Fig. 2.

Past–future information bottleneck framework results on alanine dipeptide. a Unbiased simulation trajectory for dihedral angles Φ and Ψ, in blue triangles and orange pluses, respectively. The alanine dipeptide molecule is shown in inset. b Absolute weights for different order parameters in the first training round as a function of the predictive time delay Δt. Blue triangles, orange triangles, green circles, and red squares correspond to cos ϕ, sin ϕ, cos ψ, and sin ψ, respectively. c–f Free energy along the adaptively learnt reaction coordinate (RC) along with the corresponding biased trajectories for different training rounds. g Free energy along ϕ, ψ after the RC has converged with energy contours every 4 kJ mol⁻¹. h Kinetics from the post-training biased runs as well as reference unbiased runs. The two are essentially indistinguishable. Orange and blue dashed lines denote unbiased data, respectively, while red and black solid lines show corresponding best fits