PaperlayerView exact revision on arXiv

Source-linked AI summary

Modelling Chemical Reasoning to Predict Reactions

Marwin H. S. Segler, Mark P. Waller

arXiv:1608.07117v1cs.AIphysics.chem-phq-bio.MN

TL;DR

Evaluating predicted nonreactions is difficult because failed reactions and limits of synthetic methods are rarely reported. The paper models reaction prediction as link finding in a molecule–reaction graph, reporting 67.5% accuracy on 180,000 held-out reactions versus 52.7% for a rule-based expert system, and correct products for 35% of 13,000 challenging novel reactions.

Abstract

from arXiv · show

The ability to reason beyond established knowledge allows Organic Chemists to solve synthetic problems and to invent novel transformations. Here, we propose a model which mimics chemical reasoning and formalises reaction prediction as finding missing links in a knowledge graph. We have constructed a knowledge graph containing 14.4 million molecules and 8.2 million binary reactions, which represents the bulk of all chemical reactions ever published in the scientific literature. Our model outperforms a rule-based expert system in the reaction prediction task for 180,000 randomly selected binary reactions. We show that our data-driven model generalises even beyond known reaction types, and is thus capable of effectively (re-) discovering novel transformations (even including transition-metal catalysed reactions). Our model enables computers to infer hypotheses about reactivity and reactions by only considering the intrinsic local structure of the graph, and because each single reaction prediction is typically achieved in a sub-second time frame, our model can be used as a high-throughput generator of reaction hypotheses for reaction discovery.

Molecule-Reaction-Graphs as a Knowledge Representation

The paper represents chemical knowledge as a role-typed molecule–reaction graph and uses paths through it to infer reactivity relationships. Those paths support reaction prediction, including product structures and proposed conditions.

Molecule-Reaction-Graphs as a Knowledge Representation

Reaction Prediction as Link Prediction

Illustrative Examples

  • A path links molecules with analogous reactivity to a reaction partner, then extends to a complementary partner to support a predicted reaction.Multiple paths can provide different explanations, and may reflect the same reaction under different conditions.Source: Illustrative Examples · pp. 5–6

Results and Discussion

The model uses a large reaction knowledge graph to predict products and novel reaction types, assess negative predictions, and propose reaction conditions. Its paths also provide inspectable, heuristic explanations of inferred reactivity.

Data

  • The graph contains 14.4 million molecules and 8.2 million binary reactions; two filters retain analogous paths and reject mismatched reaction sites.In the illustrated example, the retained path has reaction-fingerprint Tanimoto similarity 0.41, above the 0.2 threshold; the rejected path has similarity 0.0.Source: Data · pp. 6–7

Quantitative Validation of Reaction Prediction

Predicting Novel Reaction Types

  • 35% of 13,000 reactions not explainable by pre-2014 rules were predicted correctly, including examples of transition-metal-catalysed C-H functionalisation.The authors characterize the graph-based approach as complementary to rule-based systems for reactions existing rules cannot predict.Source: Predicting Novel Reaction Types · p. 9

Negative validation

  • The model classified 94% of plausible, generated wrong products as non-occurring.Negative evaluation is constrained by the lack of published failed reactions and methodological limitations.Source: Negative validation · pp. 9–10

Empirical Validation of Condition Prediction

  • Across 11 recent reactions, products were correct in every case and proposed conditions were similar or functionally similar to reported conditions.Examples of functional similarity include substituting potassium carbonate for caesium carbonate as a base, or copper triflate and TBHP for BQ as an oxidant.Source: Empirical Validation of Condition… · p. 11

Interpreting Link Prediction as Chemical Reasoning

  • Paths recombine known reactions to infer analogous or complementary reactivity and offer inspectable explanations, but are heuristic rather than transitive laws.The model considers complete molecules without predefined functional groups or encoded rules, implicitly capturing functional-group tolerance and regio-/chemoselectivity.Source: Interpreting Link Prediction as… · p. 13

Conclusion

The model uses knowledge-graph link prediction and path-finding to generate reaction hypotheses, including beyond known reaction types. The authors also identify the model’s current scope and the need for further quantitative work on catalyst and reagent prediction.

Conclusion

  • The model generalises beyond its knowledge to reaction types it has not seen, while quantitatively predicting products of binary reactions and detecting unlikely reactions.The authors report that predicting catalysts and reagents for novel reactions also appears feasible, but say more quantitative research is needed.Source: Conclusion · p. 14
  • Path-finding relates molecular reactivity as analogous or complementary, supporting link prediction from essentially the complete published knowledge of organic chemistry.The approach generates hypotheses through knowledge-graph link prediction.Source: Conclusion · p. 14
  • The model was deliberately restricted to binary reactions and molecules with known reactions to introduce and validate the concept.The authors anticipate overcoming these limitations, including by augmenting the graph with hierarchical knowledge about molecules and reactions.Source: Conclusion · p. 14
  • The authors anticipate using the approach as an inspiration tool for unprecedented reactions and plan to examine combinations with machine learning and low-cost quantum mechanics.They describe the hierarchical-knowledge extension as preliminary laboratory work.Source: Conclusion · p. 14
Loading 1608.07117v1…