In the early 2000s pharmaceutical drug discovery was beginning to use computational approaches for ADME/Tox prediction in an effort to reduce the risk of later stage failures. Much has been written in the intervening twenty-plus years and significant expenditure has occurred in companies developing these in silico capabilities. It is therefore an appropriate time to assess where these tools can fit in today’s generative AI paradigm for drug discovery.

Significant quantities of ADME data are collected throughout early discovery to inform molecular designs and mitigate risk. Quantitative structure-property relationship (QSPR) models aim to extract maximal value from these datasets by learning relationships between molecular structures and the ADME properties of interest. We present findings from a study focused on modeling historical ADME datasets with multitask neural networks, using both fully internal datasets as well as hybrid internal/external datasets.

Recursion’s integrated operating system combines proprietary in-house data generation and advanced computational tools to generate novel insights to initiate and accelerate programs. Using our platform we follow a mapping and navigating approach that enables us not only to unravel the complexity of biology but also to identify chemical starting-points and drive SAR. Following this novel approach we efficiently advance projects from initiation through different stages of pre-clinical development.

In silico modeling has aided drug development for nearly 50 years, but remains handicapped by speed/throughput and predictability of properties that lead to drug success. We're combining both physics-based modeling and machine learning to create an end-to-end drug discovery pipeline comprising generation and filtering of new chemical entities for a target in days. Our tools have outperformed other publicly-known tools on benchmarks and have successfully identified true binders from false positives for JAK2.

This will talk will describe the use of active learning and free energy perturbation (FEP) to optimize chemical series on 2 active drug discovery programs. Using generative design we are able to generate a large set of analogs for any chemical series, and identify the highest probability ideas using FEP and ML ADME models. This approach is the first described to combine AI and established structure-based methods to accelerate optimization of a chemical series.

We used a digital assay and generative chemistry to improve the physicochemical properties of our beyond-rule-of-5 lead, increasing oral exposure by a million-fold. The clinical WRN inhibitor HRO761 has the best physico-chemical profile of all de novo designed oral drugs above 700 Da, resulting in excellent human pharmacokinetics.

Iambic has created a cutting-edge AI-driven platform to tackle the most challenging design problems in drug discovery and address unmet patient need. Our platform enables us to widely explore chemical space, while also sampling a wide range of target product profiles. We have demonstrated our platform on initial programs, with our first scheduled for clinical entry just two years after launch.

Traditional drug discovery platforms, developed by technical experts, often lack user-friendly designs and the expertise of medicinal chemists. CEEK-CURE, a novel medicinal chemistry-centric explainable AI (XAI) platform, bridges this gap. In this presentation, we will demonstrate the transformative impact of a medicinal chemistry-centric AI platform on enhancing hit rates and selectivity profiles. We will showcase two different projects focusing on oncology and neuropathic pain targets.

Rapidly evolving computer hardware, software, and machine learning algorithms, rapidly growing databases related to chemical compounds, biomolecules and biomedicine offer a unique opportunity to dramatically accelerate lead discovery for unmet medical needs, rare and neglected diseases, and emerging threats. We will describe the recent advances in searching billions of compounds for challenging tasks and new targets by using combining large-scale docking, modeling, GPU-algorithms, with AI and machine learning in one pipeline.

With the surge in targeted protein degradation as a therapeutic strategy, there is an increasing demand for novel molecules capable of selectively targeting and degrading disease-associated proteins. Cyclin E and CDK2, key regulators of the cell cycle, have been implicated in various malignancies and represent promising therapeutic targets. We describe the prospective design of degrader molecules with high specificity for Cyclin E and CDK2 enabled by the Auto/dx platform.

At Serna Bio, we’re investigating the potential of AI to rapidly accelerate the discovery and development of small molecule modulators of RNA function. To train our ML models, we’ve generated a proprietary dataset of ~2.5 million RNA-small molecule binding data points. Using these models, we can computationally learn features of small molecules that bind to different RNA motifs and identify distinct chemical features in different subsets of RNA binders.

Since its inception, Atomic AI has made substantial advances in its AI platform for RNA-targeted small molecule drug discovery. These advances have been enabled by large-scale in-house data collection and the development of new ML models. The talk describes ATOM-1, a large language model trained on billions of experimental data points, and its applications in drug discovery at Atomic AI.

This presentation will highlight the Koliber AI platform's progress in minimizing input dataset requirements for predicting peptide and protein properties such as potency, stability and permeability. We will explore a range of applications, including anti-microbial and immune-modulating peptides, as well as various datasets containing non-canonical amino acids and cyclic peptides. We will present examples demonstrating de novo predictions of substitutions for peptides and proteins that influence potency and substrate specificity.