Fluxion

Ask why one cytochrome P450 hydroxylates arachidonate at one carbon while a close relative attacks another, and an Enzyme Commission number is not an answer. Most machine-learning tools for enzyme function return exactly that: they assign a sequence to a closed set of EC classes, or retrieve the nearest annotated reaction from a database. Both are bounded by what has already been catalogued, which is the opposite of what is needed when hunting promiscuous enzymes for unrecorded transformations.

Fluxion, developed at AITHYRA with the University of Queensland and collaborators, replaces retrieval with generation. It models an enzymatic step as electron redistribution — the arrow-pushing chemists draw by hand — and samples trajectories through that space instead of picking from a list. Reactant and product states are encoded as bond-electron (BE) matrices, whose diagonal counts non-bonding electrons per atom and whose off-diagonal entries hold bond orders; a step is a stochastic transport between two such matrices, learned as a Schrödinger bridge. Because the representation counts electrons, every generated step conserves them by construction.

What makes the output enzyme-specific rather than merely chemically plausible is the conditioning. Substrates, cofactors and user-designated catalytic or binding residues enter as BE matrices; the surrounding protein enters as ESM-2 3B embeddings injected by cross-attention. The same substrate presented with two different P450 sequences therefore yields different predicted reaction sites. Catalytic-residue annotations can come from tools such as Squidly, built by the same group.

#Key Features

• Generation rather than classification: Inference draws 64 samples per reactant and returns a frequency-ranked distribution over products, so competing plausible outcomes are ranked instead of collapsed into one call.
• Electron conservation by construction: The bond-electron formalism makes mass and electron balance a property of the representation, not something a post-hoc filter must enforce.
• Enzyme-conditioned regioselectivity: Cross-attention over protein language model embeddings lets sequence identity shift probability mass between candidate reaction sites on a single substrate.
• Multi-step mechanisms: A beam search over iterated single-step predictions assembles complete arrow-pushing pathways, ranked by mean per-step log-probability.
• Reusable embeddings and auxiliary heads: Encoder representations feed downstream specificity prediction, and optional reactivity and activity heads score enzyme-substrate pairs without disturbing product generation.

#Technical Details

The encoder is a GraphGPS-style graph transformer: six pre-norm layers, hidden width 256, 16 attention heads, feed-forward width 1024, combining edge-aware message passing over the BE graph with global self-attention. ESM-2 3B residue embeddings of dimension 2560 enter through eight-head cross-attention re-injected every second layer. Training runs in two stages — pretraining on the atom-mapped USPTO mechanistic corpus curated for FlowER, with its published partition adopted unchanged, then fine-tuning on 2,747 steps across 599 enzyme mechanisms reconstructed from M-CSA by an agentic annotation pipeline that passed 599 of 734 entries (81.6%) through chemical validation. A separate reactant-to-product model trains on 237,207 USPTO reactions and 28,837 EnzymeMap enzyme-reaction pairs.

On held-out M-CSA mechanisms split by mechanistic similarity to training, top-1 stepwise accuracy is 45%, 32% and 22% on the easy, medium and hard splits, exceeding the coverage ceiling of rules derived from the same training data at k=3. Product-informed pathway accuracy at k=5 is 20.0%, 11.3% and 6.3%. Across five held-out CYP substrates, site-ranking AUC averages 0.709, reaching 0.897 on arachidonate against a substrate prior of 0.811. On esterase and phosphatase specificity screens, a gradient-boosting model over Fluxion embeddings performs on par with EzSpecificity and above ESP — though the same model over plain ESM-2 embeddings and a Morgan fingerprint matches both, which the authors read as evidence the signal may not be Fluxion-specific.

#Applications

The intended user is an enzyme engineer deciding where to begin a directed-evolution campaign. Given a target transformation and a candidate enzyme, it returns ranked products with an inspectable mechanism, and screens at a practical rate — roughly 2,000 enzyme-substrate pairs in under an hour at eight samples per pair. Tested against 206 experimental engineering campaigns drawn from EnzEngDB, it generated the observed non-native product for 33 of them within a 64-sample budget. The mechanistic output is useful in its own right: a chemist can read the proposed arrow-pushing and judge it, which neither an EC label nor a binary activity call permits.

#Impact

The authors frame Fluxion as a proof of concept, and its limits are worth stating plainly. Without fine-tuning, generation is driven by substrate and cofactor chemistry more than by enzyme identity — randomly permuting which enzyme accompanies a reaction does not reduce the number of correct non-native products across EnzEngDB. Training carries no negative examples, so the model never learns when to abstain, and applying it to screening requires fine-tuning on experimental negatives. The atom-conserving formalism also obliges the user to supply every species contributing atoms, cofactors included. What it does establish is narrower and still consequential: mechanistic generative modelling recovers transformations outside any rule set derived from the training mechanisms, and its representations carry transferable signal. Code, data and weights are stated as pending release at publication, and none are public yet.