Autocatalytic cycles are reaction loops whose products help make more of themselves. They are central to origin-of-life chemistry. Graph methods can now list such loops exactly, but they say nothing about whether the loops can actually run. A loop is only as fast as its hardest step, set by that step's activation barrier: the energy hill the molecules must climb, found by tracing the lowest-energy route between reactant and product (the minimum energy path). Computing every barrier with quantum chemistry is expensive. Machine-learned shortcuts are cheap, but their barrier errors are large (above 5 kcal/mol on multi-step networks) and uneven across types of chemistry.
We propose to fold barriers directly into the search for autocatalytic loops. The method finds the loop whose worst step is lowest, then spends expensive quantum-chemistry calls only on the few steps that decide whether a loop passes or fails. We borrow an evaluation design from human genetics, where risk models built on European-ancestry cohorts degrade when applied to other ancestries. Each verdict is reported separately for chemistry that resembles the machine-learning training data and for chemistry that does not, and a verdict counts only if it replicates under two independently trained barrier models.
Core Claim
Whether an autocatalytic cycle is kinetically viable depends on a maximum, not a sum. Write $E_a(r)$ for the activation barrier of reaction $r$, $M$ for a candidate autocatalytic flow (a set of reactions with integer multiplicities that returns more of a target molecule than it consumed), and $\tau$ for a feasibility threshold. The gate is
$$\text{viable}(M) \iff B(M) = \max_{r \in M} E_a(r) \le \tau .$$
The task is then to find, among flows satisfying the exact autocatalysis constraints, the one minimizing $B(M)$. This has two consequences. First, average per-reaction barrier error is the wrong figure of merit. A verdict flips only when the error lands on a flow's bottleneck step and that step lies within the error band of $\tau$. Second, for the same reason, a small budget of quantum-chemistry calls aimed at bottleneck steps should recover nearly all the verdicts that full quantum-chemistry exploration produces.
We predict that bottleneck-targeted refinement matches the full-DFT cycle set at a small fraction of the DFT cost. DFT here means density functional theory, the standard quantum-chemistry method. We also predict that the gain over uniform refinement grows as the chemistry moves away from the machine-learned model's training distribution.
Research Program
The structural side reuses the integer-hyperflow formalism. Chemical Transformation Motifs - Modelling Pathways as Integer Hyperflows models reactions as hyperedges, meaning edges with several inputs and several outputs, and solves an integer linear program over whole-number stoichiometries. Defining Autocatalysis in Chemical Reaction Networks adds "exclusive" autocatalysis: the molecule cannot be made from the food set without a seed. It showed this filters 736 formal autocatalysts in E. coli down to 580. On the Realisability of Chemical Pathways checks with Petri-net reachability whether a balanced flow can actually fire in causal order, possibly by borrowing an intermediate and returning it.
All three papers state that they ignore rates and thermodynamics. That omission is the gap we fill. We add the bottleneck objective to their integer program, using an auxiliary variable bounded below by the barrier of every reaction with nonzero flow. We then pass the optimal flow, and the next-best alternatives, to the Petri-net realisability check. Alternatives matter because Zubarev and Rappoport found that autocatalytic optima such as the reverse TCA cycle are degenerate, so many near-equivalent loops exist.
The energetic side follows the network-exploration approach of Context-Driven Exploration of Complex Chemical Reaction Networks. That work generates reactive complexes from electronic-structure heuristics, finds transition states and confirms them by following the path downhill to both endpoints, then detects cycles by tree traversal on a directed graph. It is our ground-truth method, and its directed-graph cycle detection is one of our baselines.
Cheap barriers come from two sources benchmarked in Harnessing Machine Learning to Enhance Transition State Search with Interatomic Potentials and Generative Models. One is a fine-tuned interatomic potential, LEFTNet: a neural network trained to predict quantum-chemistry energies and forces. The other is React-OT, a generative model that proposes transition-state geometries directly. That paper motivates the design as well as supplying it. It shows that foundation potentials fail to locate transition states zero-shot, that direct-force architectures corrupt second derivatives badly enough to derail transition-state optimization, and that barrier error exceeds 5 kcal/mol even for correctly identified reactions on a 131-reaction network.
From disagreement between the two estimators we derive an uncertainty band for each barrier. A reaction is sent to DFT refinement only if two conditions hold: it is the bottleneck, or near-bottleneck, of some candidate flow, and its band straddles $\tau$. After each refinement the integer program is re-solved, because a corrected barrier can move the bottleneck to a different step. We call this loop active verification.
Continuous path representations such as AdaPath, from PINN-MEP, are a secondary option for floppy or solvated steps where discrete methods fail. They are not the core of the proposal.
The population-generalizability lens enters as evaluation discipline, not as a claim about human biology. Cross-ancestry Mendelian randomization studies such as the psychiatric–glaucoma analysis report effects separately per ancestry cohort and trust only conclusions that replicate in both. PRISM showed that a small set of relevant, tissue-specific annotations transferred better to African-ancestry cohorts than a large generic set, improving accuracy by 13.10% with 102-fold fewer annotations.
We carry over both habits. We stratify reactions into chemical "populations" by their similarity to the Transition1x training reactions. We require verdict replication across the two barrier models. And we test whether a small, targeted fine-tuning set beats a larger generic one, mirroring PRISM. We do not use the speculative link between autocatalytic metabolism and ancestry-dependent causal estimates sketched in the file, because nothing in the file tests it.
Baselines
The do-nothing reference is the structural pipeline alone. It runs hyperflow autocatalysis detection plus Petri-net realisability with no barriers and declares every realisable autocatalytic flow viable. Against the DFT reference it should show high recall and poor precision, which quantifies how much chemically inert autocatalysis purely combinatorial methods report.
The reference, and the strongest existing method, is full DFT exploration in the style of Chemoton (Simm and Reiher). On the Formose reaction restricted to molecules of at most $C_4H_8O_4$, it reported 934 molecular configurations, 6,871 minimum energy paths and 10 autocatalytic cycles with low barriers. We treat its cycle set, recomputed by us if the network files are unavailable, as ground truth. We treat its DFT call count as the cost to beat. Beating it means recovering at least 9 of the 10 cycles, with no extra cycles that fail DFT verification, at under a fifth of its DFT calls.
The strongest cheap competitor is React-OT followed by DFT refinement of every proposed transition state. On the Transition1x subset of 960 reactions it identified 836 intended transition states, against 720 for LEFTNet followed by DFT, as reported in Harnessing Machine Learning…. For cycle recovery on any autocatalytic network, we found no reported result on Lacuna for React-OT, LEFTNet or any interatomic potential. Producing that number is part of the contribution.
We run these models with the published checkpoints. Fine-tuning data, the DFT functional and basis set, and the threshold $\tau$ are identical across all arms. Every arm gets the same DFT budget, reported as calls and GPU/CPU hours.
Two further cheap arms gate on GFN2-xTB barriers (a fast semi-empirical quantum method) and on unrefined LEFTNet barriers. Two ablations isolate the key ingredient. One spends the same DFT budget on uniformly random reactions. The other spends it on the reactions with the highest uncertainty regardless of their bottleneck status. A third ablation replaces Petri-net realisability with directed-graph cycle detection, measuring how many barrier-feasible loops cannot actually fire without borrowing.
Why not just…?
Why not just run full DFT, as Chemoton did? For a Formose-sized network this is affordable, and we do it, because it is the ground truth. The problem is scaling. Graph-grammar expansions such as the glyoxylate scenario generate networks far larger than 6,871 paths. The experiment that matters is the cost curve: DFT calls needed to recover the reference cycle set as network size grows.
Why not just gate with a cheap semi-empirical method such as GFN2-xTB and skip DFT? This is a real threat on in-distribution chemistry. If xTB-gated verdicts match DFT on Formose, the refinement loop is unnecessary there. We report this honestly and restrict our claim to chemistry where xTB or the learned potentials fail. The deciding evidence is the sulfur-containing HCN–cysteamine chemistry of Novichkov et al., for which we have no evidence of cheap-method accuracy.
Why not just fine-tune the potential on more reactive data until barriers are accurate? Even the best fine-tuned model in the benchmark kept barrier errors above 5 kcal/mol on a network. That spread is wide enough to flip verdicts near any plausible threshold. Following PRISM's finding that targeted data beats volume, we compare two fine-tuning sets at matched size: a generic Transition1x-style set and one built from bottleneck steps in the target chemistry. If generic fine-tuning closes the verdict gap, targeting adds little and we will say so.
Why not just use directed-graph cycle search or flux balance analysis on the barrier-filtered network? The realisability work showed that perfectly balanced pathways, such as the non-oxidative pentose phosphate pathway, cannot fire without a borrowed intermediate. The ablation above measures how often this happens among low-barrier Formose loops.
Proposed Experiments
The first decisive check is on Formose. We plot recall and precision of the DFT-verified low-barrier cycles against DFT calls spent, for every arm, at three thresholds spanning roughly ±5 kcal/mol around the value consistent with the reference's "low barrier" cycles. The method succeeds if its curve dominates uniform and uncertainty-only refinement at every budget. It fails if uncertainty-only refinement matches it, which would mean bottleneck structure adds nothing.
The second check repeats the protocol on the KHP decomposition network (131 reactions, 1,065 conformations). That network has DFT barriers but no autocatalysis, so here the bottleneck objective reduces to minimax pathway selection. It tests whether the refinement policy transfers to a network built by someone else.
The third check stratifies all reactions by distance from the training distribution. We expect cheap-gate error to rise with distance. We also expect the advantage of targeted refinement to grow with distance, mirroring the ancestry-transfer pattern. Absence of that gradient would falsify the generalization half of the claim.
The fourth check computes verdict replication rates between the LEFTNet and React-OT gates before refinement. It asks whether non-replicating verdicts are where DFT disagrees, which would make cross-model replication a free reliability signal.
The final check runs exploratorily on a small expansion of HCN–cysteamine chemistry, with DFT spot checks only on the cycles that pass. It reports which autocatalytic loops survive, as a first barrier-checked map of that experimentally observed system.
References
- Chemical Transformation Motifs - Modelling Pathways as Integer Hyperflows
- Defining Autocatalysis in Chemical Reaction Networks
- On the Realisability of Chemical Pathways
- Uncertainty of Prebiotic Scenarios: The Case of the Non-Enzymatic Reverse Tricarboxylic Acid Cycle
- Context-Driven Exploration of Complex Chemical Reaction Networks
- Harnessing Machine Learning to Enhance Transition State Search with Interatomic Potentials and Generative Models
- PINN-MEP: Continuous Neural Representations for Minimum-Energy Path Discovery in Molecular Systems
- Mendelian randomization study shows no causal relationship between psychiatric disorders and glaucoma in European and East Asian populations
- PRISM: ancestry-aware integration of tissue-specific genomic annotations enhances the transferability of polygenic scores
- In silico Support for Eschenmoser’s Glyoxylate Scenario
- Autocatalytic Reaction between HCN and Cysteamine Produces Hydrophobic, Catalytic, Liquid Compartments