Meetings

Autumn UK-QSAR 2024 Meeting

Exscientia is very happy to welcome you to the Autumn 2024 UK-QSAR meeting, to be held on 17th October 2024 at Lady Margaret Hall, Oxford, UK.

Exscientia Logo

As always, the event is free of charge to attend. However it will be held in person and will not be streamed online or recorded for later access.

Registration

The registration is now closed.

Posters

Abstract submission is now closed.

Please follow these links to download the posters list and abstracts.

All posters need to be pre-printed as A0 (33.1 inches x 46.8 inches / 841 mm x 1189 mm) or A1 size (23.4 x 33.1 inches / 594 x 841 mm).


Agenda

09:00-10:00Open registration, coffee/tea
10:00-10:15Welcome and opening remarks
Session 1: Generative AI in drug discovery
10:15-10:45Morgan Thomas (Universitat Pompeu Fabra) – “Structure-based generative molecular design: from 2D or 3D?”
10:45-11:15José Jiménez-Luna (Microsoft Research) – “Learning chemical intuition from chemist on the loop”
11:15-11:45Finton Sirockin (Novartis) – “Generative Chemistry at Novartis: a view from the trenches”
11:45-12:05Marcel Verdonk and Noj Malcolm (Astex Pharmaceuticals & Schrodinger) – “Richard Hall: A reflection on a life cut short”
12:05-13:30Lunch and poster session
Session 2: Machine Learning in drug discovery
13:30-14:00Liam Wilbraham (Exscientia) – “Active learning on synthons for molecular design”
14:00-14:30Lindsay Willmore (Google DeepMind) – “AlphaFold 3: What’s new and improved and what’s left to do”
14:30-15:00Break and poster session
Session 3: Computational toxicology
15:00-15:30Vigneshwari Subramanian (AstraZeneca) – “Predicting safety liabilities of small molecules with machine learning: An industry perspective”
15:30-16:00Layla Hosseini-Gerami (Ignota Labs) – “Harnessing AI for Mechanistic Insights into Drug Failures: A Path to Salvageability”
16:00-16:30Daniella Hares (Institute of CancerResearch) – “Enhancing Small Molecule Binding through Computational Analysis of Water Networks” Poster winner of the Spring 2024 event
16:30-17:00Concluding remarks, poster prize winner announcement and close


Abstracts

Structure-based generative molecular design: from 2D or 3D?
Morgan Thomas, Universitat Pompeu Fabra

Several approaches can be taken on the path to the automated design of small molecules as candidate drugs that bind to therapeutically relevant protein structures. ML-based generative molecular design approaches can be mostly reduced to structure-implicit or structure-explicit, i.e., molecule building in 2D or in 3D. Recent comparisons between these approaches have highlighted limitations with regards to protein-ligand interactions and ligand strain. In this work this comparison has been extended to also compare the quality of chemistry proposed, revealing that structure-implicit generative design from 2D remains dominant. This re-iterates the challenges that 3D models still face including training data quantity, quality and required aspects such as novelty. Furthermore, this poses questions with regards to research direction in the field, and the quality of evaluations currently published. As an example of successful structure-based generative design from 3D, we have used reinforcement learning and molecular docking utilising domain expertise to identify novel hits against a GPCR target.

Learning chemical intuition from chemist on the loop
José Jiménez-Luna, Microsoft Research

The lead optimization process in drug discovery campaigns is an arduous endeavour where the input of many medicinal chemists is weighed in order to reach a desired molecular property profile. Building the expertise to successfully drive such projects collaboratively is a very time-consuming process that typically spans many years within a chemist’s career. In this talk I’ll show how we replicated this process by applying artificial intelligence learning-to-rank techniques on feedback that was obtained from 35 chemists at Novartis over the course of several months. I’ll also exemplify the usefulness of the learned proxies in routine tasks such as compound prioritization, motif rationalization, and biased de novo drug design.

Generative Chemistry at Novartis: a view from the trenches
Finton Sirockin, Novartis

This presentation will discuss the integration of Generative Chemistry (GenChem) methods into medicinal chemistry projects at Novartis. Specifically, it will delve into the lessons learned over the past years. The presentation will emphasise the practical aspects of integrating GenChem methods into the day-to-day ideation and prioritisation processes of medicinal chemistry teams.

Richard Hall: A reflection on a life cut short
Marcel Verdonk and Noj Malcolm, Astex Pharmaceuticals & Schrodinger

Personal and professional tribute to Richard Hall, who died 15th December 2023.

Active learning on synthons for molecular design
Liam Wilbraham, Exscientia

Generative design and library enumeration are two important compound generation strategies in computational small molecule drug discovery. The former aims to refine small molecule structures to align with predefined multi-parameter objectives while maintaining sample efficiency but can lead to compounds that are challenging to synthesize due to the absence of explicit chemistry constraints. Library enumeration strategies apply common medicinal chemistry reactions to commercially-available building blocks, creating a synthetically-tractable chemical space. However, as this method is frequently used to generate compounds combinatorially, scoring the enumerated library beyond targeted enumeration schemes can become computationally intractable. To achieve the sample efficiency enjoyed by generative approaches while maintaining the synthetic tractability offered by library enumeration, we introduce Synthon Active Learning (SAL). SAL bridges both methods via a factored active-learning approach, building models to rank synthons according to a multi-parameter objective, the best of which are exhaustively enumerated to create high-scoring, synthesizable molecules. We demonstrate SAL’s efficiency experimentally: by evaluating just 6% of a space containing 1M molecules, we recover 82% and 100% of the top 100 molecules for a docking and 3D similarity objective, respectively. Further, we show our method scales beyond the domain of exhaustive screening, finding increasingly high-scoring molecules in spaces of up to 10^12 molecules. Finally, we validate SAL for molecule design tasks on three distinct protein targets using ligand-based, structure-based, and drug-likeness scoring functions. We show that SAL-generated molecules obtain comparable or better ADMET, synthesizability, and chemical property distributions compared to known bioactive compounds while successfully optimising for the target objective.

AlphaFold 3: What’s new and improved and what’s left to do
Lindsay Willmore, Google DeepMind

AlphaFold 3 represents a significant advancement in the field of structural biology, enabling the simultaneous prediction of complex structures involving diverse biological components such as proteins, nucleic acids, small molecules, ions, and modified residues. Furthermore, the model has improved accuracy in predicting the structures of protein-ligand and protein-nucleic acid interactions compared to specialized tools. I will take you through the model updates and training schemes used to achieve this enhanced model coverage and accuracy. And I will also discuss some of the unexpected wins and shortcomings of the new AlphaFold model.

Predicting safety liabilities of small molecules with machine learning: An industry perspective
Vigneshwari Subramanian, AstraZeneca, Gothenburg, Sweden

Ensuring right safety has been a part of AstraZeneca’s 5R framework for the past decade and adoption of this framework has had a significant impact on improving our project success rates and reducing drug attrition. Screening cascades involving in vitro and in vivo assays have been developed to assess safety liabilities, however, in silico approaches have also been exploited extensively to support decision-making in projects. Having the potential to predict safety-related end points early on would not only prevent compounds with safety liabilities progressing further, but also support the design of molecules with improved safety profiles. This talk will primarily focus on our machine learning modeling efforts to support safety and organ-level toxicity assessments, including predicting off-target pharmacology, phospholipidosis, Drug Induced Liver Injury and QTc prolongation (a potential indicator of cardiac toxicity).

Harnessing AI for Mechanistic Insights into Drug Failures: A Path to Salvageability
Layla Hosseini-Gerami, Ignota Labs

The pharmaceutical industry witnesses the attrition of over a thousand drug candidates annually during pre-clinical or Phase I studies, due to safety concerns. The limited mechanistic understanding of these failures poses a significant hurdle in diagnosing root causes and identifying potential avenues for salvaging these compounds. This talk will explore the innovative approach developed by Ignota Labs, which leverages artificial intelligence to transform the landscape of drug development by providing mechanistic insights into drug failures. Ignota Labs employs a causal knowledge graph methodology, integrating multimodal data sources to create a comprehensive, granular understanding of drug actions and their associated adverse effects. This  approach enables the generation of detailed hypotheses about the mechanisms underpinning observed safety issues. By systematically analysing data from various biological, chemical, and clinical domains, the AI-driven platform can identify critical causal events and suggest modifications to improve the safety profile of drug candidates. The presentation will delve into the technical aspects of the causal knowledge graph, highlighting how we process and integrate diverse datasets to construct a detailed mechanistic map. A case study will be presented to illustrate the practical application of this methodology, demonstrating how our approach can diagnose specific mechanistic failures and propose actionable strategies for drug redesign. By elucidating the root causes of drug failures, this AI-based approach not only enhances the understanding of drug safety but also opens new pathways for rescuing potentially valuable therapeutics. This talk aims to inspire new thinking and innovation in drug development, giving a glimpse into a future where AI-driven insights significantly reduce attrition rates by offering a second chance to promising drug candidates.

Enhancing Small Molecule Binding through Computational Analysis of Water Networks
Daniella Hares, Institute of Cancer Research

Water networks can have a critical role in small molecules binding to a target protein and are an important consideration in structure-based drug design. Previous computational studies have predicted protein-bound water molecules to contribute between -3 to -6 kcal mol-1 towards binding affinity [1]. When modifying the molecular structure of a lead compound, the rearrangement of surrounding water networks in the binding site can impact potency but this contribution is challenging to measure experimentally. Computational methods are ideally suited to study the interplay between ligand optimisation and water displacement by predicting the effect of structural changes on both the activity of the compound and the stability of neighbouring water molecules. We used Grand Canonical Monte Carlo simulations [2] and alchemical free energy calculations to retrospectively rationalise the trends in potency observed in a set of B-cell Lymphoma 6 (BCL6) inhibitors [3]. As part of the drug design process, the inhibitor molecular structure was modified to displace water molecules that were part of a network within the protein binding site. Using the BCL6 project as an example, this talk will demonstrate how  computational approaches can shine a light into an aspect that is often overlooked but essential to guide the design of better compounds when working with hydrated protein pockets. We show the power of these methods and encourage their use more widely, particularly in prospective applications on drug discovery projects.

[1] Mobley, D. L. and Dill K. A. (2009). Structure, 17 (4), 489–498.
[2] Ross, G. A., Bodnarchuk, M. S. and Essex, J. W. (2015). J. Am. Chem. Soc., 137 (47), 14930–14943.
[3] Lloyd, M. G., Huckvale, R., Cheung, K.-M. J., Rodrigues, M. J., Collie, G. W., Pierrat, O. A., Gatti Iou, M., Carter, M., Davis, O. A., McAndrew, P. C., Gunnell, E., Le Bihan, Y.-V., Talbot, R., Henley, A. T., Johnson, L. D., Hayes, A., Bright, M. D., Raynaud, F. I., Meniconi, M., Burke, R., van Montfort, R. L. M., Rossanese, O. W., Bellenie, B. R., Hoelder, S. (2021). J. Med. Chem., 64 (23), 17079–17097.