Predicting the formation of NADES using a transformer-based model

What if we could predict new green solvents the way we predict the next word in a sentence? A transformer-based AI model is doing just that for Natural Deep Eutectic Solvents (NADES), unlocking a treasure trove of 337 new mixtures.

Natural Deep Eutectic Solvents, or NADES, are the rising stars of green chemistry—promising alternatives to traditional organic solvents in pharmaceuticals, agriculture, and food. But finding new ones has been a slow, empirical grind, often just tweaking known recipes. Enter a transformer-based neural network, the same architecture behind language AI, now trained to read chemical structures as if they were sentences.

The model was first pre-trained on vast unlabeled chemical data using SMILES notation, then fine-tuned to classify whether a mixture forms a stable NADES or not. This clever adaptation from language learning means it needs only small datasets and modest computing power. The result? The algorithm predicted 337 new stable eutectic mixtures from a database of natural compounds.

Even more impressive, it can suggest the exact components and molar ratios needed to create NADES with entirely new molecules—not just those in its training data. This was validated using known NADES and by crafting novel solvents containing ibuprofen. The potential is transformative: instead of treating bioactive compounds as mere solutes, they could become functional parts of liquid formulations, streamlining drug development and green chemistry alike.

Key Points

• Transformer AI predicts 337 new NADES from natural compounds.
• Pre-trained on SMILES, fine-tuned for binary classification.
• Predicts components and molar ratios for new NADES, validated with ibuprofen.