Chemical models of cold interstellar clouds underpredict how much of certain molecules are actually there. In the Taurus Molecular Cloud 1 (TMC-1), ring-shaped molecules such as cyanocyclopentadiene and 2-cyanoindene are observed at abundances 100 times or more above what standard reaction networks predict. The leading explanation is that the networks are missing reactions, not that they have the wrong rates for the reactions they already contain. These networks are hand-assembled catalogs of thousands of gas-phase and grain-surface reactions. If a reaction is absent, no amount of rate tuning can supply it. This proposal treats "which reactions are missing" as a search problem with two parts. A graph model learns, from the known catalog and from each molecule's chemical structure, which unlisted reactions look plausible. A differentiable chemistry simulator then tests whether adding the top-ranked candidates actually closes the gap between model and telescope.
The core bet is that the graph model is useful as a prior that narrows the search. The real constraint is observational: TMC-1 gives column densities (total amount of a molecule along the line of sight) for a few dozen relevant species. Any balanced reaction among the network's species is a candidate, so these few measurements cannot single out the right additions among thousands of options. An attribute-aware graph embedding is a vector representation of each molecule learned from both its reaction neighbors and its intrinsic chemical features. Such an embedding can say which candidates resemble known interstellar chemistry, including for newly detected ring species that have almost no catalogued reactions. The new element is the coupling between this learned plausibility and a simulator fit to observations. Neither the existing astrochemical machine-learning work nor the existing reaction link-prediction work makes this coupling.
Where the existing work stops
Two lines of work come close, and each covers only half of the problem.
The first line is astrochemical machine learning, which predicts abundances while ignoring reactions. Explaining the Chemical Inventory of Orion KL through Machine Learning encodes molecules as strings, compresses them with an autoencoder, and regresses column densities with gradient boosting. Implementation of rare isotopologues into machine learning of the chemical inventory of the solar-type protostellar source IRAS 16293-2422 does the same with Mol2vec embeddings and Gaussian processes. ASTROCHEMICAL FORECASTING WITH MACHINE LEARNING forecasts TMC-1 abundances from pretrained molecular embeddings and argues explicitly that hand-curated networks bias the result. These papers motivate the proposal, because they show molecular structure predicts abundance. Their models, however, discard the reaction network and so cannot say which pathway is missing.
The second line is reaction link prediction, which models connectivity but only in terrestrial chemistry. Learning graph representations of biochemical networks and its application to enzymatic link prediction (ELP) fuses KEGG metabolic connectivity with fingerprint attributes and predicts missing enzymatic reactions, including for molecules unseen in training. Chemical-Reaction-Aware Molecule Representation Learning (MolR) trains a graph neural network so that reactant embeddings sum to product embeddings, which turns reaction plausibility into a vector-arithmetic test. These two papers supply the method. Both are trained on enzyme or patent chemistry near room temperature. Neither ever checks a predicted reaction against a physical simulation.
The simulator role is filled by Carbox: an end-to-end differentiable astrochemical simulation framework. It turns UMIST, KIDA-style, PRIZMO and UCLCHEM networks into GPU-differentiable stiff equation systems, so gradients of abundances with respect to rate coefficients are available. It operates on a fixed set of reactions.
Core Claim
Missing-reaction discovery in interstellar networks becomes much better posed if the candidate reactions are ranked by a reaction-aware, attribute-aware embedding before a differentiable simulator fits their rates to observations. The claim is specific and falsifiable. We will deliberately delete reactions from a known network and generate synthetic observations from the complete network. An embedding-ranked candidate set should then recover the deleted reactions at higher precision than either a uniform candidate set or a structure-only ranking, at equal simulator budget. On real TMC-1 data, the imputed network should predict held-out species better than the original network and better than flat regression.
One observational fact sharpens the hypothesis. Discovery of Interstellar 2-Cyanoindene (2-C9H7CN) in GOTHAM Observations of TMC-1 reports that models get the ratios between pure hydrocarbon rings and their cyano-substituted forms roughly right while missing absolute abundances. The cyanation step therefore looks fine, and the deficit should sit upstream, in ring formation from flexible chains. Detection of interstellar H2CCCHC3N and Detection of Interstellar E-1-cyano-1,3-butadiene in GOTHAM Observations of TMC-1 identify such chains as plausible precursors. If the method's top fitted additions land downstream of cyanation, that is evidence against the method rather than a discovery.
Research Program
The first component is a reaction-plausibility model. Reactions are hyperedges: two reactants yield one or more products, so pairwise links are not enough. Each molecule node carries MACCS fingerprints, which ELP found often beat larger fingerprints. Each node also carries charge, radical character, and ring count.
A graph neural network propagates information over the known catalog and produces molecule embeddings $h$. We score a candidate reaction $r$ with reactants $R$ and products $P$ as
$$s(r) = -\Big\lVert \sum_{i \in R} h_i + t_{c(r)} - \sum_{j \in P} h_j \Big\rVert,$$
where $t_{c(r)}$ is a learned offset vector for the reaction's mechanism class $c(r)$. The classes are ion-molecule, radical-neutral, dissociative recombination, grain-surface, and so on. The offset lets the conservation rule of MolR differ by mechanism instead of assuming ambient organic chemistry. Training uses a margin loss against corrupted reactions in which one product is swapped for another mass- and charge-balanced species. Pretraining on USPTO with the released MolR code, followed by fine-tuning on the astrochemical catalog, is one arm to test. The other arm trains on the catalog alone, to see whether terrestrial chemistry transfers or misleads.
The second component is candidate generation and fitting. We enumerate all mass- and charge-balanced two-body reactions among network species, with a cap on product count. The plausibility model keeps the top $K$. In Carbox, each kept candidate receives a rate coefficient with a log-normal prior. The prior's location comes from the typical rate of its mechanism class, and its inclusion penalty is inversely proportional to $\exp(s(r))$. The rates are then fit to observed column densities by gradient descent under a TMC-1-like physical model. The output is a short ranked list of reactions whose inclusion most reduces the discrepancy. Each comes with a sensitivity, which tells laboratory chemists what to measure.
The third component is the controlled benchmark that makes this testable. We will use the KROME cloud network from Systematic selection of surrogate models for nonequilibrium chemistry (37 species, 287 reactions) and the larger UMIST- or KIDA-format network loaded through Carbox. We delete subsets of reactions under three schemes: at random, by mechanism class, and all reactions touching a chosen species. That last scheme mimics a newly detected molecule with no catalogued chemistry. Synthetic observations come from the full network, restricted to the species a survey like GOTHAM would actually see. Because the deleted reactions are known, both ranking and recovery can be scored exactly.
Nuclear reaction networks for neutron-capture processes are structurally similar, but they are out of scope here. They are a natural second application once the method is validated on the astrochemical case.
Baselines
The naive reference point is the original network with its rates refit in Carbox and no added reactions. This is the strongest version of "the network is fine, only the numbers are off." If refitting existing rates alone closes the TMC-1 ring deficit without implausible rate values, the missing-reaction premise is wrong. Every method gets the same observations, physical model, and optimizer budget.
The flat-regression baseline is the pipeline of ASTROCHEMICAL FORECASTING WITH MACHINE LEARNING, trained on the same TMC-1 species, together with the Gaussian-process variant of Fried et al. using Mol2vec features. Lee et al. report reproducing unseen molecules' column densities "to within an order of magnitude without any tuning parameters." That is the bar on held-out-species prediction; beating it means a lower median log error under the identical leave-one-species-out split. This baseline cannot propose reactions at all, so it competes only on abundance prediction.
The strongest existing reaction-ranking method is ELP-style embedding propagation over attributed molecular graphs, from Learning graph representations of biochemical networks and its application to enzymatic link prediction. It reports a 24% AUC improvement over fingerprint-similarity link prediction on KEGG. No absolute AUC is given in the file, and there is no reported result found on Lacuna for link prediction on any astrochemical network. ELP, MolR without mechanism offsets, and our scorer will be trained on identical catalog splits with matched hyperparameter search budgets. Each will be scored on held-out reactions by area under the curve and by Hit@10 among balanced candidates. For calibration, MolR reports Hit@1 of 0.882 on USPTO-479k against 0.614 for Mol2vec, but that is a different task and dataset. We will not treat it as a target, only as evidence that reaction-aware embeddings beat isolated-molecule embeddings at product ranking.
Two ablations isolate the key ingredients. The first replaces the embedding prior with a uniform prior over the same $K$ candidates, and separately over all balanced candidates. This measures whether the learned ranking adds anything beyond sparsity. The second replaces attributes with identity features, which is equivalent to node2vec-like structure-only embeddings. This measures whether chemical attributes matter, and it should matter most in the species-deletion scheme. Beating the uniform-prior ablation on synthetic recovery precision at matched simulator calls is the result the whole proposal rests on.
Why not just…?
Why not just regress column densities with gradient boosting or Gaussian processes, which already reach $R^2$ of 0.883 and 0.920? Those numbers come from Orion KL and IRAS 16293-2422B, and on abundance prediction alone these models may well match ours; we will report it honestly if they do. They answer a different question, though. They cannot name a reaction, which means they offer no laboratory measurement to make and no correction to the physical model. The comparison that matters for this proposal is reaction recovery, where they have no output.
Why not just embed the known network with node2vec or DeepWalk? Newly detected rings such as 2-cyanoindene have few or no catalogued reactions, so a structure-only embedding has no signal for them. In other domains, attribute fusion helped most on sparse graphs: CARL: Content-Aware Representation Learning for Heterogeneous Networks reported 10.9–41% gains over structural baselines, and An attribute-enhanced relationship-aware neighborhood matching model with dual attention reported 83.91% against 41.25% Hits@1 for structure-only GCN-Align on DBP15K. Those are not chemistry results. The species-deletion ablation tests the claim directly in this setting.
Why not just add every balanced candidate reaction to Carbox and fit with a sparsity penalty? This is the real threat, and it is why the proposal centers on the simulator rather than on link prediction alone. If observations constrain the network well, sparse fitting over all candidates may recover the missing reactions without any learned prior. Our expectation is that GOTHAM-scale observations underdetermine thousands of candidate rates. Under that condition, sparse fitting will find many equally good but chemically arbitrary solutions. The synthetic deletion benchmark settles this. If the uniform-prior arm matches the embedding-prior arm in recovery precision, we will drop the embedding and report that the simpler method suffices.
Why not just train a fast neural surrogate of the chemistry? The CODES benchmark found that fully connected surrogates have the best one-shot accuracy, but that they accumulate errors under repeated rollout. More basically, a surrogate maps abundances to abundances and contains no explicit reactions to add. A surrogate could later speed up the fitting loop, but it cannot replace the search.
Proposed Experiments
The decisive experiment is synthetic recovery. On the KROME cloud network and the larger catalog, we delete 5–20% of reactions under each of the three schemes and generate observations for survey-visible species only. We then report precision at 10 and 50 for recovered reactions, plus the log abundance error on unobserved species, using the mean log absolute error from CODES. Each scheme is run across the embedding-prior, uniform-prior, structure-only, ELP, and no-addition arms at a fixed number of simulator gradient steps. Varying the number of observed species should show the embedding prior's advantage growing as observations become scarcer. If that trend does not appear, the core claim fails.
The second experiment is TMC-1 itself. We fit the imputed network to GOTHAM column densities while holding out one ring species at a time: cyanocyclopentadiene, 2-cyanocyclopentadiene, 2-cyanoindene, and the related chains. We then predict the held-out species' abundance and compare against the refit-only network and the regression baselines. We will also check where the top fitted reactions sit relative to the cyanation step. Upstream placement is consistent with the ratio evidence of Sita et al.; downstream placement counts against the method. The candidate list will also be screened against the N-, O-, and S-bearing heterocycle non-detections in Barnum et al.. An imputed network that overproduces species the survey failed to detect is penalized.
The final output is a short list of proposed low-temperature reactions, each with the abundance change it explains and the rate range it requires. These are claims a laboratory kinetics group can confirm or refute, which is the only validation that ultimately counts.
References
- Explaining the Chemical Inventory of Orion KL through Machine Learning
- Implementation of rare isotopologues into machine learning of the chemical inventory of the solar-type protostellar source IRAS 16293-2422
- ASTROCHEMICAL FORECASTING WITH MACHINE LEARNING
- Learning graph representations of biochemical networks and its application to enzymatic link prediction
- Chemical-Reaction-Aware Molecule Representation Learning
- Carbox: an end-to-end differentiable astrochemical simulation framework
- Discovery of Interstellar 2-Cyanoindene (2-C9H7CN) in GOTHAM Observations of TMC-1
- Detection of interstellar H2CCCHC3N
- Detection of Interstellar E-1-cyano-1,3-butadiene in GOTHAM Observations of TMC-1
- Systematic selection of surrogate models for nonequilibrium chemistry
- CARL: Content-Aware Representation Learning for Heterogeneous Networks.
- An attribute-enhanced relationship-aware neighborhood matching model with dual attention
- A Search for Heterocycles in GOTHAM Observations of TMC-1