Fig. 3: Main architecture of OmniMol and performance on chirality-aware tasks. | Nature Communications

Fig. 3: Main architecture of OmniMol and performance on chirality-aware tasks.

From: Unified and explainable molecular representation learning for imperfectly annotated data from the hypergraph view

Fig. 3

a The overall architecture of OmniMol. b The core component, OmniMol block, consists of multiple OmniMol Layers, interspersed with the positional update blocks. c Each OmniMol layer incorporates a task-routed Mixture of Experts (MoE) layer, facilitating task adaptivity. d In chirality-aware tasks, OmniMol surpassed previous methodologies. Abbreviations and symbols: FFN = FeedForward Neural Network, MAE = Mean Absolute Error, μ = mean value, σ = standard deviation, BCD = best chiral descriptors, AP = AtomPairs Fingerprints, LR = logistic regression, RF = random forest, GB = gradient boosting, XGBoost = Extreme Gradient Boosting, SVR = support vector regression. Other reference methods: SchNet42, MPNN60, AttentiveFP58, DRFormer17.

Back to article page