EvoEGF-Mol: Geodesic Flow in SBDD

• The paper demonstrates how EvoEGF-Mol uses composite exponential-family distributions to unify continuous atomic coordinates and discrete chemical types via exponential geodesics.
• It introduces dynamic endpoint scheduling and progressive-parameter refinement, which stabilize training by preserving gradient information and ensuring high-quality molecular samples.
• Benchmark results reveal that EvoEGF-Mol outperforms traditional methods in terms of geometric precision, binding affinity, and scaffold recovery in SBDD applications.

Evolving Exponential Geodesic Flow for SBDD (EvoEGF-Mol) is a generative modeling framework for structure-based drug design (SBDD) that constructs probability flows in the space of molecular structures via information-geometric principles. The method addresses the mismatch present in prior approaches between probabilistic modeling of continuous atomic coordinates and discrete chemical categories by representing molecules as composite exponential-family distributions and defining generative trajectories as exponential geodesics under the Fisher–Rao metric. EvoEGF-Mol leverages dynamically concentrating endpoint distributions, progressive refinement, and a unified probabilistic manifold approach, achieving high geometric fidelity and target specificity, as demonstrated on major molecular design and docking benchmarks (Jin et al., 30 Jan 2026).

1. Composite Exponential-Family Representation of Molecules

EvoEGF-Mol treats a molecule $M$ as a single probabilistic object, whose joint distribution over atomic coordinates $X\in\mathbb R^{N\times 3}$, atom types $V$, and bond types $E$ factorizes multiplicatively into exponential-family densities. The general exponential-family form is

$$p(x;\,\theta) = h(x) \exp\left(\theta^\top T(x) - A(\theta)\right)$$

where $\theta$ are natural parameters, $T(x)$ are sufficient statistics, $A(\theta)$ is the log-partition function, and $h(x)$ is the base measure.

• Continuous atomic coordinates: Modeled with independent isotropic Gaussian factors for each atom:

$$p_{\rm coord}(X\mid\mu,\sigma^2) = \prod_{i=1}^N \frac{1}{(2\pi\sigma^2)^{3/2}} \exp\left(-\frac{\|x_i-\mu_i\|^2}{2\sigma^2}\right)$$

with natural parameters $\eta_1 = \mu/\sigma^2, \eta_2 = -1/(2\sigma^2)$.

• Discrete atom types: Modeled using the Dirichlet-multinomial relaxation; for atom $i$:

$$p_{\rm atom}(v\mid\alpha) = \frac{1}{B(\alpha)} \prod_{k=1}^{K_a} v_k^{\alpha_k - 1},\quad v\in\Delta^{K_a-1}$$

A one-hot type is approximated by making a particular $\alpha_k$ large.

• Bond types: Treated analogously to atom types.

The overall molecular distribution is: $$p(M\mid\theta) = p_{\rm coord}(X\mid\mu,\sigma^2) \times p_{\rm atom}(V\mid\alpha^{\rm atom}) \times p_{\rm bond}(E\mid\alpha^{\rm bond})$$ ensuring coverage of both continuous and categorical aspects within a unified exponential-family product manifold.

2. Exponential-Geodesic Flow and the Fisher–Rao Metric

On the product exponential-family manifold, the Fisher–Rao metric is

$$I(\theta) = \mathbb E_{p(\cdot;\theta)}[\nabla_\theta \log p\, \nabla_\theta \log p^\top] = \nabla^2 A(\theta)$$

The exponential-geodesic (e-geodesic) between distributions $p_0 = p(\cdot;\theta_0)$ and $p_1 = p(\cdot;\theta_1)$ is a linear path in natural-parameter space: $$\theta(t) = (1-t)\theta_0 + t\theta_1,\quad t\in[0,1]$$ This trajectory corresponds to the solution of the information-geometric gradient flow ODE: $$\frac{d\theta(t)}{dt} = -I(\theta(t))^{-1} \nabla_\theta D\big(p(\cdot;\theta(t))\|q\big)$$ where $D(\cdot\|\cdot)$ is the Kullback–Leibler divergence, and for $q = p(\cdot;\theta_1)$, the gradient reduces to $I(\theta)(\theta-\theta_1)$.

This construction ensures geodesics respect the geometry induced by the underlying probabilistic manifold, unifying both continuous and discrete factors within a single information-geometric framework.

3. Evolving Endpoints to Prevent Instantaneous Trajectory Collapse

Directly targeting Dirac distributions as endpoints results in instantaneous variance collapse of the geodesic for $t>0$, collapsing all learning signal near $t=1$. EvoEGF-Mol circumvents this by replacing static Dirac targets with dynamically concentrating endpoints:

• For Gaussians (coordinates):

$$\sigma_1(t) = \sigma_{\max}(1-t) + \varepsilon,\quad \varepsilon\ll1$$

with instantaneous variance schedule for the geodesic:

$$\sigma_t^2 = \left(\frac{1-t}{\sigma_0^2} + \frac{t}{\sigma_1(t)^2}\right)^{-1}$$

• For Dirichlet variables (types/bonds): Given $\alpha_0 = \mathbf{1}$ (uniform), the endpoint for target category $k^*$ evolves as:

$$\alpha_1(t) = \left[1-\rho(1-t)\right]e_{k^*} + \rho(1-t)\frac{1}{K}\mathbf{1},\quad \rho\in(0,1)$$

and

$\alpha_t = (1-t)\alpha_0 + t\,\alpha_1(t)$

The time-dependent target parameter $\theta_1(t)$ yields a generative trajectory in parameter space: $\theta(t) = (1-t)\theta_0 + t\,\theta_1(t)$ with

$$\dot\theta(t) = (\theta_1(t) - \theta_0) + t\,\dot\theta_1(t)$$

This scheduling stably guides generative flows towards sharply concentrated distributions without losing gradient information prematurely.

4. Progressive-Parameter-Refinement Architecture

The framework employs a progressive-parameter-refinement network, leveraging a Bayesian-Flow-Network-style sampler. For each of $n$ discrete time steps:

1. Given $\theta(t)$, intermediate “noisy” samples $m_t \sim p(\cdot;\theta(t))$ are drawn.
2. The tuple $(m_t,\,t,\,{\rm protein\ pocket})$ is processed by a unitransformer (4 layers, hidden size 128, gated attention, kNN=32).
3. The terminal natural parameters $\hat\theta_1$ are predicted.
4. The next-step parameters are set by

$$\theta(t+\Delta t) = (1-(t+\Delta t))\,\theta_0 + (t+\Delta t)\,\hat\theta_1$$

By updating via progressive refinement, the network avoids direct inversion of increasingly ill-conditioned variance parameters, providing stable learning and high-quality sample generation.

5. Loss Functions and Optimization

All loss functions derive from the local Kullback–Leibler divergence between evolving and predicted distributions along the geodesic:

• Coordinates (Gaussian):

$$\mathcal L_x = \mathbb E_{t,X} \left[\,\frac{1}{2\,\sigma_t^2}\|\mu_1 - \hat\mu_1\|^2\,\right]$$

• Types/Bonds (Dirichlet):

$$\mathcal L_{\rm type} = D_{\rm KL}(\mathrm{Dir}(\alpha_t)\|\mathrm{Dir}(\hat\alpha_t)) = \ln\frac{B(\hat\alpha_t)}{B(\alpha_t)} + \sum_k (\alpha_{t,k} - \hat\alpha_{t,k})\left(\psi(\alpha_{t,k}) - \psi(\sum_j\alpha_{t,j})\right)$$

• Total Loss:

$$\mathcal L = \mathcal L_x + \mathcal L_{\rm atom} + \mathcal L_{\rm bond}$$

This principled local-KL supervision ensures information-geometric consistency between generated molecules and targeted distributions across all domains.

6. Empirical Performance and Benchmark Analysis

On CrossDock2020 for de novo pose generation, EvoEGF-Mol approaches reference-level geometric precision and interaction fidelity:

• PoseBusters passing rate: 93.4% (baseline: MolCRAFT 84.6%, AR 59.0%, Pocket2Mol 72.3%)
• Binding affinity (AutoDock Vina):
  • Vina Score avg/median: -6.14 / -6.89 (MolCRAFT: -6.55 / -6.95)
  • Vina Min avg/median: -6.98 / -7.12 (best across baselines)
  • Vina Dock avg/median: -7.72 / -7.88 (best across baselines)
• Strain-energy (25/50/75-percentiles): 8.94 / 25.96 / 56.65 (substantially lower than competitors)
• Chemical properties: QED = 0.53, SA = 0.75, Clash-ratio = 0.24

On MolGenBench (scaffold-level, 120 targets, 1000 samples/target):

• Scaffold pass (MedChem filter): 31.79% (MolCRAFT: 23.07%)
• Scaffold Hit Recovery (In-distribution): 29.61% (MolCRAFT: 23.65%)
• Hit Fraction: 1.08% (MolCRAFT: 0.70%)
• Target-Aware Score (TAScore): higher than all baselines

On molecule-level MedChem filters:

• Pass Rate: 37.52% (MolCRAFT: 25.93%)
• Hit Recovery (In): 7 targets (0 for most diffusion methods)
• Hit Fraction: 0.03% (MolCRAFT: 0.01%)

EvoEGF-Mol thus achieves state-of-the-art geometric plausibility, binding affinity, and filtering compliance, markedly outperforming autoregressive, diffusion, and Bayesian-flow-based SBDD baselines (Jin et al., 30 Jan 2026).

7. Significance and Methodological Distinctions

EvoEGF-Mol introduces a unified information-geometric approach to molecular generative modeling for SBDD, seamlessly bridging continuous atomic and categorical chemical spaces within a single exponential-family framework. The core innovation of evolving exponential geodesic flow permits stable training, precise generative targeting, and avoidance of gradient collapse—limitations that challenge traditional methods that separately treat Euclidean and probability simplex variables or naively impose endpoint Dirac constraints. By integrating progressive-parameter refinement and dynamic endpoint scheduling, EvoEGF-Mol enhances sample quality and target conditioning, which is substantiated by high geometric validity, improved pharmacological profiles, and effective scaffold recovery benchmarks. The method represents a significant advancement in SBDD generative methodology, providing a foundation for further information-geometric modeling applications within scientific generative modeling (Jin et al., 30 Jan 2026).