Abstract
Drug response prediction (DRP) is a crucial phase in drug discovery, and the most important metric for its evaluation is the IC50 score. DRP results are heavily dependent on the quality of the generated molecules. Existing molecule generation methods typically employ classifier-based guidance, enabling sampling within the IC50 classification range. However, these methods fail to ensure the sampling space range's effectiveness, generating numerous ineffective molecules. Through experimental and theoretical study, we hypothesize that conditional generation based on the target IC50 score can obtain a more effective sampling space. As a result, we introduce regressor-free guidance molecule generation to ensure sampling within a more effective space and support DRP. Regressor-free guidance combines a diffusion model's score estimation with a regression controller model's gradient based on number labels. To effectively map regression labels between drugs and cell lines, we design a common-sense numerical knowledge graph that constrains the order of text representations. Experimental results on the real-world dataset for the DRP task demonstrate our method's effectiveness in drug discovery.
Introduction
The search for novel drug-like compounds is a monumental task, with the chemical search space estimated between $10^{23}$ and $10^{60}$ molecules. A primary objective in this search is optimizing the IC50 score, which measures the concentration of a drug required to inhibit a biological process by half. Traditional generative models for drug discovery often rely on classifier-based guidance, categorizing molecules into broad bins such as "active" or "inactive."
However, this approach is often insufficient. Because biological responses are continuous and quantifiable, classification creates a sampling space that is too broad, leading to the generation of many ineffective candidates. As shown in Figure 1, regressor-based guidance allows for targeting specific points rather than broad ranges, significantly narrowing the search space to more effective regions.
Methodology
To address the limitations of classification, we propose a framework centered on three core innovations:
1. Regressor-free Guidance
Inspired by classifier-free guidance in image synthesis, this technique enables a diffusion model to generate molecules conditioned on continuous numerical IC50 targets. By blending conditional and unconditional scores, the model navigates the chemical space toward specific potency targets without requiring a separate, gradient-based regressor during inference.
2. Common-Sense Numerical Knowledge Graph (CN-KG)
Deep learning models often struggle to interpret the inherent order of numerical values represented as text. We introduce a CN-KG to constrain text representations of numbers, ensuring the model understands that a value like 0.5 is quantitatively less than 0.6. This bridges the gap between discrete text encoders and continuous biological metrics.
3. DBControl (Dual-Branch Controlled Noise Prediction)
Biological datasets like GDSCv2 are often small, posing a risk of overfitting. Our DBControl architecture utilizes two branches: one trained on large-scale unconditional data (e.g., QM9) to learn general chemical validity, and a second branch that learns specific drug-cell line response controls. These are integrated via "zero convolution" layers to maintain chemical "sanity" while being highly responsive to specific biological conditions.
Experimental Results
The model was evaluated on the GDSCv2 dataset for DRP tasks, using unconditional datasets like QM9, ZINC250k, and GDB-17 for general molecular training.
Quantitative Performance
Our method outperformed state-of-the-art baselines (GDSS, EDM, MOOD) across key metrics:
- FCD (Fréchet ChemNet Distance): Improved by 2.68%, indicating higher chemical plausibility.
- MMD (Maximum Mean Discrepancy): Improved by 2.23%, showing a better match between generated and target distributions.
Qualitative Analysis
UMAP visualizations confirmed that molecules generated via regressor-free guidance cluster more tightly around target biological features compared to unconditional or classifier-based methods. We found that a guidance strength ($w$) of 1.0 to 5.0 offers the optimal balance between target adherence and molecular diversity.
Conclusion
By shifting from discrete classification to continuous regression targets, our framework provides a more precise and effective sampling space for drug discovery. While the computational intensity of diffusion models and the need for wet-lab validation remain challenges, this work demonstrates a significant step toward more targeted and efficient molecule generation for drug response prediction.
Code available at: https://anonymous.4open.science/r/RMCD-DBD1
Summary updated