Reaction models judged only by endpoints miss the point: UniRxnAxis splits the latent update, and its orthogonal residual carries most retrosynthesis utility
Synopsis
Reaction-prediction models are usually judged by endpoint accuracy, which does not show how a model represents a chemical transformation. UniRxnAxis unifies forward prediction and retrosynthesis as direction-conditioned endpoint redistribution in a shared latent space, making source states, transformed states and encoded target endpoints directly comparable, and decomposes the learned displacement into an endpoint-parallel component and an endpoint-orthogonal residual. On USPTO-50K retrosynthesis, decoding only the parallel state gives top-1 0.0997, while decoding only the orthogonal residual reaches 0.6303, close to the full state's 0.6441.
Interpretation
A reaction model's learned displacement can be split into a part pointing along the direction of the true product and a perpendicular part, and the latter still supports retrosynthesis decoding. Reaction models have been evaluated by endpoint accuracy, leaving little handle for inspecting how a model organizes a chemical transformation or where it fails. On the USPTO-50K retrosynthesis test split, decoding the endpoint-parallel state alone gives top-1 0.0997, decoding the residual-only state gives 0.6303, and the full transformed state gives 0.6441.
The split relies on a shared encoder that places reactants and products in one latent coordinate system, with a direction-conditioned module separating forward prediction from retrosynthesis. The direction-conditioned transformation is a shared update module followed by a lightweight direction adapter, and the decoder is conditioned on the source state and predicted displacement, predicting only residual bond-order edits relative to the source graph. The audit uses standard USPTO-50K splits; the jointly trained checkpoint reaches top-1 64.4 on retrosynthesis, and forward prediction reaches 0.8832 under joint training and 0.8816 under forward-only training.
The orthogonal residual is not a norm or template artifact; it selectively pushes bond edits. Replacing the residual with norm-matched random directions, token or sample shuffles, same-class shuffles, or class prototypes does not recover the true residual's performance. The true residual gives top-1 0.6441, norm-matched random directions give 0.1007 and 0.1022, class prototypes give 0.2606, 0.1460 and 0.1558, and shuffle controls give 0.0855, 0.1162 and 0.1188; the residual pushes 52.17% of target edit edges across the decoding threshold while only 0.04% of wrong edges cross.
Perspective
The demonstrated scope is decoder-executability evidence on USPTO-50K retrosynthesis with two architectures, UniRxnAxis and Graph2Edits; on USPTO-MIT only the cross-dataset stability of the parallel and orthogonal geometry is tested, not decoder executability. The intended readers are cheminformatics and machine-learning researchers who want to audit how a reaction model transforms its input, and only deterministic top-1 results are reported.
The decomposition uses the true endpoint, so the orthogonal residual is an audit-time diagnostic rather than an inference-time control. The size of the parallel-versus-residual gap depends on the reference axis, the alignment weight and the decoder interface, and only its sign stays stable across the tested settings; reaction-center and failure-category labels are automatic proxies that have not been validated by chemists.
