By: Bohdan Waszkowycz, Senior Principal Scientist; Ting Qin, Senior Principal Scientist; Jack Hoffman, Senior Scientist 2, Sygnature Discovery
AI and ML technologies have become firmly established across all stages of drug discovery, delivering proven benefits in accelerating timeframes and de-risking the expensive journey from early-stage drug design through to clinical candidate. This has been exemplified by early clinical success stories from biotech companies with pioneering AI platforms, including InSilico Medicine, Recursion/ExScientia and Schrödinger, as well as recent collaborations between large pharma and AI developers.
CROs supporting early-stage drug discovery projects have been quick to establish their own internal AI platforms to address the needs of their clients in a rapidly evolving field. This brings its own challenges, as CROs typically work with a wide range of clients, from virtual companies and small biotechs through to large pharma, each with their own internal discovery platforms and preferred ways of working. CROs aim to offer state-of-the-art experimental and computational technologies at every stage of the drug discovery pipeline, typically encompassing target validation, hit identification, hit-to-lead and late-stage lead optimisation.
A CRO requires an AI platform to be adaptable and extendable, offering pragmatic and cost-effective solutions depending on the specific goals of each client. Examples of particular focus areas for AI include the use of generative molecular design to discover new hit matter and accelerate the subsequent Design-Make-Test-Analyse (DMTA) cycles which drive lead optimisation. As synthetic chemistry is often a bottleneck in early discovery, AI and ML can reduce timescales to reach clinical development by selecting compounds for synthesis with the greatest chance of success.
A further consideration for a CRO is ensuring effortless data sharing across internal and external project teams, which is essential to support timely, data-driven decision-making. This requires adherence to the highest standards of data security and integrity (e.g. ISO27001). In some instances, this may require offering solutions hosted on internal, firewall-protected servers as well as via external cloud platforms.
From our experience as a CRO, a key learning is that no single platform available in the market currently, can address every project requirement – rather, success depends on having a wide range of tightly integrated technologies, closely coupled with human expertise and creativity, to adapt to the specific challenges that each project brings.
Computational Approaches in Drug Design
Computer-aided drug design (CADD) has had a long history in drug discovery, pre-dating the recent boom in AI and ML technologies. The goal has always been to offer a rational and focused approach to drug discovery by exploiting whatever data is available to guide the design process. Typically, two broad categories of approaches are commonly pursued:
- Structure-based drug design – exploiting experimental or modelled structures of the target protein or nucleic acid to guide the design and optimization of ligands – for example using X-ray or cryo-EM structures of the protein-ligand complex as the basis for docking and molecular dynamics (MD) simulations
- Ligand-based drug design – exploiting data from ligand structure-activity relationships (SAR) to infer the key features for binding (i.e. the pharmacophore), designing new analogues (scaffold hopping, library design) and building predictive models for affinity and ADMET-related properties (quantitative structure-property/activity relationships, QSAR/QSPR)
In recent years, these methods have been expanded in scope by wide-ranging advances in AI and ML, as well as accompanying developments in high performance computing to enable large-scale calculations to be routinely performed. This extends the impact of computational methods, which will be discussed in the following sections.
Generative Molecular Design
The application of Generative AI (GenAI) to molecular design has become a widespread and effective approach to address the challenge of designing novel chemical structures with appropriate druglike properties. Compared to traditional computational approaches, GenAI offers the advantages of accessing a vast chemical space, much greater than the pool of commercially available compounds routinely used for virtual screening. Importantly, GenAI methods also allow for focussing of design towards desired goals, enabling multiparameter optimisation of affinity, selectivity and ADMET profile, which is a key requirement of successful DMTA cycles. However, the practical exploitation of GenAI comes with its own considerations and challenges, which are explored below.
Balancing chemical diversity with synthetic pragmatism
Although GenAI methods allow extensive sampling of a potentially limitless chemical space, this is not without some serious drawbacks in terms of project timelines. Depending on the AI software used, the designs generated can be non-druglike in terms of unsuitable property profiles or the presence of unstable (even nonsensical) chemistry. Some of this can be resolved by extensive triaging of designs by standard cheminformatics workflows. A more significant limitation is synthetic tractability – a fast-moving hit-finding project cannot afford to be slowed down by extensive multistep syntheses of a large set of diverse, speculative designs. This can be addressed to some extent by introducing retrosynthetic analysis or calculated accessibility scores as part of the triage process. Some well known synthetic accessibility scores can be used to filter out complex synthetic routes.1 However, because individual projects often depend on project specific intermediates to make final compounds, project dependent synthetic feasibility becomes more relevant with the project progressing. Consequently, an AI platform that incorporates iterative human in the loop interaction, particularly involving bench chemists who design compounds with real time consideration of synthetic feasibility, is important for effectively guiding drug design.
An additional approach gaining ground is to build some synthetic awareness into the GenAI, typically in terms of training the generator on known synthetic chemistry routes and providing it with a database of commercial building blocks. In this way, the GenAI is constrained to search a narrower chemical space but one in which there is a much greater chance that designs can be readily synthesized. This approach often represents the best balance between sampling broad chemical novelty while minimizing the overall synthetic cost. In practice, we have found it to be a reliable approach for rapidly expanding SAR around early hits and accelerating lead optimization through focused library design, where limiting design to a small number of robust synthetic routes has a direct impact on the turnaround and throughput of the DMTA cycle.
Multiparameter optimization - integrating ML into the design process
Successful drug design always requires the balance of multiple desired properties – potency alone is never sufficient without appropriate selectivity, physicochemical properties and ADMET profile. GenAI can tackle this challenge by integrating ML models into the design process, effectively driving the design towards the desired region of property space.
As with any other application of ML, success ultimately depends on a model’s ability to generalise to unseen data. Assessing this reliably requires careful consideration of aspects such as data curation, dataset splitting and the adoption of statistically rigorous protocols for method comparison.2 Efforts should also continue after model deployment, for example through quantifying the uncertainty of predictions and monitoring dataset drift.
From our experience, approaches for explaining model predictions, such as SHAP analysis, as well as data-visualisation tools, such as UMAP, can be powerful in combination with human expertise. By highlighting the patterns a model has learned from its training data, these tools help to better understand the strengths and limitations of our models and support the creation of new hypotheses to drive projects forward.
Typically, models are trained for potency against the desired target and against one or more antitargets (to address selectivity), as well as ADMET-related properties to give a handle on permeability, clearance and metabolic stability. These models may be built on internal project data (“local” models, e.g. for potency relating to the current chemical series) or on external data (“global” models, commonly for ADMET properties). In the early stages of a project, model building is often not straightforward - there may be too few internal datapoints to build a robust local model, while public domain databases (e.g. ChEMBL) present issues of data standardization across diverse assay platforms. We have found that a few well-chosen properties (e.g. LogD and solubility) can usually be very effective in driving design into a more successful property space, even when the desired ADMET endpoints prove difficult to model.
A further use of MPO to guide design is to incorporate 3D scores, e.g. 3D similarity to a reference ligand or a docking score into the protein structure of interest. Reinforcement learning methods allow the GenAI stage to learn how 2D chemical structure relates to 3D features, enabling the modeller to drive design towards a feature of interest (e.g. a particular polar or lipophilic contact in the protein binding site) or to ensure that 2D designs are in a similar 3D space (in terms of shape and pharmacophores) as the existing chemical series.
Generative molecular design in practice
A number of software platforms for GenAI are available, mostly in the public domain and often requiring some technical programming expertise to install and run. REINVENT (AstraZeneca) is one of the most well-known examples that we have implemented in house. Of the commercial offerings, we have extensive experience of Makya from Iktos, which covers all the above features of MPO- and retrosynthesis-driven design, including 3D shape-based and docking-based scoring, accessible from an easy-to-use graphical interface. It is worth emphasizing that GenAI typically generates many thousands of designs, so extensive triaging of the output is an essential step to ensure the best candidates are selected for synthesis. This typically requires close collaboration with chemoinformaticians (for property filtering and ML model building) and molecular modellers (for additional 3D analysis in terms of docking, molecular dynamics or prediction of free energy of binding, which can give more robust predictions of activity but at significant computational cost). This further emphasizes the multidisciplinary environment that is essential for efficient drug design.
AI-Enabled Drug Discovery Platform Example - SygDesign
To address the challenges of a AI platform within a CRO environment, we leverage the well-established Electronic Laboratory Notebook (ELN) framework to build an in-house, AI enabled drug discovery platform, namely SygDesign.3 It is an in house platform that transforms complex, heterogeneous data into high quality insights to support drug discovery programmes. It incorporates a variety of supervised learning techniques, ranging from ‘conventional’ machine learning architectures (linear regression, random forest, xgboost, etc.) as baselines, to modern and more complex approaches including MPNNs and transformer-based systems. These models may make use of public data and/or a project’s own data to capture local structure-activity-relationships, and, where feasible, exploit information from broader domains. These models may either be used directly by scientists to make predictions, or in conjunction with reinforcement learning to generate new designs. Operating securely behind the company firewall, it adheres to FAIR principles (Findable, Accessible, Interoperable, and Reusable). By embedding synthetic feasibility and medicinal chemistry expertise, the platform empowers bench chemists with state-of-the-art AI capabilities that accelerate discovery.
Summary
As a CRO, we need to build a broad and robust AI/ML platform to support wide-ranging client needs. The platform has to cover:
- A real-world AI platform should transition from disconnected data repositories and isolated AI workflows to a centralized, integrated system that enables reusable, flexible, and project configurable workflows, addressing the diverse requirements of multiple projects within a CRO.
- As AI is a rapidly evolving field, with new algorithms and models being published frequently, validation and deployment of new model methods through an open and flexible platform are necessary to bring cutting edge technologies into the drug discovery process.
- AI models should be transparent and interpretable, enabling understanding of structure–property relationships and informing subsequent iterations of compound design.
- The platform should be intuitive and easy to use for chemists, enabling an expert in the loop workflow. This allows chemists to design compounds with real time consideration of synthetic feasibility and to actively contribute to the drug design process through their practical expertise.
- The platform should be established as an open, extensible foundation that enables seamless integration with external platforms through robust APIs and complementary interfaces. It should provide a scalable foundational framework capable of both leveraging and advancing alongside evolving AI technologies, including emerging paradigms such as agentic AI. By establishing an interoperable ecosystem, the platform should serve as a long-term innovation backbone, continuously expanding to support both current and future frontiers in drug discovery, including but not limited to target validation4 and peptide design.
The overarching goal is to achieve close integration of AI/ML tools into the DMTA cycle – we need to apply the best tool appropriate for each task to speed up the DMTA cycle and reduce the number of cycles to reach our goal. Although no predictive methods are perfectly accurate, over time they improve the quality of the molecules that are designed and hence improve the chance of clinical success.
References
- Ertl, P. and Schuffenhauer, A., 2009. Estimation of synthetic accessibility score of drug-like molecules based on molecular complexity and fragment contributions. Journal of cheminformatics, 1(1), 8.
- Ash, J.R., Wognum, C., Rodríguez-Pérez, R., Aldeghi, M., Cheng, A.C., Clevert, D.A., Engkvist, O., Fang, C., Price, D.J., Hughes-Oliver, J.M. and Walters, W.P., 2025. Practically significant method comparison protocols for machine learning in small molecule drug discovery. Journal of chemical information and modeling, 65(18), pp.9398-9411.
- Qin T, Chandrasekaran A, Hoffman J, Shiers J, Jain T, Sambrook Smith C, Empowering chemists in drug design: Delivering AI solutions through an ELN framework at the enterprise level, SLAS Technology, Volume 37, March 2026, 100392
- Rosenstock, T.R., Ohzeki, H., Kumagai, S., Suda, N. and Smith, C.S., 2026. Multidisciplinary review method for novel target identification and prioritization for neurodegenerative diseases. Drug Discovery Today, 104671.
About the Authors
Bohdan Waszkowycz is a Senior Principal Scientist at Sygnature Discovery. As a computational chemist, he supports multiple client projects, with a particular interest in structure-based drug design and its integration with generative AI. With a PhD in theoretical chemistry from the University of Manchester, he has worked in the pharmaceutical industry for over 35 years in a range of biotech and CRO companies including C4X Discovery, Cancer Research UK, Argenta Discovery and Tularik.
Ting Qin is a Senior Principal Scientist at Sygnature Discovery. Ting leads an AI team developing SygDesign, an AI-enabled drug discovery platform. His team leverage the ELN framework to ensure SygDesign adheres to FAIR data principles, making data Findable, Accessible, Interoperable, and Reusable. Ting holds a PhD in computational chemistry from the University of Oxford and has over a decade of experience in the drug discovery industry. He has publications in the fields of artificial intelligence and lab digitalization, presenting practical solutions that support and accelerate the drug discovery process.
Jack Hoffman is a Senior Scientist 2 at Sygnature Discovery, where he develops automated systems to support drug discovery across scientific domains. As part of the SygDesign team, he focuses on the design and deployment of tools for use by both scientists and AI agents. Jack holds an MChem from the University of Oxford and has contributed to projects spanning machine learning, cheminformatics andbioinformatics.