Answers are grounded in this paper and cite their sources.
Flow-based generative models learn to transform random noise into molecular conformations by following a prescribed interpolation path between the two. Straight-line (linear) interpolation is the usual choice, but it is only one of many possible paths. We compare four paths within a single affine framework, x t = α t x+ β t z : linear, trigonometric, spherical and a new operator-based path in which each degree of freedom follows its own schedule. Linear interpolation is a limiting case of both the operator-based and spherical paths. The operator-based path bends inside it (contracted), whereas the trigonometric and spherical paths bend outside it (expanded). We train the same Transformer velocity field with each path on a three-well potential, the mini-protein Trp-cage and the intrinsically disordered protein α-synuclein. All four paths reproduce the reference distributions with comparable accuracy, but they get there differently. The trigonometric and spherical paths resolve conformational basins early and approach the reference smoothly. The operator-based path delays part of the transport, so some features stall or briefly diverge before converging sharply. An exact relation between the rate of change of the KL divergence and the alignment of the learned velocity with the score difference, together with closed-form moments of each path, explains these differences. The choice of interpolation path therefore controls how a flow model converges, not only where it ends.
No takes yet. Share an insight, caveat, or question.
E et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: