Highlighting methodological differences and superiority of UniGuide
Thank you for your kind answer and for increasing the score. We are glad that our clarifications and additional results have addressed most of the reviewer's concerns. Below, we would like to offer further clarifications on UniGuide's superiority and distinct advantages over alternative guidance methods that control generation during inference:
> Highlight the difference to other controlling techniques
- **UniGuide maps the condition to the configuration space**:
- Prior works [1,2,3], centred around classifier-guidance [1], guide towards conditions that describe global input properties. For example, [2,3] control the generation of molecules given a desired quantum property. We highlight that such prior works are limited in their applications and cannot (directly) condition on geometric conditions as introduced in our work.
- [1,2,3] control the generation during inference usually by mapping the output of the diffusion model to the condition's space $(Z \to S)$, by using a classifier/regressor $f_{\theta}(z_t)$. In other words, the distribution $p_{\theta}(c |z_t)$ takes the form $p_{\theta}(c|z_t) = \mathcal{N}(c |f_{\theta}(z_t), I).$ In contrast, UniGuide maps the source condition to the configuration space $(S \to Z)$ which simplifies the gradient signal computation $\nabla_{z_t} p_{\theta}(c |z_t)$ to a squared error between the sample $z_t$ and the target condition $c = C({s}, \hat{z}_t)$ and eliminates the need for additional networks, see Eq.15.
- **UniGuide is a self-guiding method**: UniGuide modifies the reverse process of the unconditional diffusion model without relying on additional networks to guide the generation, unlike [1,2,3] that use additional differentiable networks for guidance. This highlights the flexibility of UniGuide as it does not require additional training: neither training a specialised conditional diffusion model nor training additional classifiers/regressors for guidance.
- **UniGuide is a gradient-based method**: In our experimental evaluations, we include other controlling techniques during inference, such as inpainting or position guidance.
- For SBDD, inpainting directly inpaints the noised condition to the generation, leading to a perturbed denoising process through the used hard constraints.
- For LBDD, ShapeMol+g directly leverages the positions of the reference ligands and adds additional position corrections to the conformations based on distance measures.
> Elaborate on the superiority of UniGuide
- **Unification** through the condition map: UniGuide leverages a novel condition map to incorporate diverse geometric conditions, such as protein structures, molecular fragments, and molecular surfaces. This extends unconditional base models to diverse drug discovery tasks. Beyond established drug discovery tasks, UniGuide enables tackling entirely novel tasks where neither baselines nor sufficient data exists, e.g. the tasks of density-based drug design or symmetric proteins. We kindly refer the reviewer to the general response and the figures of the PDF.
- **Equivariance**: Different from the text and image domains, equivariance is an important inductive bias of molecular generative modelling. In Theorem 4.1, we derive the conditions on UniGuide's condition map that enable guidance while respecting the molecular domain by maintaining equivariant updates.
- Empirically, our results show that through UniGuide's soft gradient-based updates, we achieve state-of-the-art performance by providing the best trade-off between how well we incorporate the condition into the generative process (LBDD: shape similarity and graph similarity, SBDD: Vina) and the overall quality of the molecules (QED, SA, validity). We kindly point the reviewer to Tables 2 & 3 in the PDF.
- Finally, we believe our follow-up response shows UniGuide's superiority over validity guidance. However, we are unsure if the reviewer has seen it. We provide results demonstrating that our condition map is more reliable and outperforms validity guidance regarding shape similarity, highlighting its effectiveness.
In summary, like classifier guidance, UniGuide controls the generation without additional training. In addition, UniGuide extends this approach [1] to the molecular domain by mapping the source condition to the configuration space. This enables direct self-guidance without additional networks while preserving equivariance. Our experimental evaluations show that UniGuide outperforms various baselines, including other techniques that control the generation solely during inference across multiple downstream tasks. We are thankful for the reviewer’s feedback and hope that our answers sufficiently clarify the remaining concern.
[1] Dhariwal et al. Diffusion models beat gans on image synthesis
[2] Equivariant Energy-Guided SDE for Inverse Molecular Design
[3] Training-free Multi-objective Diffusion Model for 3D Molecule Generation