Forward synthesis analysis

In the context of ASKCOS, we generally view forward synthesis analysis as supplementary studies for one-step retrosynthetic suggestions. Accessible from the side bar under Forward Synthesis, such analyses include:

• Condition Recommendation which suggests best reaction contexts given the reactant(s) and the target
• Product Prediction which predicts the major product given the reactants(s)
• Impurity Prediction which predicts impurities from different modes like over-reaction
• Regioselectivity Prediction
• Site selectivity prediction for Aromatic C-H Functionalization reactions.

In general, major Product Prediction is the most well-defined task with the most quantitatively robust models, easily reaching >80% top-1 accuracy on common benchmarking datasets. In contrast, the other four are more specialized models trained with smaller or noisier data, and the results should be viewed with more of a grain of salt than definitive.

Let's do a quick run of Product Prediction.

1. Enter CC(=O)O.CCCO as the Reactants, the SMILES for the mixture of acetic acid and propan-1-ol
2. Click on settings>model selections and check the following are set:
   • Condition Recommendation Model = Neural Network
   • Forward prediction model = wldn5
   • Forward model training set = pistachio
3. Click GET RESULTS

The Product Prediction result table will be populated shortly after.

In the results, the esterification product is predicted as the top-1, with a high probability of 0.9404. Depending on the exact model used, the top products may or may not make chemical sense. More details on the different models, as well as on the other forward synthesis analysis tasks will be discussed in [04-More forward synthesis analysis examples](04-More forward synthesis analysis examples).