Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Efficient Protein Optimization via Structure-aware Hamiltonian Dynamics

Jiahao Wang, Shuangjia Zheng

TL;DR

Protein design in large sequence spaces with rugged epistasis is addressed by HADES, which combines structure-aware Bayesian optimization with Hamiltonian dynamics to sample from a continuous state $q$ (discretized to a one-hot sequence) and a momentum $p$. The method uses a potential $U(q)=-\log P(f(q))$ and kinetic energy $K(p)=\frac{1}{2m}\|p\|^2$, with leapfrog integration and a virtual barrier to enforce valid probability vectors, plus Metropolis corrections. A two-stage surrogate learns structure perturbations (RMSD) via ESMFold priors and fitness in a joint encoder-decoder framework, enabling a smooth, structure-informed landscape for sampling. In silico benchmarks on GB1 and PhoQ show HADES achieving higher cumulative max fitness and diversity than baselines, validating the benefit of structure-aware priors and uncertainty-aware Hamiltonian sampling for efficient exploration and design. The approach promises substantial reductions in wet-lab cost while enabling design of proteins with similar structures and enhanced properties.

Abstract

The ability to engineer optimized protein variants has transformative potential for biotechnology and medicine. Prior sequence-based optimization methods struggle with the high-dimensional complexities due to the epistasis effect and the disregard for structural constraints. To address this, we propose HADES, a Bayesian optimization method utilizing Hamiltonian dynamics to efficiently sample from a structure-aware approximated posterior. Leveraging momentum and uncertainty in the simulated physical movements, HADES enables rapid transition of proposals toward promising areas. A position discretization procedure is introduced to propose discrete protein sequences from such a continuous state system. The posterior surrogate is powered by a two-stage encoder-decoder framework to determine the structure and function relationships between mutant neighbors, consequently learning a smoothed landscape to sample from. Extensive experiments demonstrate that our method outperforms state-of-the-art baselines in in-silico evaluations across most metrics. Remarkably, our approach offers a unique advantage by leveraging the mutual constraints between protein structure and sequence, facilitating the design of protein sequences with similar structures and optimized properties. The code and data are publicly available at https://github.com/GENTEL-lab/HADES.

Efficient Protein Optimization via Structure-aware Hamiltonian Dynamics

TL;DR

Protein design in large sequence spaces with rugged epistasis is addressed by HADES, which combines structure-aware Bayesian optimization with Hamiltonian dynamics to sample from a continuous state (discretized to a one-hot sequence) and a momentum . The method uses a potential and kinetic energy , with leapfrog integration and a virtual barrier to enforce valid probability vectors, plus Metropolis corrections. A two-stage surrogate learns structure perturbations (RMSD) via ESMFold priors and fitness in a joint encoder-decoder framework, enabling a smooth, structure-informed landscape for sampling. In silico benchmarks on GB1 and PhoQ show HADES achieving higher cumulative max fitness and diversity than baselines, validating the benefit of structure-aware priors and uncertainty-aware Hamiltonian sampling for efficient exploration and design. The approach promises substantial reductions in wet-lab cost while enabling design of proteins with similar structures and enhanced properties.

Abstract

The ability to engineer optimized protein variants has transformative potential for biotechnology and medicine. Prior sequence-based optimization methods struggle with the high-dimensional complexities due to the epistasis effect and the disregard for structural constraints. To address this, we propose HADES, a Bayesian optimization method utilizing Hamiltonian dynamics to efficiently sample from a structure-aware approximated posterior. Leveraging momentum and uncertainty in the simulated physical movements, HADES enables rapid transition of proposals toward promising areas. A position discretization procedure is introduced to propose discrete protein sequences from such a continuous state system. The posterior surrogate is powered by a two-stage encoder-decoder framework to determine the structure and function relationships between mutant neighbors, consequently learning a smoothed landscape to sample from. Extensive experiments demonstrate that our method outperforms state-of-the-art baselines in in-silico evaluations across most metrics. Remarkably, our approach offers a unique advantage by leveraging the mutual constraints between protein structure and sequence, facilitating the design of protein sequences with similar structures and optimized properties. The code and data are publicly available at https://github.com/GENTEL-lab/HADES.
Paper Structure (11 sections, 6 equations, 5 figures, 3 tables, 2 algorithms)

This paper contains 11 sections, 6 equations, 5 figures, 3 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Background
  4. Method
  5. Sampling with Hamiltonian Dynamics
  6. Two-stage Surrogate Model Learning
  7. Experiments
  8. Experimental Setup
  9. Results and Analysis
  10. Conclusion and Future Work
  11. Acknowledgments

Figures (5)

  • Figure 1: The overall architecture of the proposed HADES framework.
  • Figure 2: Virtual barrier and position discretization.
  • Figure 3: Cumulative maximum fitness scores by experiment rounds on two protein engineering benchmark tasks. Green triangles represent the mean values.
  • Figure 4: Cumulative maximum fitness scores for K={16, 32, 64, 128} on two protein engineering benchmark tasks. All curves are derived from 10 runs with random network initialization. The shaded regions represent the standard deviation.
  • Figure 5: High-fitness structures sampled from EvoPlay and HADES on GB1 task. The RMSD* score (the lower the better) calculates the root mean square deviation over the 4 mutation sites.