Unified Biomolecular Trajectory Generation via Pretrained Variational Bridge
Ziyang Yu, Wenbing Huang, Yang Liu
TL;DR
This paper introduces the Pretrained Variational Bridge, a novel deep generative model that efficiently produces molecular trajectories by unifying training on diverse data and incorporating reinforcement learning for protein-ligand complexes.
Contribution
The paper presents a unified framework for molecular trajectory generation that leverages cross-domain structural knowledge and reinforcement learning, improving fidelity and efficiency over prior models.
Findings
Faithfully reproduces thermodynamic and kinetic observables.
Provides stable and efficient generative dynamics.
Supports post-optimization of docking poses.
Abstract
Molecular Dynamics (MD) simulations provide a fundamental tool for characterizing molecular behavior at full atomic resolution, but their applicability is severely constrained by the computational cost. To address this, a surge of deep generative models has recently emerged to learn dynamics at coarsened timesteps for efficient trajectory generation, yet they either generalize poorly across systems or, due to limited molecular diversity of trajectory data, fail to fully exploit structural information to improve generative fidelity. Here, we present the Pretrained Variational Bridge (PVB) in an encoder-decoder fashion, which maps the initial structure into a noised latent space and transports it toward stage-specific targets through augmented bridge matching. This unifies training on both single-structure and paired trajectory data, enabling consistent use of cross-domain structural…
Peer Reviews
Decision·ICLR 2026 Poster
The idea of pretraining a bridge to recapitulate the initial state, and then fine-tuning it to produce the evolved state, is quite interesting. The work also touches upon simulation of protein-ligand simulations, which have been somewhat neglected in the ML for MD literature, despite their significant practical importance.
**Method** * The RL formulation of the holo complex finetuning task seems gratuitous. In particular, if the reward is the RMSD to the holo state, why can't the holo state be used in a supervised fine-tuning fashion? If would seem that if the reward is simply the similarity to an explicit, known state, that is the setting of supervised learning, not reinforcement learning. **Experiments** * There are missing controls that make the value of the pretraining bridge hard to interpret. What if we pre
- The paper proposes a novel methodology to include pretraining on datasets with static structures but diverse chemical space, and then finetuning on dynamical data. This enables the model to achieve better chemical transferability despite the limited chemical space coverage of the dynamical data. - The integration of RL with adjoint matching for pose-optimization in docking is a novel application. And the authors have shown the improvement in the ligand pose after the finetuning. - The model's
- While the paper evaluates against other trajectory-based models, it assesses performance on free energy landscapes. While most of the metrics compared in the paper are actually thermodynamic properties, they can be evaluated with i.i.d. (time-agnostic) sampling models. It will be helpful to benchmark against those methods as well. - In the meanwhile, although the time-dependent model describes dynamics, it's not obvious from the benchmarks and applications shown in the paper why the time-depen
- The paper works on a relevant problem. - Optimal control methods are leveraged for more efficient training. - The method is evaluated on relevant benchmarks.
- The explanation of the theory and the notation is quite confusing. I sympathize that this is not trivial, especially with having in a mind a relatively broad target audience from diverse research backgrounds. But considerable effort should be made to improve the writing. I try to make some concrete suggestions below. I'm certainly willing to raise my score if readability is improved!
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsMachine Learning in Materials Science · Protein Structure and Dynamics · Computational Drug Discovery Methods