predictions come with mechanistic hypotheses about why patients respond or not, which can be experimentally validated and potentially targeted therapeutically.

As organoid technology advances to incorporate immune cells and stromal components of the tumor microenvironment, this approach is expected to become even more powerful. The authors envision using time-paired molecular datasets before and after drug treatment, and integrating multiple molecular data layers into the network-based framework to further improve precision oncology for bladder and colorectal cancer patients.

TL;DR: Patient-derived organoids produce more clinically relevant biomarkers than conventional cell lines, and the network-guided framework is interpretable and extensible to multi-omic data and more complex organoid systems.
Citation: Open Access, 2020. Available at: PMC7599252.