Overview
Flow matching with stochastic splits and deletions.
Branching Flows augments a base Markov generator with split and deletion events. A forest of binary trees couples roots sampled from a simple initial distribution to leaves from the data distribution.
Along each branch, elements evolve according to the chosen base process. At split events an element is duplicated in place; at deletion events it is removed. The model learns the base generator together with the split and deletion rates.
QM9
Small molecules.
QM9 provides a benchmark for multimodal (continuous atom positions and discrete atom types) variable-length distribution learning.
Proteins
Branching Flows finetunes into protein models.
We finetune split and deletion heads into ChainStorm (Oresten et al., 2025), an SE(3) equivariant Protein backbone model over positions, rotations, and residue labels. The base model exhibits spontaneous chain symmetry, and this is preserved in the Branching Flows finetune.
Plausible backbone geometry
Unknown-length infix sampling
This allows completing parts of proteins without knowing a priori the segment length.
New results: Branching La Proteina
Branching Flows finetunes into La Proteina.
La Proteina (Geffner et al., 2025) uses a latent side chain representation for all-atom structure generation. Branching Flows finetunes retain exceptionally high designability. Examples below from a finetune of the LD3 variant.
Resources