Combining two distinct AI models designed to ‘think’ in completely different manners minimizes the failures and hallucinations that may occur when executing a significant chemical operation – retrosynthesis.
‘We discover that various models have unique strengths, and their integration can lead to substantial performance enhancements compared to any individual model by itself,’ states Marwin Segler, senior director at Microsoft Research AI for Science in Cambridge, UK. The integrated tool, Retrochimera, can also be inexpensively and swiftly adjusted to operate with new reaction databases.
Segler and his team surveyed PhD chemists regarding their preference for retrosynthetic steps proposed from a reference dataset or from Retrochimera. The chemists favored Retrochimera’s recommendations approximately two-thirds of the time. ‘To our knowledge, this is the first instance where a model is shown to generate predictions that align more closely with chemists’ preferences than its test-set reference reactions,’ the researchers document.
AI interprets the language of chemistry
Retrosynthetic analysis aids chemists in strategizing molecule synthesis of interest by working backward from the target structure. Researchers have historically investigated AI tools’ application in this area, yet their effectiveness had previously declined rapidly with molecules significantly different from their training data.
By examining various AI types to assess their potential to expedite molecular design, Segler’s team noticed that their strengths and weaknesses could complement one another. One tested type is a transformer, prevalent in widely used large language models like ChatGPT. In written language, transformers grasp a word’s meaning based on its contextual frequency among other words.
In the recent study, one of Retrochimera’s AI ‘personalities’ or submodels is a retrosynthetic transformer named R-Smiles 2. The term Smiles refers to Simplified Molecular Input Line Entry System, a method of employing letters and symbols to denote chemical structures.
‘Smiles, akin to natural language, is a sequence of symbols dictated by rules,’ Segler explains. ‘In natural language, the symbols are words and the rules are grammar. In Smiles, the symbols are atoms, bonds, ring closures, and more, while the Smiles syntax determines which sequences are valid molecules.’ Therefore, transformers ‘can comprehend the structure of both natural language and Smiles’.
The retrosynthetic tools Segler and his colleagues evaluated included an entirely different kind of AI known as graph neural networks (GNNs). GNNs are frequently utilized in social networks and in chemistry applications because they excel at portraying structured objects. Consequently, for the other AI submodel of Retrochimera, the researchers devised a GNN referred to as Neuralloc.
Reducing hallucinations by over half on industrial reactions
Both submodels accept a single molecule as input and yield a prioritized list of reactions that could synthesize that molecule, Segler informs Chemistry World. Following extensive training on both common and rare reactions, Retrochimera learns to rank the integrated lists as well.
Collaboration with team members from Novartis in Basel, Switzerland, allowed the researchers to apply Retrochimera directly for retrosyntheses on 10,444 reactions sourced from the drug company’s internal database. The partnership also involved scientists from GSK in Stevenage, UK, who supplied over 650,000 reactions from their database for training, validation, and testing. With the GSK data, the researchers investigated whether retraining or fine-tuning Retrochimera would enhance its performance.
The combined AI tool consistently outperformed each individual submodel, particularly for rare transformations. The cross-checking mechanism between submodels minimizes the number of hallucinations, or odd suggestions, by more than half.
‘We’ve also discovered that fine-tuning the model on internal datasets can further elevate performance, which is crucial for real-world applications in pharmaceutical and chemistry companies,’ Segler remarks. Additionally, due to Retrochimera’s modular structure, new component models can be introduced over time, which Microsoft anticipates will further boost performance.
‘Retrochimera is a remarkable, meticulously validated achievement and arguably the most robust single-step retrosynthesis model to date,’ states Philippe Schwaller from EPFL in Switzerland. ‘The advancements concerning rare reactions and the reduced occurrence of hallucinated reactions are significantly important in practice.’
He observes that previous evaluations of retrosynthetic AI were based on the fidelity of reproduced answers. ‘That’s an inadequate measure, as there is seldom a singular correct answer,’ Schwaller asserts. He wishes for blind assessments by expert chemists to ‘become the standard’.
In the year following the submission of the newly published paper, Segler notes, Microsoft has progressed its internal models in collaboration with partners. Concurrently, RetroChimera ‘continues to be highly competitive in independent benchmarks despite its release last year, in a rapidly evolving field, while also being very cost-effective,’ he adds. Microsoft is additionally working on other tools, such as Qfang for forecasting reaction conditions, and an AI agent named Clio that incorporates Retrochimera, Segler emphasizes.