Figure 2. Comparison of experimental, structural, and model-predicted 1E6 responses to ALW peptide substitutions.

(a) Experimental TNF release for all 190 single–amino–acid variants. (b) RACER-m predictions using the baseline 69-structure training set, capturing the central hot spot (positions outlined in red) but showing weaker agreement at flanking and MHC-facing positions. (c) Correlation between experimental TNF responses in panel (a) and RACER-m Δz predictions in panel (b). Each point represents a single peptide substitution, colored by peptide position. The red dashed lines denote the experimental weak-activity threshold (vertical) and the WT z-baseline (horizontal). The fitted regression line with 95% confidence band illustrates the modest global monotonic trend between experimental activity and predicted binding score.