ActivityDiff: Classifier-Guided Diffusion for Drug Design

• ActivityDiff is a generative framework for de novo drug design that employs classifier-guided diffusion modeling to balance positive efficacy and negative toxicity signals.
• It utilizes separately trained drug-target classifiers to provide dual guidance, enabling both on-target activation and off-target suppression during molecule generation.
• Experimental validation demonstrates ActivityDiff's capability for multi-target design, fragment-constrained optimization, and enhanced selectivity in drug development.

ActivityDiff is a generative framework for de novo drug design that achieves fine-grained, integrated control over multiple molecular activities—enabling targeted activation or inhibition, cooperative multi-target modulation, and explicit minimization of off-target effects through classifier-guided diffusion modeling. Unlike conventional generative methods that optimize for a single desired activity, ActivityDiff employs separately trained drug-target classifiers to provide both positive (efficacy-enhancing) and negative (toxicity- or liability-reducing) signals during molecule generation, thus balancing efficacy and safety in a unified workflow.

1. Motivation and Problem Statement

In de novo drug design, precise modulation of molecular biological activity—across primary efficacy targets, combination targets, and unintended off-targets—remains a central challenge. Most existing generative models (including VAE, GAN, and standard diffusion-based techniques) focus on enriching a single pharmacological attribute, lacking mechanisms to jointly balance positive (on-target efficacy) and negative (off-target or toxicity) requirements within a single, integrated generative process. Unmanaged off-target effects are a primary cause of attrition in drug development. ActivityDiff is designed to surmount this limitation by providing explicit, independent guidance for multiple targets during molecular graph generation, enabling combinatorial and selectivity-driven optimization.

2. Discrete Diffusion Model and Classifier-Guidance Mechanism

ActivityDiff operates within a discrete denoising diffusion probabilistic model (DDPM) over molecular graphs. Each molecule is represented as a fully connected graph parameterized by node and edge one-hot vectors (encoding atomic and bond types). The forward process systematically corrupts these structured features through iterative addition of Gaussian noise.

Guidance is incorporated during the reverse (denoising) diffusion process via classifier gradients:

• At each reverse step $t$, an intermediate molecular graph $\mathcal{G}_{t-1}$ is decoded.
• For property $y$, a pre-trained drug-target classifier $p_0(y\,|\,\mathcal{G}_{t-1})$ computes the probability of $y$ (e.g., activity against a target).
• The denoising transition is adjusted as:

$$p(\mathbf{x}_{t-1} \mid \mathbf{x}_t, y) \propto p_\epsilon(\mathbf{x}_{t-1} \mid \mathbf{x}_t) \cdot p_0(y \mid \mathcal{G}_{t-1})$$

or, with first-order Taylor expansion over the molecule representations,

$$p(\mathbf{x}_{t-1} \mid \mathbf{x}_t, y) \approx p_\epsilon(\mathbf{x}_{t-1} \mid \mathbf{x}_t) \cdot \left[p_0(y \mid \mathcal{G}_{t-1}) + (\mathcal{G}_{t-1}-\mathcal{G}_{\mathrm{ref}}) \cdot \nabla_{\mathcal{G}_{t-1}} p_0(y \mid \mathcal{G}_{t-1})\right]$$

where $p_\epsilon$ is the unconditional denoising prediction, $\mathcal{G}_{\mathrm{ref}}$ is a reference molecule, and the classifier gradient $$\nabla_{\mathcal{G}_{t-1}} p_0(y \mid \mathcal{G}_{t-1})$$ directs the generation towards (or away from) regions with desired properties.

This scheme allows rapid reconfiguration for different property objectives, as classifiers for new targets or liabilities can be swapped or combined without retraining the entire generative network.

3. Positive and Negative Activity Guidance Paradigm

ActivityDiff uniquely supports simultaneous positive and negative property guidance:

• Positive Guidance: Drug-target classifiers are trained to recognize binding (or activation) for one or more efficacy targets. During generation, positive guidance gradients encourage the model to yield structures with high predicted activity (e.g., $p_0(y^+)\sim1$).
• Negative Guidance: Separate classifiers are trained on binding data for liability or off-targets, or on inactivity data. During generation, their gradients are used to suppress the probability of unwanted interactions (e.g., minimize $p_0(y^-)$).

This dual guidance enables scenarios including:

• Dual-target (polypharmacology) design: Simultaneously maximize predicted affinity for two efficacy targets.
• Selectivity enhancement: Maximize predicted activity for a primary target while minimizing the same for close off-targets, e.g., suppressing EGFR activity while enriching HER2.
• Fragment-constrained design: Conditioning on the presence of structural motifs required for either activity or synthetic tractability.

In each case, the classifier-guided gradients provide direct control over the drift direction in the generative trajectory, yielding molecules whose property distributions match precise, user-specified target profiles.

4. Experimental Validation

The framework was validated on multiple tasks:

• Single-target generation: Molecules generated under positive guidance for a given target exhibited property prediction scores and docking statistics comparable to reference actives, verifying the efficacy-enhancing capability.
• Dual-target and combination design: By combining two classifiers, ActivityDiff generated molecules predicted to co-target proteins (e.g., BRAF/MEK). Statistical metrics, such as the Bhattacharyya coefficient between model and reference property distributions, confirmed alignment with the desired multi-property constraints.
• Fragment-constrained dual-target design: When constrained to retain a functional fragment for one target, ActivityDiff successfully generated molecules optimized for a second, distinct target without sacrificing the primary motif.
• Selectivity and off-target reduction: The integration of negative guidance led to notable reductions in off-target scores (e.g., fewer molecules with high EGFR activity in a HER2 task), resulting in sharper predictive separability compared to baseline or positive-only guidance.

Throughout, the combination of positive and negative guidance allowed fine regulation of generated property distributions as evidenced by the resulting prediction histograms and structural diversity analyses.

5. Impact and Unification in Drug Design Pipelines

ActivityDiff constitutes a unified generative paradigm delivering customizable, property-balanced libraries without post hoc sorting. It:

• Integrates property control directly into sampling, bypassing the inefficiency and potential bias of filter-based approaches.
• Enables extensibility—new pharmacological criteria or toxicity profiles can be incorporated merely by training new classifiers rather than rerunning model optimization.
• Directly supports the medicinal chemistry workflow, as property-guided generation facilitates scaffold hopping, multi-target design, and avoidance of chemical liability.

This suggests rapid ideation cycles for both single- and multi-target lead optimization, lowering the risk of late-stage clinical attrition due to off-target effects or inadequate selectivity.

6. Future Directions and Limitations

The authors point to several avenues for further work:

• Improved property balancing: Methodological refinements could more precisely control the relative strength of multiple objectives, potentially via higher-order guidance terms or dynamic weighting schemes.
• Expanded use of inactivity/negative data: Enhanced training of negative classifiers or direct incorporation of off-target liability datasets can improve discrimination in complex, promiscuous chemical spaces.
• System-level integration: Coupling with systems pharmacology models—encompassing network-based activity, metabolism, and toxicity—may further improve practical predictive performance.
• Diffusion model refinements: Advanced noise scheduling or the inclusion of more flexible discrete noise processes could improve sample diversity and convergence of the guided generation.

A plausible implication is that, as classifier-guided diffusion modeling matures, the ActivityDiff approach will serve as a generalizable engine for multi-property molecular design, adaptable to emerging biological and chemical data modalities.