4.8. Retrosynthesis score (SynthScore)
The retrosynthesis score predicts the feasibility of synthesis of a compound from 0-2, with 2 being the most likely to be synthesizable. The model is built from the results of over 400,000 compounds evaluated by AiZynthFinder (Copyright 2020 Samuel Genheden and Esben Bjerrum). AiZynthFinder is a tool for retrosynthetic planning. The default algorithm is based on a Monte Carlo tree search that recursively breaks down a molecule to purchasable precursors. The tree search is guided by a policy that suggests possible precursors by utilizing a neural network trained on a library of known reaction templates. AiZynthFinder returns several pieces of information based on the input of a SMILES string, including whether a retrosynthesis route can be solved or not and a “top score” from 0-1, which relates to the solvability; in practice, top scores of > 0.93 tend to be solvable by AiZynthFinder.
We analyzed > 400,000 unique compounds from various internal and public sources in AiZynthFinder and created an ANNE classification model from the solved/not solved results. This [0,1] binary output was not as useful as a continuous score between 0 and 1. Since each ANN classification model has a single logistic output node it was easy to take the averaged raw output of all ANNE nodes and add the resulting numerical values to the “top score” value, also 0-1, for each compound, resulting in a range of 0-2. An ANNE regression model was then built on this value for the > 400,000 compounds. The performance metrics are shown below.
We sent a set of novel compounds designed in the AIDD module for various projects to several different chemical synthesis vendors and created a database of the average price and number of steps for synthesis that were proposed for each compound. 29 compounds were deemed to be synthesizable by at least one of the companies. We then performed a correlation between the Retrosynthesis score and the average price or average number of steps. We found a high degree of correlation for both (r2 = 0.565, SRCC = -0.691 for the price and r2 = 0.524, SRCC = -0.776 for the number of steps).