Out-of-Distribution Generalization from Labeled and Unlabeled Gene Expression Data for Drug Response Prediction

Hossein Sharifi-Noghabi, Parsa Alamzadeh Harjandi, Olga Zolotareva, Colin C. Collins, Martin EsterbioRxiv 2021 · DOI

TL;DR

This paper introduces Velodrome, a novel semi-supervised method for predicting drug response from gene expression data. Velodrome tackles the challenge of data discrepancy between preclinical and clinical settings by employing out-of-distribution generalization, allowing it to learn from diverse, unlabeled datasets and generalize to unseen patient data without requiring target domain access during training.

What Problem Does This Paper Address?

Accurately predicting how a patient will respond to a drug based on their genetic makeup (pharmacogenomics) is crucial for personalized medicine. However, a major hurdle is the "data discrepancy" between preclinical data (like cell lines and patient-derived xenografts) and real-world patient data. Current transfer learning methods often assume access to target domain data during training, which is unrealistic in clinical settings, or rely solely on limited labeled preclinical data. This paper addresses the need for a method that can leverage abundant unlabeled preclinical data and generalize effectively to unseen patient populations without requiring target domain information during training.

Methods

Velodrome is a deep neural network that combines two main approaches to out-of-distribution generalization: learning domain-invariant features and learning hypothesis-invariant features. It uses a shared feature extractor to map gene expression data from various sources into a common latent space. This is complemented by domain-specific predictors for each source domain. The core innovation lies in its novel objective function, which comprises three components:

  1. Supervised Loss: Ensures the learned features are predictive of drug response using labeled data.
  2. Alignment Loss (CORAL loss): Encourages domain-invariant representations by minimizing the difference between covariance matrices of features from different domains, promoting generalization.
  3. Consistency Loss: Exploits unlabeled data by ensuring that predictions for unlabeled samples are consistent across different predictors, promoting hypothesis-invariant representations.

This combination allows Velodrome to learn from both labeled and unlabeled data from diverse sources, aiming for robust generalization.

Datasets

The study utilizes a variety of publicly available datasets:

  • Patients without drug response (Source Domain): TCGA (Cancer Genome Atlas Research Network) data from multiple cancer types (breast, lung, pancreatic, kidney, prostate, myeloid, lymphoma).
  • Cell lines with drug response (Source Domain): CTRPv2, GDSCv2, and gCSI pan-cancer datasets.
  • PDX samples with drug response (Target Domain): PDX Encyclopedia (PDXE) dataset.
  • Patients with drug response (Target Domain): Several cancer-specific datasets and a pan-cancer dataset from TCGA.

The paper also provides supplementary tables detailing the characteristics of these datasets, including tissue type, number of samples, and number of genes.

Key Insights and Ideas

  • Unlabeled Data is Key: The paper strongly suggests that effectively leveraging unlabeled data is crucial for improving drug response prediction, especially when dealing with diverse and potentially imbalanced datasets.
  • Semi-Supervised Out-of-Distribution Generalization: Velodrome is presented as the first method to combine semi-supervised learning with out-of-distribution generalization for drug response prediction, addressing a significant gap in existing approaches.
  • Hypothesis vs. Domain Invariance: The study highlights the importance of balancing both domain-invariant (alignment loss) and hypothesis-invariant (consistency loss) features for robust generalization, with the consistency loss showing a particularly significant impact on patient predictions.
  • Biological Relevance: Velodrome's predictions are not just statistically significant but also show associations with known target genes of the studied drugs, suggesting it captures underlying biological mechanisms.

Key Results

  • Superior Performance: Velodrome consistently outperformed state-of-the-art methods (including supervised OOD generalization, domain adaptation, and semi-supervised learning baselines) across cell lines, PDX samples, and patient data in terms of Pearson/Spearman correlations, AUROC, and AUPR.
  • Cell Lines: Achieved the second-best performance, demonstrating strong competitiveness.
  • PDX Samples: Achieved the best performance in terms of AUROC and AUPR.
  • Patients: Achieved the best performance in terms of AUROC and AUPR, outperforming competitors.
  • Tissue Generalization: Velodrome demonstrated the ability to generalize to tissue types that were well-represented, under-represented, or completely absent in the training data, including non-solid tissues.
  • Biological Associations: Predictions showed statistically significant associations with known target genes for various drugs and cancer types, confirming biological relevance.

Limitations and Future Work

  • Output Space Discrepancy: The paper acknowledges a limitation in the difference between how drug response is measured in cell lines (concentration-based) versus PDX/patient data (tumor volume change). While they used AAC for cell lines, this difference can affect prediction accuracy. A recent method addresses this but requires target domain access during training, which violates the OOD generalization assumption.
  • Dataset Specificity: The poor/negative correlation for Docetaxel and Paclitaxel on non-solid tissues might be dataset-specific and requires further investigation.
  • Lack of Drug Response Data in TCGA Association Studies: The biological significance analysis relied on TCGA data without drug response labels, meaning direct experimental validation of these specific predictions is a future research direction.
  • Future Directions:
    • Extending Velodrome to incorporate other omics data types (e.g., proteomics, germline variants) and drug information (e.g., chemical representation).
  • Developing explainable versions of Velodrome by replacing the feature extractor with knowledge-based networks.
  • Applying Velodrome to other clinical problems beyond drug response prediction, such as metastasis prediction or rare cancer analysis.
  • Further investigating cross-tissue generalization, especially for rare tissue types.

Summary updated

Abstract

Abstract Data discrepancy between preclinical and clinical datasets poses a major challenge for accurate drug response prediction based on gene expression data. Different methods of transfer learning have been proposed to address this data discrepancy. These methods generally use cell lines as source domains and patients, patient-derived xenografts, or other cell lines as target domains. However, they assume that they have access to the target domain during training or fine-tuning and they can only take labeled source domains as input. The former is a strong assumption that is not satisfied during deployment of these models in the clinic. The latter means these methods rely on labeled source domains which are of limited size. To avoid these assumptions, we formulate drug response prediction as an out-of-distribution generalization problem which does not assume that the target domain is accessible during training. Moreover, to exploit unlabeled source domain data, which tends to be much more plentiful than labeled data, we adopt a semi-supervised approach. We propose Velodrome, a semi-supervised method of out-of-distribution generalization that takes labeled and unlabeled data from different resources as input and makes generalizable predictions. Velodrome achieves this goal by introducing an objective function that combines a supervised loss for accurate prediction, an alignment loss for generalization, and a consistency loss to incorporate unlabeled samples. Our experimental results demonstrate that Velodrome outperforms state-of-the-art pharmacogenomics and transfer learning baselines on cell lines, patient-derived xenografts, and patients. Finally, we showed that Velodrome models generalize to different tissue types that were well-represented, under-represented, or completely absent in the training data. Overall, our results suggest that Velodrome may guide precision oncology more accurately.
BibTeX
@misc{sharifinoghabi2021out,
  author = {Hossein Sharifi-Noghabi and Parsa Alamzadeh Harjandi and Olga Zolotareva and Colin C. Collins and Martin Ester},
  title = {{Out-of-Distribution Generalization from Labeled and Unlabeled Gene Expression Data for Drug Response Prediction}},
  year = {2021},
  doi = {10.1101/2021.05.25.445658},
  howpublished = {bioRxiv},
}
Figure