Discovering Synergistic Drug Combinations for COVID with Biological Bottleneck Models

Wengong Jin, Regina Barzilay, Tommi S. JaakkolaarXiv 2020 · arXiv

Introduction

Finding synergistic drug combinations—where two drugs together achieve higher efficacy than the sum of their individual effects—is a primary goal in pharmacology. Synergy allows for higher efficacy with lower dosages, which in turn reduces toxic side effects. However, the search space for these combinations is vast, and experimental screening is prohibitively expensive. While deep learning has shown promise in predicting synergy for well-studied diseases like cancer, these models typically require hundreds of thousands of data points. For emerging diseases like COVID-19, where clinical data is extremely scarce, standard "black-box" neural networks often fail due to overfitting and a lack of biological context.

Researchers from MIT CSAIL have addressed this challenge by developing a "biological bottleneck" model. This framework forces the neural network to first predict drug-target interactions (DTI) before estimating antiviral activity, effectively using biological principles as a regularizer to improve generalization from small datasets.

The Biological Bottleneck Framework

The core innovation of this study is the decomposition of the synergy prediction task into two biologically-inspired stages, as illustrated in the architecture below.

Schematic of the Biological Bottleneck Model architecture
  1. Drug-Target Interaction (DTI) Module: A Graph Convolutional Network (GCN) processes a drug's molecular structure to predict its interaction with a set of biological targets (proteins). This creates a "DTI vector" ($z$), which serves as the biological bottleneck.
  2. Target-Disease Association Module: A linear layer maps these biological footprints to the final antiviral effect.

To predict the combined effect of two drugs, the model employs a Bliss Layer. This layer uses "Bliss independence" logic to combine the target profiles of two drugs ($z_A$ and $z_B$) into a single vector ($z_{AB} = z_A + z_B - z_A \odot z_B$), representing the union of their biological effects. Furthermore, the model incorporates "latent targets"—hidden variables that allow the AI to learn unknown biological mechanisms not yet documented in literature. This approach ensures that the model's predictions are grounded in protein interactions rather than random statistical correlations.

Data and Training Strategy

To overcome the scarcity of COVID-19 data (only 90 training combinations), the researchers utilized multi-disease training. The model was co-trained on HIV data, leveraging shared antiviral mechanisms. The training set included data on approximately 20,000 compounds from the ChEMBL database and 42 known biological targets, such as the viral protease 3CLpro and the host protein ACE2, supplemented by 58 latent targets.

Experimental Validation and Results

The biological bottleneck approach significantly outperformed standard models. The full model achieved a 0.78 AUC on the COVID-19 test set, whereas a standard GCN without the bottleneck performed at 0.537 AUC, barely better than random chance. Ablation studies confirmed that the inclusion of latent targets was crucial, as their removal dropped performance to 0.718.

Following the computational ranking of 11,600 potential combinations, the top 30 candidates were tested in a high-security (BSL-3) laboratory. This experimental validation identified two novel synergistic combinations that showed significantly stronger inhibition of SARS-CoV-2 in Vero E6 cells than the individual drugs alone.

Validated synergistic drug combinations: Remdesivir + Reserpine and Remdesivir + IQ-1S

Conclusion and Future Outlook

The success of the biological bottleneck model demonstrates that incorporating domain-specific constraints can enable deep learning to succeed even in data-poor regimes. By identifying combinations like Remdesivir + Reserpine and Remdesivir + IQ-1S, the researchers have provided new leads for COVID-19 treatment.

While these results are promising in vitro, further in vivo studies are necessary to confirm safety and efficacy in humans. Nevertheless, this framework serves as a blueprint for rapid-response drug discovery for future emerging infectious diseases, where initial data will inevitably be limited.

Summary updated

BibTeX
@misc{jin2020discovering,
  author = {Wengong Jin and Regina Barzilay and Tommi S. Jaakkola},
  title = {{Discovering Synergistic Drug Combinations for COVID with Biological Bottleneck Models}},
  year = {2020},
  eprint = {2011.04651},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2011.04651},
}
Figure