Answers are grounded in this paper and cite their sources.
To bridge the fidelity gap between high-fidelity quantum mechanics (QM) calculations and low-cost, noisy molecular dynamics (MD) trajectories, this paper proposes a dynamic uncertainty-weighted multigranularity coattention framework (DUW-MGCA), which jointly models QM/MD multimodal information within an SE(3) equivariant graph neural network. This framework adaptively adjusts the weight of the two modalities using heteroscedastic uncertainty and soft gating and improves the physical realism of the energy landscape through weak physical regularization constraints. Experimental results show that DUW-MGCA significantly outperforms existing baselines on the HiQBind-MISATO data set, achieving RMSE, mean absolute error, and Pearson R values of 0.979 kcal/mol, 0.823 kcal/mol, and 0.931, respectively, and achieving lower calibration metrics with expected calibration error (0.475) and negative log-likelihood (1.398). The method achieves an inference time of about 1 min per complex on a single GPU, balancing high accuracy with acceptable computational overhead.
No takes yet. Share an insight, caveat, or question.
Tan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: