Drug Discovery Cost Reduction with Uncertainty-Guided Predictions

Drug Discovery Cost Reduction with Uncertainty-Guided Predictions

Key Outcomes

75% Reduction in drug discovery costs
10x Faster discovery
60% Less training data

Context

The global pharmaceutical industry is a 1T market that has been experiencing a development decrease for the past two decades. Large pharmaceutical companies are spending nearly 80B per year to develop fewer successful drugs. The return for every dollar invested in research and development decreased from 10 cents in 2010 to 2 cents today. Roughly 1 out of 10 drugs are approved after clinical trials with the R&D process for each drug costing approximately 2.17B, compared to 1.19B in 2010, and an average time from discovery to launch of 10-12 years.

AI Drug Discovery is an emerging sector that intends to address these shortcomings. In fact, 72% of professionals think the industry is behind in AI development and 44% cite lack of talent as the main barrier. This type of technology can in some cases speed up the discovery stages by a factor of 15. The size of this market increased from 200M in 2016, to 1.5B in 2019, and is expected to reach 7.1B by 2030 with a CAGR of 23.72%. Moreover, at least 20 partnerships have been reported between pharmaceutical companies and AI drug discovery startups. Additionally, some of the key incumbents, such as Pfizer, GlaxoSmithKline, and Novartis, are developing in-house AI expertise.

Our Solution

Themis AI’s technologies are based on over five years of foundational research carried out at MIT CSAIL that led to proprietary advancements in uncertainty estimation. These techniques have been successfully applied for molecular property prediction and discovery. In 2021, evidential deep learning was used to demonstrate a new approach to uncertainty quantification for neural network-based molecular structure−property prediction at no additional computational cost. The approach enabled calibrated predictions where uncertainty correlated with error, allowed for sample-efficient training through uncertainty-guided active learning, and improved experimental validation rates.

Additional Resources

A. P. Soleimany, A. Amini, S. Goldman, D. Rus, S. N. Bhatia, and C. W. Coley, “Evidential Deep Learning for Guided Molecular Property Prediction and Discovery,” ACS Central Science, vol. 7, no. 8, pp. 1356–1367, Jul. 2021..

Q. Xu, E. Ahmadi, A. Amini, D. Rus, and A. W. Lo, “Identifying and Mitigating Potential Biases in Predicting Drug Approvals,” Drug Safety, vol. 45, no. 5, pp. 521–533, May 2022.