Fig. 5: Graph neural network model for DNA folding thermodynamics.
From: High-throughput DNA melt measurements enable improved models of DNA folding thermodynamics
a Schematic representation of the graph neural network (GNN) architecture. b Mean absolute error (MAE) of melting temperature on Array Melt data as a function of the number of graph convolution layers in the GNN. c MAE of ∆G37 on Array Melt data as a function of the percentage of training data used to fit the GNN. d Pearson’s R correlation coefficient between GNN predictions and held-out Array Melt data, literature UV melting data, or Oliveira et al. duplex melting data. e Scatter plot comparing GNN predictions with held-out Array Melt test data, both combined and split by variant class. Each dot represents a single sequence variant. f Scatter plot comparing GNN predictions with literature UV melting data or Oliveira et al. duplex melting data. g Pearson’s R between principal components of the learned variant embeddings and measured thermodynamic parameters or variant properties. h Comparison between the most correlated principal components and ∆H or Tm. i UMAP visualization of learned variant embeddings after aggregation, colored by dataset (Array Melt, literature UV melting, or Oliveira et al. duplex melting data, color scheme the same as in (d–f)). j Benchmark of all models on held-out Array Melt data, compared to the measurement error of Array Melt. k Benchmark of all models on the orthogonal Oliveira et al. dataset, plotted by the number of mismatches. Colors correspond to models as in Fig. 5j.
