Research map / Probabilistic and scientific modeling / Molecules and proteins
Predicting Dose-Response Curves with Deep Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ARCANet, a neural network that predicts entire drug dose-response curves from drug and cell-line transcriptome embeddings, using a KL-plus-MSE loss on softmax-normalized curves.
Problem
Dose-response curves are usually modeled by parametric Hill-type curves fitted per drug-cell pair, which cannot handle biphasic drugs, unseen drugs or cell lines, and reduce curves to scalar IC50 values that can mislead drug selection.
Approach
A non-parametric deep network encodes drug (Morgan fingerprint or molecular graph) and cell-line gene expression into an interaction embedding, then produces a mixture of latent Hill responses per concentration, trained with a variational KL loss over whole curves plus MSE.
Key results
- ARCANet (fingerprint) gives the lowest extrapolation MSE on three of four datasets (e.g., PRISM 0.854 vs 1.503 for best individual 4P), with significance in those cases.
- For precision oncology on unseen cell lines, ARCANet (GNN) reduces PRISM MSE from 1.185 (FunFor) to 0.425.
- Parametric 3P/4P individual models only beat ARCANet when training data is reduced by roughly two orders of magnitude (to a few thousand measurements).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ARCANet with Morgan fingerprints instead of the strongest individual Hill baseline | Dose-response prediction · CTRPv2, GDSC2, NCI60, and PRISM smoothing | MSE: 0.022 -> 0.011; 0.005 -> 0.003; 0.099 -> 0.047; 1.055 -> 0.271 | Helped |
| use ARCANet with Morgan fingerprints instead of the strongest reference baseline | Dose-response prediction · CTRPv2, GDSC2, and PRISM interpolation | MSE: 0.013 -> 0.012; 0.004 -> 0.003; 1.166 -> 0.68 | Helped |
| use the 4P individual model instead of ARCANet for interpolation | Dose-response prediction · NCI60 | MSE: 0.051 -> 0.045 | No clear effect |
| use ARCANet with Morgan fingerprints instead of the strongest reference baseline | Dose-response prediction · CTRPv2, GDSC2, NCI60, and PRISM extrapolation | MSE: 0.116 -> 0.053; 0.011 -> 0.008; 0.309 -> 0.201; 1.957 -> 0.854 | Helped |
| use ARCANet instead of FunFor for unseen-cell-line predictions | Precision oncology · CTRPv2, GDSC2, and PRISM | MSE: 0.063 -> 0.03; 0.009 -> 0.005; 1.185 -> 0.422 | Helped |
| use ARCANet (GNN) instead of FunFor for unseen-drug predictions | Drug discovery · NCI60 | MSE: 0.14 -> 0.12 | Helped |
| replace the mixture of Hill equations with a concatenation-based treatment model (ablation) | ARCANet (fingerprint) · PRISM precision oncology and smoothing | MSE: 0.422 -> 0.423; 0.271 -> 0.272 | Hurt |
| set α to zero, removing the variational loss component (ablation) | ARCANet (fingerprint) · PRISM precision oncology and smoothing | MSE: 0.422 -> 0.431; 0.271 -> 0.281 | Hurt |
| replace the regression module with a fully connected layer (ablation) | ARCANet (fingerprint) · PRISM precision oncology and smoothing | MSE: 0.422 -> 0.427; 0.271 -> 0.273 | Hurt |
| replace the mixture of Hill equations with a concatenation-based treatment model (ablation) | ARCANet (GNN) · PRISM precision oncology and smoothing | MSE: 0.425 -> 0.434; 0.335 -> 0.337 | Hurt |
| set α to zero, removing the variational loss component (ablation) | ARCANet (GNN) · PRISM precision oncology and smoothing | MSE: 0.425 -> 0.435; 0.335 -> 0.329 | Mixed |
| replace the regression module with a fully connected layer (ablation) | ARCANet (GNN) · PRISM precision oncology and smoothing | MSE: 0.425 -> 0.422; 0.335 -> 0.327 | Mixed |
Surprising findings
- Parametric 3P/4P models outperform ARCANet only when training data shrinks by about two orders of magnitude, indicating the high-capacity network is data-hungry.
- For NCI60 interpolation, the 4P individual baseline has lower MSE than ARCANet (fingerprint), though not significantly.
Limitations
- Parametric baselines are better when training data is very small
- Graph representation has lower performance on seen cell-line and drug combinations than fingerprints
- 2P mixed-effect model is numerically unstable and only converges on GDSC2
Experimental setup
Datasets: CTRPv2 NCI60 GDSC1 GDSC2 PRISM
Benchmarks: CTRPv2 GDSC2 NCI60 PRISM
Metrics: MSE standard error two-tailed t-test
Baselines: 2P individual 3P individual 4P individual 2P mixed-effect logistic model (Vis et al., 2016) FunFor (Fu et al., 2021)
Builds on: Hill equation mixed-effect logistic model (Vis et al., 2016) functional random forest (Fu et al., 2021)
Closest papers by meaning
- Contrastive Balancing Representation Learning for Heterogeneous Dose-Response Curves Estimation (AAAI 2024 · similarity 0.83)
- Knowledge-Guided Domain Adaptation Model for Transferring Drug Response Prediction from Cell Lines to Patients (AAAI 2025 · similarity 0.83)
- WISER: Weak Supervision and Supervised Representation Learning to Improve Drug Response Prediction in Cancer (ICML 2024 · similarity 0.82)
- Learning Identifiable Factorized Causal Representations of Cellular Responses (NeurIPS 2024 · similarity 0.82)
- ScreenShot: A Foundation Model for Few-Shot Combination Drug Screening (NeurIPS 2026 · similarity 0.82)
- Customized Subgraph Selection and Encoding for Drug-drug Interaction Prediction (NeurIPS 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.