Accelerating early stage Drug Discovery with low-data Generative-AI for potency and ADMET optimisation

Drug Discovery (DD) involves costly-and-lengthy experimental cycles to design novel drugs, synthesise them and then test their properties (e.g.toxicity/absorption/target-binding). DD success rate is extremely low: only 1:10,000 clinical-trial drug-candidates are FDA approved\[Boyd,2021\].

Predicting drug-properties is key in de-risking of assets before expensive drug-synthesis/testing. ~1/3 of drug development is dedicated to DD \[NatureBiotechnology,2018\]. Greater modelling accuracy in drug-characteristics and drug-candidate selection will accelerate time-to-first-in-human-trials by ~3-years; even a 10% prediction improvement could save billions \[Deloitte,2019\].

AI has the potential to accelerate drug-candidate optimisation with an estimated 25-50% speed increase and ~£25M potential savings per active drug programme \[WellcomeTrust,2023\]. However, AI-powered drug-design requires extensive data curation/generation and technical know-how. Additionally, designing truly novel-drugs with AI that satisfy medicinal chemists remains an unsolved problem and modelling drug potency is often difficult as not much data is available in the early stages of a drug programme.

**DeepMirror's innovative solution will generate drug-like molecules for the early stages of drug programmes where limited data exists, through an easy-to-use application.** This increases agency of R&D teams by reducing reliance on in-house AI-teams or partnerships with AI-companies. DeepMirror's app is estimated to accelerate drug-optimisation 2-4-fold, \[DeepMirror,2023\] and has already been validated in a case study with Tes Pharma, where the platform sped up DD by 6x (AppendixQ5).

**This project will develop three novel, IP-generating technologies following extensive user feedback of the current platform:**

**Low-Data-AI**: A novel end-to-end hybrid technology that combines physical simulations with AI to enable effective potency modelling in low data scenarios. This will enable users, regardless of their AI-expertise, to prioritise potent compounds with almost no input or technical jargon.

**Generative-AI Algorithm:** An advanced algorithm to develop bespoke, novel drug-molecules given specific conditions that satisfies medicinal chemistry curiosity (e.g. molecule-diversity/stability/chemical group)

**Certified Models:** A library of AI models trained of a mixture of private and public data that are easily finetuned onto specific use cases in an automated fashion and further reduce user data requirements.

These innovations will accelerate generation of novel drug-candidates, reduce experimental data-requirements, provide actionable AI predictions, streamline internal R&D processes and improve accessibility, speed and accuracy of AI in DD.

More efficiently iterating suitable molecule candidates will reduce R&D time, thereby increasing sector productivity.

Looking ahead, AI-accelerated DD could improve patient outcomes and provide increased efficiencies in identifying novel treatments for currently untreatable diseases and provide patients with accelerated access to new treatment modalities.