Fig. 4. Distinguishing sensors from binders.

(A) Nonbinders do not bind to membranes; sensors only bind to curved membranes; binders bind to any membrane, independent of curvature. (B) ΔΔF_adj for 19 benchmark peptides (table S4) predicted with the CNN model (top) and calculated from MD simulations (bottom). (C) Linear correlation between CNN-predicted ΔΔF_adj and the flat membrane binding free energy ΔF_sm(R = ∞). Circles indicate peptides for which ΔF_sm(R = ∞) was calculated by thermodynamic integration (section S10). The Evo-MD optimum (black) is the sequence in Fig. 2D. For the remaining peptides (crosses), ΔF_sm was derived from the linear fit ΔF_sm = 3.83 × ΔΔF_adj + 12.27. (D) Highlighted ΔΔF_adj values for ALPS GMAP-210 variants. (E) The membrane binding probability P_m as a function of the membrane binding free energy ΔF_sm (section S10 and Eq. 2), at a vesicle radius R = 50. In (D) and (E), the orange area indicates the likely regime (0.05 ≤ P_m ≤ 0.95) where peptides are empirically classified as sensors. (F) Membrane binding probability P_m as a function of the vesicle radius R for a range of sensing free energies ΔΔF (section S10).