Symmetry-Constrained Gaussian Processes for Sample-Efficient Molecular Property Prediction

Kaustubh S. Bukkapatnam, Siddharth Karuturi, Laksh PatelAI4Science @ ICML 2026 · OpenReview

TL;DR

SYMGP is a new Gaussian Process framework that bakes physical symmetries—like rotation, translation, and atom-swapping—directly into its mathematical core. By enforcing these constraints, the model achieves the same accuracy as state-of-the-art deep learning models using up to $5\times$ less data, all while providing the reliable uncertainty estimates needed for autonomous lab discovery.

What Problem Does This Paper Address?

In the race to find better drug candidates or more efficient solar cell materials, data is the primary bottleneck. Calculating a molecule's properties using quantum chemistry (like DFT) or measuring them in a wet lab is incredibly expensive and time-consuming.

To speed this up, researchers use Active Learning (AL): an iterative loop where an AI model predicts properties, identifies where it is most uncertain, and asks for a new "label" (a lab test or simulation) for the most informative candidate. However, standard models often "waste" data learning basic physics. For example, a standard model might not realize that a molecule rotated 90 degrees or with its identical carbon atoms swapped is actually the same molecule.

While deep learning models can be made "symmetry-aware," they often struggle to provide the mathematically rigorous uncertainty estimates required for effective Active Learning. This paper bridges that gap by bringing physical symmetry to Gaussian Processes (GPs), the gold standard for uncertainty-quantification in small-data regimes.

Methods

The authors introduce SYMGP, which uses a technique called Orbit Averaging to construct a symmetry-constrained kernel ($k_{Sym}$).

  1. Symmetry Groups: The model enforces invariance under the Euclidean group $E(3)$ (rotations, reflections, translations) and the permutation group $S_N$ (swapping identical atoms).
  2. Kernel Construction: Instead of a standard kernel, they average a base kernel over the "orbit" of the symmetry group. Mathematically, this ensures that $f(X) = f(g \cdot X)$ for any symmetry operation $g$.
  3. Theoretical Proofs: The authors prove that this isn't just a heuristic; it fundamentally reduces the "effective dimension" ($d_{eff}$) of the function space the model has to search. They provide a new sublinear regret bound, proving that SYMGP converges to the correct answer faster than symmetry-unaware models.
  4. Practical Implementation: To avoid the high computational cost of literal averaging, they use invariant descriptors—specifically, the sorted eigenvalues of the interatomic distance matrix—which achieve the same effect with much higher efficiency ($O(N^2 \log N)$).

Datasets

The framework was evaluated on two major benchmarks in computational chemistry:

  • QM9: A dataset of ~134k small organic molecules. The task was predicting the HOMO-LUMO gap, a key electronic property for materials science.
  • FreeSolv: A smaller, more challenging dataset of 642 molecules used to predict hydration free energy. This served as a "low-data" stress test for the model's efficiency.

Key Insights and Ideas

  • Symmetry as Dimension Reduction: The most significant "aha moment" is the formal proof that enforcing symmetry is equivalent to reducing the complexity of the Reproducing Kernel Hilbert Space (RKHS). By shrinking the "search space," the model gains more information from every single labeled example.
  • The Cost of Symmetry Violations: The authors show that when a model isn't symmetry-aware, it suffers from "epistemic inflation"—it remains uncertain about things it should already know, which destabilizes the active learning process.
  • Calibration Matters: Because SYMGP respects the underlying physics, its uncertainty estimates are "well-calibrated." When the model says it is 90% sure, it is actually right 90% of the time, which is critical for trusting autonomous discovery systems.

Key Results

  • $5\times$ Data Efficiency: On the QM9 dataset, SYMGP reached an accuracy level (0.108 eV RMSE) with only 800 labeled molecules that standard baselines (like GAUCHE-WL) required roughly 1,600 to 3,200 molecules to achieve.
  • Outperforming Baselines: SYMGP systematically outperformed both standard RBF kernels and specialized graph kernels (GAUCHE) across all tested data sizes.
  • Ablation Success: The researchers found that permutation invariance (atom swapping) provided the biggest boost in performance (+24.9% error reduction), followed by rotation and reflection.
  • Near-Perfect Calibration: SYMGP achieved an Expected Calibration Error (ECE) of 0.021, significantly better than the 0.089 seen in standard GP-RBF models.

Limitations and Future Work

  • Computational Scaling: Like all exact Gaussian Processes, SYMGP scales at $O(n^3)$ relative to the number of labeled points. This limits its use to datasets of roughly 5,000 points or fewer before it becomes computationally sluggish.
  • Approximations: The authors suggest that future work could integrate Sparse GP approximations (like inducing points) to scale the method to much larger datasets.
  • Periodic Symmetries: The current model handles isolated molecules. Extending this to "space groups" (periodic symmetries) would allow it to be used for crystal structure prediction and solid-state materials discovery.

Summary updated

Abstract

Molecular property prediction from limited labeled data is a central bottleneck in computational chemistry and materials discovery. We introduce SYMGP, a Gaussian process framework whose kernel provably enforces the physical symmetries of molecular property functions: permutation invariance over atom indices, and invariance under the Euclidean group $E(3)$ of rotations, reflections, and translations. We show that symmetry-averaging reduces the effective dimension of the reproducing kernel Hilbert space, yielding tighter information-gain bounds and sublinear regret guarantees for the resulting active learning acquisition strategy. Concretely, we prove that the maximum information gain $\gamma_T$ for SYMGP with a squared-exponential base kernel scales as $\mathcal{O}((\log T)^{d_{\text{eff}}+1})$ where $d_{\text{eff}} < d_{\text{ambient}}$, giving a provably faster convergence rate than a symmetry-unaware baseline. Experiments on QM9 and FreeSolv demonstrate that SYMGP matches the accuracy of fully supervised deep models using up to $5\times$ fewer labeled examples, while producing well-calibrated predictive uncertainties that are required for closed-loop autonomous discovery pipelines.
BibTeX
@inproceedings{bukkapatnam2026symmetry,
  author = {Kaustubh S. Bukkapatnam and Siddharth Karuturi and Laksh Patel},
  title = {{Symmetry-Constrained Gaussian Processes for Sample-Efficient Molecular Property Prediction}},
  booktitle = {AI4Science @ ICML},
  year = {2026},
  url = {https://openreview.net/forum?id=XNb0fwIGGJ},
}
Figure