DiFfER: Discrete Flows for Enzymatic Reaction Mechanism Step Prediction

Enzymatic reaction mechanisms specify the electron movements that convert substrates into products, yet sparse annotations limit mechanism-step prediction. Current flow matching-bsed methods operate on continuous bond-electron (BE) matrices, producing fractional electron states that require rounding. To mitigate this, we introduce DiFfER, a Discrete Flow matching model for Enzymatic Reaction mechanism-step prediction that directly generates sparse, integer-valued BE matrix updates while enforcing symmetry by construction. We extend the representation to radical and metal-ion chemistry and curate atom-mapped trajectories for nine radical mechanisms. On 185 held-out enzymatic steps, DiFfER improved coverage and exact recovery over a compute-matched continuous model and more than doubled Top-1 recovery. Electron-conservation regularization increased conserving outputs without improving correct-step recovery, while few-shot transfer remained limited to related radical chemistries. DiFfER establishes discrete flow matching as a stronger basis for enzymatic mechanism prediction and identifies mechanistic data coverage as a central barrier to broader generalization.