Who should jump in now? Teams with proprietary data and a need for diverse, strategy-aware planning.
The clearest early winners are organizations that hold proprietary reaction datasets. A 2023 study from IBM Research showed that retraining language models on a mix of patent and proprietary data produced a 'considerable boost in accuracy' for both reaction outcome prediction and single-step retrosynthesis [3]. That means if your team has years of in-house reaction records, you can customize a model to your specific chemistry and see immediate gains—something a generic public model can't offer.
The second group that benefits early are teams whose work demands diverse disconnection strategies—that is, finding multiple chemically different ways to break a molecule apart. A 2023 ACS Central Science paper found that adding a 'disconnection prompt' (a text hint about where to break the molecule) improved prediction performance by 39% over the baseline and produced a broader set of precursor suggestions [2]. Another 2023 study showed that prepending a classification token to the molecule's text representation lets you steer the model toward different reaction families, which helps recursive synthesis tools avoid dead ends and find routes for more complex molecules [4]. For medicinal chemists exploring many synthetic options, this diversity is a game-changer.
A 2026 paper in Matter goes further, showing that large language models can guide search algorithms toward 'strategy-aware' retrosynthetic plans—letting chemists specify in natural language things like protecting-group strategies or global feasibility [1]. That means teams doing complex, multi-step synthesis planning—where strategic thinking matters as much as single-step accuracy—will see value sooner than teams doing routine, well-templated reactions.
Who should hold off? Teams with narrow, well-templated chemistry or no proprietary data.
If your team works on a narrow set of reactions with established templates, the added complexity of a language model may not pay off. The same 2023 study that showed gains from proprietary data also noted that the lack of publicly available datasets with universal coverage is often the limiting factor for higher accuracy [3]. In other words, if you don't have proprietary data to customize the model, you're stuck with the biases and gaps of public training sets.
Another reason to wait: the models still have diversity problems. A 2023 Digital Discovery paper explicitly called out that a common issue is 'a lack of diversity in the proposed disconnection strategies'—the models tend to suggest precursors from the same reaction family [4]. While prompts help, they require extra engineering and a human-in-the-loop to choose the right prompt. If your team isn't ready to invest in that workflow, you might not see the benefit.
Also, the technology is evolving fast. A 2024 Nature Machine Intelligence paper introduced ChemCrow, an LLM agent that integrates 18 expert-designed tools, and showed it could autonomously plan and execute syntheses [6]. That suggests that the next generation of tools will be more capable and easier to use. Waiting a year or two could mean adopting a more mature, integrated system rather than building your own from scratch.
What's been overturned? The old view that retrosynthesis models are just black-box translators.
The older view, from around 2016-2020, was that retrosynthesis language models were essentially translation tools: you feed in a molecule's SMILES string and get out precursor strings, with little user control. That's been overturned by the prompt-based and strategy-aware approaches. The 2023 disconnection prompt paper showed you can steer the model's output by adding a text hint, giving chemists 'greater control over the disconnection predictions' [2]. The 2026 Matter paper goes further, showing LLMs can reason about chemical strategy and guide search algorithms, not just predict single steps [1].
Another revision: the idea that you need a huge, universal dataset to get good performance. The 2023 customization study showed that combining a modest amount of proprietary data with public patent data can yield a significant accuracy boost, even for out-of-distribution reactions [3]. That means teams with niche chemistry can build useful models without waiting for a perfect public dataset.
Finally, the role of the chemist has shifted from passive user to active director. The 2025 ChatChemTS paper showed that an LLM-powered chatbot can let chemists design molecules through chat interactions, automatically building reward functions for desired properties [5]. That's a far cry from the early days where you needed to be an AI expert to use these tools. The trend is clear: the barrier to entry is dropping, but the teams that benefit most are those that can provide their own data and are willing to engage with the model's suggestions.
About These Sources
This answer is built on 6 peer-reviewed studies — published from 2023 to 2026, 3 from 2024 or later, 6 in Q1 journals, collectively cited 527 times — selected as the most relevant from 8 studies that passed quality screening, drawn from 41 papers retrieved from a database of over 500 million.
Sources used in this answer
Chemical reasoning in LLMs unlocks strategy-aware synthesis planning and reaction mechanism elucidation
Showed that LLMs, when integrated with Monte Carlo Tree Search, can guide strategy-aware retrosynthetic planning and mechanism elucidation, with newer and larger models showing increasingly sophisticated chemical reasoning.
Unbiasing Retrosynthesis Language Models with Disconnection Prompts
Demonstrated that using a disconnection prompt in retrosynthesis language models improves performance by 39% over baseline and increases diversity of precursor suggestions, applicable to traditional and enzymatic reactions.
Fast Customization of Chemical Language Models to Out-of-Distribution Data Sets
Reported a methodology for retraining language models on proprietary datasets, showing a considerable boost in accuracy when combining patent and proprietary data in a multidomain learning formulation, with guidelines for corporate customization.
Enhancing diversity in language based models for single-step retrosynthesis
Showed that prepending a classification token to the target molecule's representation allows steering a retrosynthesis Transformer toward different disconnection strategies, consistently improving diversity and enabling recursive synthesis to circumvent dead ends.
Large language models open new way of AI-assisted molecule design for chemists
Developed ChatChemTS, an LLM-powered chatbot that assists users in designing molecules via chat interactions, including automated construction of reward functions, demonstrated in de novo design of chromophores and anticancer drugs.
Augmenting large language models with chemistry tools
Introduced ChemCrow, an LLM chemistry agent integrating 18 expert-designed tools, which autonomously planned and executed syntheses and guided discovery of a novel chromophore, demonstrating new capabilities in automating chemical tasks.