Chemical reaction yield prediction
![Project image](https://blackthorn.ai/wp-content/uploads/2023/03/CHEM3.png)
Business Goals
- Develop an application for estimating the chemical yield of the target product for unexplored organic reactions.
Challenge
- Typically, product yield estimations are empirical, i.e., based on synthetic chemists’ experience. Therefore, there is no reliable way to predict yields for new chemical reactions.
Results
- Our approach outperforms or equals all known state-of-the-art AI models predicting chemical reaction yields on public datasets:
- R^2 = 0.86, RMSE = 1.35 on Suzuki-Miyaura [https://www.science.org/doi/10.1126/science.aap9112]
- R^2 = 0.93 on Buchwald-Hartwig [https://www.science.org/doi/10.1126/science.aar5169]
Implementation Details
- A novel graph neural network architecture (RD-MPNN) was designed for organic reaction yield prediction. The network combines structural information, molecular-, and reaction-level descriptors. The network's performance was compared with the performance of other machine learning models on the same data and targets: linear models, decision tree ensembles, fully-connected feed-forward, and transformer networks. The RD-MPNN outperformed all the listed approaches in single- and multi-reaction class settings.
Industry
Service
Keywords
- Pharmaceuticals
- Drug Discovery
Roadmap
/*=
$user_is_authed
? declense_numeral(get_field('duration'), 'month', 'months')
: 'X months';
*/ ?>
Literature review, solution design, graph neural network PoC. (1 month)
Core development (2 months)
Design and implementation of alternative approaches as references (1 month)
Reproducing public benchmarks and comparing with the RD-MPNN (1 month)
Sign up to receive the project description
Want to talk?
![](https://blackthorn.ai/wp-content/themes/blackthorn/assets/dist/img/team/michael.webp)
Michael Gurbych
Director,
Operations and Finance
Operations and Finance
Roadmap
/*=
$user_is_authed
? declense_numeral(get_field('duration'), 'month', 'months')
: 'X months';
*/ ?>
Literature review, solution design, graph neural network PoC. (1 month)
Core development (2 months)
Design and implementation of alternative approaches as references (1 month)
Reproducing public benchmarks and comparing with the RD-MPNN (1 month)