Predicting Dose-Response Curves with Deep Neural Networks

Pedro Alonso Campana, Paul Prasse, Tobias SchefferICML 2024 · OpenReview

In pharmacology, the "dose-response curve" is the holy grail: it tells researchers exactly how much of a drug is needed to inhibit cell growth. For decades, the industry has relied on the Hill equation, which assumes these curves always follow a simple, symmetric S-shape (sigmoidal).

However, biology is rarely that tidy. Many drugs exhibit "biphasic" behavior—where a low dose might actually stimulate growth before a high dose kills it—which the Hill equation simply cannot model. Furthermore, most current methods reduce these complex curves to a single number, the $IC_{50}$ (the concentration that inhibits 50% of growth). This oversimplification can lead to poor clinical recommendations. Finally, traditional models require experimental data for every specific drug-cell pair to work; they cannot predict how a brand-new drug will affect a specific tumor type it hasn't seen before.

To address these limitations, researchers have developed ARCANet (Anti-Cancer Response Curve Analysis Neural Network), a deep learning framework that predicts the entire dose-response curve of a drug rather than just a single summary statistic. By combining molecular representations with cell-line transcriptomes, the model enables "zero-shot" predictions for entirely new drugs and patient tissues, outperforming traditional parametric models and previous machine learning benchmarks.

The ARCANet Architecture

Unlike models that predict a single scalar value, ARCANet is a "universal approximator" designed to output a continuous functional curve.

ARCANet Architecture

Key components of the architecture include:

  • Dual Encoders: One branch processes the cell's gene expression (transcriptome), while the other processes the drug molecule. The authors tested two drug encodings: traditional Morgan fingerprints (fixed-length bit vectors) and Graph Neural Networks (GNNs) that "read" the actual chemical structure.
  • Interaction Module: A multi-head attention mechanism allows the model to learn which specific genes interact with which parts of a drug molecule.
  • Physics-Informed Architecture: The model is biased toward learning a mixture of Hill-like responses but includes residual connections that allow it to deviate from that shape if the data suggests a more complex (e.g., biphasic) behavior.
  • Hybrid Loss Function: Instead of just using Mean Squared Error (MSE), the network uses a combination of MSE and Kullback-Leibler (KL) divergence. This forces the model to pay attention to the shape of the curve, especially in high-gradient areas, rather than just the average error.

Key Insights and Results

The study utilized five of the largest publicly available repositories of dose-response data (GDSC1, GDSC2, CTRPv2, PRISM, and NCI60) to validate the model.

Escaping the $IC_{50}$ Trap: The authors demonstrate that relying on $IC_{50}$ is dangerous for drugs that don't fit the Hill model. ARCANet successfully captures the "long tail" of drug effects and non-sigmoidal behaviors that traditional 4-parameter logistic models are mathematically incapable of representing.

Zero-Shot Generalization: Because ARCANet learns from the underlying molecular structure and gene expression, it can perform "zero-shot" inference. It can predict the response of a cell line it has never seen before (Precision Oncology) or a molecule it has never seen before (Drug Discovery).

Graphs vs. Fingerprints: A fascinating finding was the trade-off between representations. Morgan fingerprints worked best for interpolating effects on known cell lines, but GNNs were significantly better at generalizing to entirely new drugs and tissues.

Data Efficiency and Limitations

While ARCANet is a high-capacity neural network, it remains highly competitive with simple parametric models even when training data is significantly reduced.

Data Efficiency Comparison

As shown in the data efficiency analysis, ARCANet's performance drops when training data becomes extremely sparse. In very small-scale lab experiments (fewer than ~2,000 data points), simple 3-parameter or 4-parameter (3P/4P) parametric models are still more robust.

Additionally, while the model uses transcriptomes, it does not yet incorporate other "omics" data (like proteomics or metabolomics) which could provide a more holistic view of cell state. Finally, while the attention mechanism provides some clues, fully "opening the black box" to understand why a specific gene-drug interaction leads to a certain curve shape remains a challenge for future interpretability research. Nonetheless, ARCANet represents a significant leap forward in predictive pharmacology, moving the field beyond oversimplified metrics and toward comprehensive, zero-shot curve prediction.

Summary updated

Abstract

Dose-response curves characterize the relationship between the concentration of drugs and their inhibitory effect on the growth of specific types of cells. The predominant Hill-equation model of an ideal enzymatic inhibition unduly simplifies the biochemical reality of many drugs; and for these drugs the widely-used drug performance indicator of the half-inhibitory concentration $IC_{50}$ can lead to poor therapeutic recommendations and poor selections of promising drug candidates. We develop a neural model that uses an embedding of the interaction between drug molecules and the tissue transcriptome to estimate the entire dose-response curve rather than a scalar aggregate. We find that, compared to the prior state of the art, this model excels at interpolating and extrapolating the inhibitory effect of untried concentrations. Unlike prevalent parametric models, it it able to accurately predict dose-response curves of drugs on previously unseen tumor tissues as well as of previously untested drug molecules on established tumor cell lines.
BibTeX
@inproceedings{campana2024predicting,
  author = {Pedro Alonso Campana and Paul Prasse and Tobias Scheffer},
  title = {{Predicting Dose-Response Curves with Deep Neural Networks}},
  booktitle = {International Conference on Machine Learning},
  year = {2024},
  url = {https://openreview.net/forum?id=MDAg5Q7IsI},
}
Figure