12 years ago, researchers tried making sulfur fluoride exchange the reliable snap-together connector click chemistry wanted. It didn't work. This paper explains why and fixes it.
Well, “didn’t work” is unfair in the same way saying a rocket “didn’t work” because it only reached orbit and not Mars is unfair. SuFEx, short for sulfur(VI) fluoride exchange, absolutely works. It has become a handy way to build sulfur-based links in chemical biology, materials science, and drug discovery. The issue is that chemistry keeps being chemistry: moody, conditional, and weirdly offended by tiny changes in context.
A reaction that looks obvious on paper can give a beautiful yield, a sad little smear, or the wrong product entirely. On one hand, this is why synthetic chemists have jobs. On the other hand, it is also why synthetic chemists occasionally stare into the middle distance like they have seen the face of God in a failed TLC plate.
Tan and colleagues’ new JACS paper asks a very practical question: can machine learning predict both yield and chemoselectivity in SuFEx reactions, while still respecting the underlying chemistry instead of acting like a spreadsheet with a lab coat? Their answer is: yes, at least enough to be useful, and maybe enough to change how chemists plan these reactions.
The Problem: Click Chemistry, But With Attitude
Click chemistry is supposed to be the neat part of chemistry. You bring two molecular pieces together, they connect cleanly, everyone claps, nobody has to purify tar for three days. SuFEx joined that family because sulfur-fluoride bonds can be stable until the right partner comes along, then exchange in useful ways. Reviews have framed SuFEx as a broadly useful platform for drug discovery, polymer science, and biochemistry DOI: 10.1038/s43586-023-00241-y.
But SuFEx has too many possible dance partners. Different SuFEx “hubs,” nucleophiles, bases, additives, solvents, and molecular neighborhoods can all change what happens. Yield asks, “How much product did we get?” Chemoselectivity asks, “Which site reacted when multiple sites could?” That second question is where the universe starts giggling.
The authors tackled this by building a literature-derived dataset using large-language-model-enabled agents. Yes, even chemistry papers now have AI assistants reading the archives, which feels efficient and faintly cursed. Importantly, the dataset included successful reactions, low-yielding reactions, and failures. That matters because a dataset containing only winners is basically a dating profile: technically data, spiritually incomplete.
The Fix: Let The Model Learn, Then Make It Explain Itself
Machine learning for reaction prediction has a reputation problem. Models can sometimes predict outcomes, but if they cannot explain anything, chemists reasonably ask whether the model learned chemistry or just memorized lab gossip. Recent reviews have warned that yield prediction remains difficult because reaction data are messy, biased, sparse, and often not transferable across settings DOI: 10.1021/acs.jcim.3c01524. Real-world electronic lab notebook data have shown the same headache: more data does not magically mean better prediction if the data are noisy or uneven DOI: 10.1039/D2SC06041H.
This paper’s interesting move is that it does not just throw descriptors into a model and hope the GPU interns sort it out. The authors pair curated reaction data with mechanistic guidance. Their models estimate reaction yields and forecast chemoselectivity across different SuFEx hubs, nucleophiles, and additives. Then density functional theory, or DFT, helps explain the chemistry underneath. DFT is the computational method chemists use when they want quantum mechanics to answer a question without asking every electron to submit a memoir.
The mechanistic story centers on three cooperating forces: SuFEx-hub electrophilicity, nucleophile acidity, and additive basicity. In normal-person terms: how eager the sulfur-containing hub is to be attacked, how acidic the reacting partner is, and how the additive nudges the whole situation. On one hand, this is elegant. On the other hand, it means your reaction outcome may depend on a tiny acid-base negotiation happening at molecular scale, which is frankly rude.
Why This Matters Beyond One Reaction Family
The result is SuFExPredictor, a public platform meant to guide reaction optimization, rescue unproductive transformations, and help with late-stage functionalization of drug-like molecules. That last part matters because medicinal chemists often want to modify complex molecules without rebuilding them from scratch. If prediction can reduce trial-and-error, it saves time, materials, and emotional bandwidth.
This also fits a broader shift in AI-for-chemistry: models are moving from “black box oracle” toward “computational coworker who gives reasons.” Work like ReaMVP has explored large-scale pretraining for reaction yield prediction DOI: 10.1186/s13321-024-00815-2, but chemistry still rewards models that know when the molecules are being sneaky.
Could this approach generalize to other reaction classes? Maybe. That is the hopeful side. The dread side is that each reaction family may demand its own carefully curated dataset, mechanistic descriptors, and ritual offering to the reproducibility gods. Still, this is the right shape of progress: not AI replacing chemical intuition, but AI making that intuition less dependent on heroic guesswork and caffeine.
On one hand, the machines are learning to predict chemistry. On the other hand, they still need chemists to tell them what reality means. I do not know if that makes me comforted or nervous. Possibly both. Definitely both.
References
-
Hao-Dong Tan, Ben Gao, Huaihai Huang, Tingjun Xu, Yao Li, Jiajia Dong, Xiao-Song Xue. “Data-Driven, Mechanistically Guided Prediction of Yield and Chemoselectivity in SuFEx Reactions.” Journal of the American Chemical Society (2026). DOI: 10.1021/jacs.6c04549. PubMed: PMID 42241638
-
“Sulfur fluoride exchange.” Nature Reviews Methods Primers (2023). DOI: 10.1038/s43586-023-00241-y
-
“When Yield Prediction Does Not Yield Prediction: An Overview of the Current Challenges.” Journal of Chemical Information and Modeling (2023). DOI: 10.1021/acs.jcim.3c01524
-
“On the use of real-world datasets for reaction yield prediction.” Chemical Science (2023). DOI: 10.1039/D2SC06041H
-
“Prediction of chemical reaction yields with large-scale multi-view pre-training.” Journal of Cheminformatics (2024). DOI: 10.1186/s13321-024-00815-2
Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.