Representations, Scoring, and Alignment

Representations, Scoring, and Alignment#

This page documents the mathematical formulations underlying ShEPhERD Score’s interaction profile representations, similarity scoring functions, and analytical gradient derivations used during alignment. Alignment has been driven by the automatic differentiation engines of Jax and PyTorch. However, since v1.3.1, analytical gradients have been implemented for PyTorch which significantly reduces memory overhead and speeds up computation; this is now the default path for PyTorch-based alignment in MoleculePair. We use the notation provided in Adams et al. (2025).

Base Representations#

  • Shape: Represented as a point cloud \(\boldsymbol{x}_2 = \boldsymbol{S}_2 \in \mathbb{R}^{n_2\times3}\) (or atomic coordinates \(\boldsymbol{C} \in \mathbb{R}^{n_1 \times 3}\)).

  • Electrostatic Potential (ESP): Represented as \(\boldsymbol{x}_3 = (\boldsymbol{S}_3, \boldsymbol{v})\) where \(\boldsymbol{S}_3 \in \mathbb{R}^{n_3 \times 3}\) is a surface point cloud and \(\boldsymbol{v} \in \mathbb{R}^{n_3}\) is the Coulombic potential at each point. The potential \(\boldsymbol{v}\) is defined by atomic partial charges \(\boldsymbol{q} \in \mathbb{R}^{n_1}\) and positions \(\boldsymbol{r}\):

\[\boldsymbol{v}[k] = \frac{1}{4 \pi \epsilon_0} \sum^{n_3}_{j=1} \frac{q[k]}{\|\boldsymbol{r}[k] - \boldsymbol{r}[j] \|^2}\]

  • Pharmacophores: Represented as \(\boldsymbol{x}_4 = (\boldsymbol{p}, \boldsymbol{P}, \boldsymbol{V})\).

    • \(\boldsymbol{p} \in \mathbb{R}^{n_4 \times N_p}\): one-hot encodings of \(N_p\) types.

    • \(\boldsymbol{P} \in \mathbb{R}^{n_4 \times 3}\): 3D coordinates.

    • \(\boldsymbol{V} \in \{ \mathbb{S}^2, \boldsymbol{0}\}^{n_4}\): relative unit vectors (\(|\boldsymbol{V}[k]| = 1\)) for directional pharmacophores, or zero vectors (\(|\boldsymbol{V}[k]| = 0\)) for directionless ones.

General Similarity and Alignment#

For any two point clouds \(\boldsymbol{Q}_A\) and \(\boldsymbol{Q}_B\) with points \(\boldsymbol{r}_k\), the first-order Gaussian overlap is:

\[O_{A,B} = \sum_{a \in \boldsymbol{Q}_A} \sum_{b \in \boldsymbol{Q}_B} w_{a,b} \left(\frac{\pi}{2 \alpha}\right)^{\frac{3}{2}} \exp{\left(-\frac{\alpha}{2}\|\boldsymbol{r}_a - \boldsymbol{r}_b\|^2\right)}\]

The Tanimoto similarity function is:

\[\text{sim}^{*}(\boldsymbol{Q}_A, \boldsymbol{Q}_B) = \frac{O_{A,B}}{O_{A,A} + O_{B,B} - O_{A,B}}\]

The optimal alignment objective is:

\[\text{sim}(\boldsymbol{Q}_A, \boldsymbol{Q}_B) = \max_{\boldsymbol{R, t}}{\text{sim}^{*} (\boldsymbol{R}\boldsymbol{Q}_A^T + \boldsymbol{t}, \boldsymbol{Q}_B)}\]

where \(\boldsymbol{R} \in SO(3)\) and \(\boldsymbol{t}\in T(3)\).

Specific Scoring Functions#

The variables \(w_{a,b}\) and \(\alpha\) from the general overlap function adapt based on the representation being scored:

Shape Scoring

  • Volumetric (using atoms \(\boldsymbol{C}\)): \(w_{a,b} = 2.7\), \(\alpha = 0.81\).

  • Surface (using points \(\boldsymbol{S}\)): \(w_{a,b} = 1\), \(\alpha = \Psi(n_2)\) (a predefined constant based on \(n_2\)).

ESP Scoring

The overlap function incorporates the potential difference:

\[O_{A,B}^{\text{ESP}} = \sum_{a \in \boldsymbol{Q}_A} \sum_{b \in \boldsymbol{Q}_B} \left(\frac{\pi}{2 \alpha}\right)^{\frac{3}{2}} \exp{\left(-\frac{\alpha}{2}\|\boldsymbol{r}_a - \boldsymbol{r}_b\|^2\right)} \exp{\left(-\frac{\|\boldsymbol{v}_{A}[a] - \boldsymbol{v}_{B}[b]\|^2}{\lambda}\right)}\]

where \(\alpha = \Psi(n_3)\) and \(\lambda = \frac{0.3}{(4 \pi \epsilon_0)^2}\).

Pharmacophore Scoring

Calculated per pharmacophore type \(m \in \mathcal{M}\) (where \(|\mathcal{M}| = N_p\)). The overlap for a specific type \(m\) is:

\[O^\text{pharm}_{A,B;m} = \sum_{a \in \boldsymbol{Q}_{A,m}} \sum_{b \in \boldsymbol{Q}_{B,m}} w_{a,b;m}\left(\frac{\pi}{2 \alpha_m}\right)^{\frac{3}{2}} \exp{\left(-\frac{\alpha_m}{2}\|\boldsymbol{r}_a - \boldsymbol{r}_b\|^2\right)}\]

The total similarity sums the overlaps across all types:

\[\text{sim}_{\text{pharm}}^{*}(\boldsymbol{x}_{4,A}, \boldsymbol{x}_{4,B}) = \frac{\sum_{m\in \mathcal{M}} O_{A,B; m}}{\sum_{m \in \mathcal{M}} O_{A,A; m} + O_{B,B; m} - O_{A,B; m}}\]

The vector weighting \(w_{a,b;m}\) depends on directionality (\(\alpha_m\) is a constant per type):

\[\begin{split}w_{a,b;m} = \begin{cases} 1 & \text{if } m \text{ is non-directional}, \\ \frac{\boldsymbol{V}[a]_{m}^\top \boldsymbol{V}[b]_{m} + 2}{3} & \text{if } m \text{ is directional}. \end{cases}\end{split}\]

(Note: For aromatic rings, the absolute value of the dot product \(|\boldsymbol{V}[a]_{m}^\top \boldsymbol{V}[b]_{m}|\) is used.)

Analytical Gradients#

Analytical gradients have been implemented for pharmacophore, shape, and shape-with-avoid alignment, replacing PyTorch autograd. All implementations live in shepherd_score/score/analytical_gradients/ (PyTorch in _torch.py, re-exported via __init__.py) and are called from the optimizer loops in shepherd_score/alignment/_torch_analytical.py. The resulting speedup is approximately 2–2.5× over autograd.

The SE(3) parameter vector is always (q_w, q_x, q_y, q_z, t_x, t_y, t_z) — 4 quaternion components followed by 3 translation components. All functions support both single (7,) and batched (B,7) inputs.

1. Tanimoto Chain Rule#

The self-overlaps \(O_{A,A}\) and \(O_{B,B}\) are invariant to any rigid SE(3) transformation. Defining \(U = O_{A,A} + O_{B,B}\), the Tanimoto gradient simplifies via the quotient rule:

\[\nabla_{\theta} S = \frac{U}{(U - O_{A,B})^2} \nabla_{\theta} O_{A,B}\]

\(U\) is computed once before the optimization loop. Only \(\nabla_{\theta} O_{A,B}\) needs to be evaluated each step. The loss is \(1 - S\).

1b. Tversky Chain Rule (pharmacophore only)#

For Tversky similarity, \(D = \sigma \cdot O_{A,A} + (1-\sigma) \cdot O_{B,B}\) is a constant w.r.t. SE(3). The similarity is:

\[\text{sim}^{*}(\boldsymbol{Q}_A, \boldsymbol{Q}_B) = \min\left(\frac{O_{A,B}}{D}, 1.0\right)\]

The gradient scales directly by a constant factor (no quotient rule):

\[\nabla_{\theta} S = -\frac{\mathbf{1}[S < 1.0]}{D} \nabla_{\theta} O_{A,B}\]

where the indicator function \(\mathbf{1}[S < 1.0]\) is 1 when not clamped, 0 otherwise. Sigma variants: 'tversky'→0.95, 'tversky_ref'→1.0, 'tversky_fit'→0.05.

2. Shape and ESP Gradients#

Let \(\boldsymbol{\Delta}_{ab} = \boldsymbol{R}\boldsymbol{r}_a + \boldsymbol{t} - \boldsymbol{r}_b\) and \(E_{ab} = \exp(-\frac{\alpha}{2}\|\boldsymbol{\Delta}_{ab}\|^2)\).

For shape, the pair constant is \(C_{ab} = w_{ab} \left(\frac{\pi}{2\alpha}\right)^{3/2}\) (with \(w_{ab}=1\) for surface, \(w_{ab}=2.7\) for volumetric). For ESP, the pair constant absorbs the potential difference: \(\tilde{C}_{ab} = \left(\frac{\pi}{2\alpha}\right)^{3/2} \exp\!\left(-\|\boldsymbol{v}_A[a]-\boldsymbol{v}_B[b]\|^2 / \lambda\right)\). The ESP potential is invariant to rigid motion, so \(\tilde{C}_{ab}\) is a pure constant during alignment.

Both cases share the same gradient structure (substituting \(C_{ab}\) or \(\tilde{C}_{ab}\)):

\[\nabla_{\boldsymbol{t}} O_{A,B} = \sum_{a,b} -\alpha\, C_{ab}\, E_{ab}\, \boldsymbol{\Delta}_{ab}\]
\[\nabla_{\boldsymbol{R}} O_{A,B} = \sum_{a,b} -\alpha\, C_{ab}\, E_{ab}\, (\boldsymbol{\Delta}_{ab}\, \boldsymbol{r}_a^\top)\]

3. Pharmacophore Gradients#

The pairwise term per pharmacophore type \(m\) is \(f_{ab} = w'_{ab;m}\, K_m\, E_{ab}\) where \(K_m = (\pi/(2\alpha_m))^{3/2}\) and \(E_{ab} = \exp(-\frac{\alpha_m}{2}\|\boldsymbol{P}'_a - \boldsymbol{P}_b\|^2)\).

Translation gradient (weight \(w'\) is rotation-only, so only \(E_{ab}\) contributes):

\[\nabla_{\boldsymbol{t}} O_{A,B} = \sum_{m} \sum_{a,b} -\alpha_m\, w'_{ab;m}\, K_m\, E_{ab}\, (\boldsymbol{P}'_a - \boldsymbol{P}_b)\]

Rotation matrix gradient (product rule over spatial and weight terms):

\[\nabla_{\boldsymbol{R}} O_{A,B} = \sum_{m} \sum_{a,b} K_m E_{ab} \left[ \nabla_{\boldsymbol{R}} w'_{ab;m} - \alpha_m\, w'_{ab;m}\, (\boldsymbol{P}'_a - \boldsymbol{P}_b)\boldsymbol{P}_a^\top \right]\]

The directional weight gradients are:

  • Non-directional (Hydrophobe, ZnBinder, Anion, Cation): \(\nabla_{\boldsymbol{R}} w' = \mathbf{0}\)

  • Directional (Acceptor, Donor, Halogen): \(w = (D_{ab}+2)/3 \implies \nabla_{\boldsymbol{R}} w' = \tfrac{1}{3}\boldsymbol{V}_b\boldsymbol{V}_a^\top\)

  • Aromatic: \(w = (|D_{ab}|+2)/3 \implies \nabla_{\boldsymbol{R}} w' = \tfrac{1}{3}\operatorname{sgn}(D_{ab})\boldsymbol{V}_b\boldsymbol{V}_a^\top\)

where \(D_{ab} = (\boldsymbol{R}\boldsymbol{V}_a)^\top \boldsymbol{V}_b\).

4. Avoid Penalty Gradients#

The linear hard-sphere avoid penalty is:

\[A = \sum_{a \in \text{fit\_avoid}} \sum_{b \in \text{avoid}} \operatorname{relu}\!\left(\frac{d_0 - \|\boldsymbol{P}'_a - \boldsymbol{P}_b\|}{d_0}\right)\]

where \(d_0\) is avoid_min_dist. The gradient is active only where \(0 < \|\boldsymbol{\Delta}_{ab}\| < d_0\):

\[\frac{\partial A}{\partial \boldsymbol{P}'_a} = -\frac{1}{d_0} \sum_b \mathbf{1}[0 < \|\boldsymbol{\Delta}_{ab}\| < d_0]\, \hat{\boldsymbol{\Delta}}_{ab}\]
\[\nabla_{\boldsymbol{t}} A = \sum_a \frac{\partial A}{\partial \boldsymbol{P}'_a}, \qquad \nabla_{\boldsymbol{R}} A = \left(\frac{\partial A}{\partial \boldsymbol{P}'}\right)^\top \boldsymbol{P}_{\text{orig}}\]

The combined loss for shape-with-avoid is \(\mathcal{L} = (1 - S_\text{shape}) + \lambda_\text{avoid} \cdot A\). Gradients add linearly before quaternion projection.

5. Quaternion Projection#

The \(3\times3\) rotation gradient \(\boldsymbol{G}\) is projected onto quaternion parameters via the Frobenius inner product with the quaternion Jacobians:

\[\frac{\partial \mathcal{L}}{\partial q_k} = \operatorname{Tr}\!\left(\boldsymbol{G}^\top \frac{\partial \boldsymbol{R}}{\partial q_k}\right)\]

The four Jacobians \(\partial\boldsymbol{R}/\partial q_k\) are linear functions of \(\boldsymbol{q}\), e.g.:

\[\begin{split}\frac{\partial \boldsymbol{R}}{\partial q_w} = \begin{bmatrix} 0 & -2q_z & 2q_y \\ 2q_z & 0 & -2q_x \\ -2q_y & 2q_x & 0 \end{bmatrix}\end{split}\]

Because the optimizer works with unnormalized raw quaternions \(\hat{\boldsymbol{q}}\), the chain rule for normalization \(\boldsymbol{q} = \hat{\boldsymbol{q}}/\|\hat{\boldsymbol{q}}\|\) is applied:

\[\nabla_{\hat{\boldsymbol{q}}} = \frac{1}{\|\hat{\boldsymbol{q}}\|}\left(\boldsymbol{I} - \boldsymbol{q}\boldsymbol{q}^\top\right)\nabla_{\boldsymbol{q}}\]

Key constants — score/constants.py#

  • P_ALPHAS: per-type \(\alpha_m\) (all 1.0 except aromatic/hydrophobe = 0.7)

  • P_TYPES: ('Acceptor', 'Donor', 'Aromatic', 'Hydrophobe', 'Halogen', 'Cation', 'Anion', 'ZnBinder', 'Dummy')

References#

Citation

If you use or adapt this work, please cite:

Adams, K., Abeywardane, K., Fromer, J., & Coley, C. W. (2025). ShEPhERD: Diffusing shape, electrostatics, and pharmacophores for bioisosteric drug design. In The Thirteenth International Conference on Learning Representations. https://openreview.net/forum?id=KSLkFYHlYg

Grant, J. A. & Pickup, B. T. (1995). A Gaussian description of molecular shape. Journal of Physical Chemistry, 99(11), 3503-3510. https://doi.org/10.1021/j100011a016

Grant, J. A., Gallardo, M. A., & Pickup, B. T. (1996). A fast method of molecular shape comparison: A simple application of a Gaussian description of molecular shape. Journal of Computational Chemistry, 17(14), 1653–1666. https://doi.org/10.1002/(SICI)1096-987X(19961115)17:14<1653::AID-JCC7>3.0.CO;2-K

Grant, J. A., & Pickup, B. T. (1997). Gaussian shape methods. In W. F. van Gunsteren, P. K. Weiner, & A. J. Wilkinson (Eds.), Computer simulations of biomolecular systems (Vol. 3, pp. 150–176). Springer. https://doi.org/10.1007/978-94-017-1120-3_5

Taminau, J., Thijs, G., & De Winter, H. (2008). Pharao: Pharmacophore alignment and optimization. Journal of Molecular Graphics and Modelling, 27(2), 161–169. https://doi.org/10.1016/j.jmgm.2008.04.003

Wahl, J. (2024). PheSA: An Open-Source Tool for Pharmacophore-Enhanced Shape Alignment. Journal of Chemical Information and Modeling, 64(15), 5944-5953. https://doi.org/10.1021/acs.jcim.4c00516