Accelerating compound synthesis in drug discovery: the role of digitalisation and automation

Abstract

The Design-Make-Test-Analyse (DMTA) cycle relies on efficient compound synthesis, yet the synthesis (“Make”) process remains a significant bottleneck, especially for complex molecules. This opinion letter explores how digitalisation and automation are accelerating the entire synthesis process. It details their current integration, from AI-powered synthesis planning and streamlined sourcing to automated reaction setup, monitoring, purification, and characterisation. FAIR data principles are emphasised as crucial for building robust predictive models and enabling interconnected workflows. Finally, the future of fully integrated, data-driven synthesis with tools like “Chemical ChatBots” and the evolving skill set required for medicinal chemists in this increasingly digital and automated landscape are discussed.

This opinion addresses how digitalisation and automation can reduce the synthesis bottleneck in the DMTA cycle. Current automated synthesis & planning, future data-rich integrated platforms, and the medicinal chemist's evolving role are reviewed.

1. Introduction

The discovery and optimisation of novel small-molecule drug candidates critically hinges on the efficiency of the Design-Make-Test-Analyse (DMTA) cycle.^1–5 This iterative process relies on rapid and reliable synthesis of a series of compounds for biological evaluation. Yet, the synthesis (“Make”) process often represents the most costly and lengthy part of the cycle. The inherent challenges are magnified when dealing with complex biological targets, which, in turn, can demand intricate chemical structures. This often necessitates multi-step synthetic routes that are labour-intensive and time-consuming, involving numerous variables in reaction scouting before the identification of a successful route. This opinion letter explores how the latest advancements in digitalisation and automation have the potential to address synthetic bottlenecks. We will also detail how these advancements are currently being integrated into the pharmaceutical industry and describe our view on future developments.

The Make step of the DMTA cycle includes synthesis planning, sourcing materials, reaction setup, monitoring, purification and characterisation as well as documentation (Fig. 1). All steps of this cycle require multiple manual operations performed by humans, generating an extensive amount of information that needs to be documented and processed. Therefore, failing to complete the synthesis process, and not obtaining the desired chemical matter for testing, inevitably wastes substantial resources. With the increase in digitalisation and automation there are several opportunities to accelerate the Make process by utilising the generated data and automating repetitive and error-prone tasks. As ever, there are many cross-overs between discovery and process research;⁶ however, we focus on solutions of most interest to medicinal chemistry concerning time and resource priorities. By highlighting previous and current synthesis approaches, this overview will lead to our perspective on the future, where the strategic development and application of digital and automation solutions are poised to accelerate the Make process and boost success rates.

Fig. 1. The compound synthesis (Make) process (r.) is part of the Design-Make-Test-Analyse (DMTA) cycle (l.). The Make process encompasses an integrated workflow, from initial synthesis planning, including literature analysis and reaction condition selection, to the sourcing of chemicals, managing inventory, and setting up reactions. Subsequent steps involve reaction monitoring, work-up, and purification, followed by analytics, registration, and documentation. Finally, the pure final analogues are handed over to the compound logistics team for submission to various profiling assays.

2. The compound synthesis process

2.1. Synthesis planning

On definition of the small molecule target structure synthesis planning involves plotting a multi-step sequence of possible chemical transformations. Historically, an overarching synthesis plan was established, applying standard reaction conditions to each discrete chemical step, sometimes necessitating lengthy optimisation. Today, a more holistic approach has been adopted, integrating more sophisticated tools and knowledge to plan specific reaction conditions, with a high probability of success, into the earliest planning stages. We believe this strategic shift substantially improves the efficiency of the compound synthesis process by identifying the most promising synthetic route from the outset.

The concept of retrosynthetic analysis, a framework formalised by E. J. Corey, has been a cornerstone of synthetic chemistry for decades.⁷ This approach involves the recursive deconstruction of a target molecule into simpler, commercially available precursors. Since the number of known reactions and potential pathways for any given molecule presents a combinatorial challenge that can overwhelm human comprehension, teaching a computer the logic of chemical synthesis route design was an obvious choice. Computer-Assisted Synthesis Planning (CASP) has evolved from early rule-based expert systems that were limited and manually curated into data-driven machine learning (ML) models.⁸ Modern CASP methodologies involve both single-step retrosynthesis prediction, which proposes individual disconnections, and multi-step synthesis planning, which chains these steps into a complete route using search algorithms like Monte Carlo Tree Search or A* Search.^7,9 While substantial progress has been made, an “evaluation gap” exists, where single-step model performance metrics do not always reflect overall route-finding success.¹⁰ These and other remaining bottlenecks,¹¹ including the prediction of viable reaction conditions and accurately handling complex stereochemistry and regioselectivity, have been comprehensively covered in recent reviews from Glorius, Grzybowski and co-workers,¹² and Guo and Schwaller.¹³

The pharmaceutical industry is actively utilising Artificial Intelligence (AI)-powered platforms for synthesis planning to generate valuable and innovative ideas for synthetic route design.¹⁴ While these tools excel in providing diverse potential transformations, unfortunately, the generated proposals are rarely ready-to-execute synthetic routes.¹⁵ These computational retrosynthesis tools are most powerful when applied to complex, multi-step routes for key intermediates or first-in-class target molecules.¹⁶ Their application to designing routes for a large series of final analogues in parallel is less common, as these campaigns often preserve a common core scaffold and vary peripheral building blocks. However, there is an opportunity to further enhance the practical applicability and feasibility of the proposed routes. This opportunity stems from the inherent data incompleteness in public literature, specifically the limited availability of negative reaction data and occasional omissions in patent information.¹¹ By enriching these models with a broader spectrum of real-world experimental outcomes – both successes and failures – we can refine their predictive power for robust, lab-ready conditions and sequences. We are confident that continuous improvement through more comprehensive data integration will enable AI tools to deliver more reliable and realistic synthetic routes, further boosting user acceptance and trust within the chemistry community and ultimately augmenting our synthesis capabilities. Even with sophisticated AI, we anticipate that human insight and practical validation will always be crucial in retrosynthesis and forward synthesis, providing the essential bridge between theoretical proposals and successful experimental outcomes.

To date, many chemists still perform manual, time-consuming literature searches using databases such as SciFinder and Reaxys to extract reaction details. The selection and optimisation of reaction conditions (e.g., temperature, solvent, catalysts, reagents) have traditionally relied on chemical intuition and manual effort.¹⁷ In recent years the ML-guided identification of suitable conditions for individual reactions has received considerable attention and delivered promising case studies.^18–20 Recent advances in ML-guided reaction prediction have been summarised by the groups of White and Schwaller,²¹ and Yang and Luo.²² At Roche, we have delved into different use cases depending on the scope of application (Fig. 2). For medicinal chemistry, we have successfully established graph neural networks capable of predicting C–H functionalisation reactions (Fig. 2A–C).^23,24 For medicinal chemistry and process chemistry we have investigated the prediction of the Suzuki–Miyaura reaction, predicting screening plate layouts for High-Throughput Experimentation (HTE) campaigns (Fig. 2D). Further examples from Roche Process Chemistry & Catalysis have demonstrated the prediction of ideal conditions for Buchwald–Hartwig reactions and a framework for batched multi-objective reaction optimisation using Bayesian methods.^26,27

Fig. 2. Examples of AI-driven reaction prediction in medicinal chemistry at Roche. (A) Machine-learning prediction of reaction yield and regioselectivity in Ir-catalysed C–H borylation, showing correlation between predicted and experimental yields across diverse substrates and accurate site-selectivity forecasts.²³ Adapted from Nippa, Atz et al.,²³ licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). (B) Prospective drug-design workflow with Minisci alkylation and an “ML-Funnel” for potency, reaction feasibility, and adsorption, distribution, metabolism, excretion, toxicity (ADMET) properties, applied to 26 375 virtual carboxylic-acid-hit enumerations, yielding 14 improved lead molecules.²⁸ (C) Automated in silico substrate selection for Minisci alkylation, using graph neural networks.²⁴ Adapted from Nippa, Atz et al.,²⁴ licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). (D) ML-guided 96-well-plate design for Suzuki cross-couplings, optimising solvent/base/catalyst combinations to maximise predicted yields across 8 × 12 conditions.²⁵.

Going forward, as computational power increases and larger curated datasets become available, we anticipate that retrosynthetic analysis and condition prediction will merge into a single task. Retrosynthesis will be driven by the actual feasibility of the individual transformation obtained through reaction condition prediction of each step. This might also include the prediction of reaction kinetics to avoid undesired by-product formation (see 2.4 Reaction monitoring) and associated purification (see 2.5 Work-up and purification). For difficult-to-predict transformations where AI models are uncertain of exact conditions, we envisage the proposal of screening plate layouts to assess route feasibility.

A critical component will also be the setup of the user interface to make data and models readily accessible to chemists. The advent of agentic Large Language Models (LLMs) is reducing the barriers to interacting with complex models. Imagine dropping an image of your desired target molecule into a chat and iteratively working through the synthesis steps with your chemical ChatBot (“ChatGPT for Chemists”).^29,30 Similarly, one could envision it assisting with core diversification (e.g., using parallel chemistry or C–H functionalisation) to explore structure–activity relationships (SAR) rapidly. These approaches could be directly incorporated into design processes. We recently disclosed such a workflow, highlighting the impact of synthetic accessibility assessment in the design process.²⁸ We are confident that such a forward-looking synthetic planning system will become a reality if fundamental changes are made in the documentation of chemical reactions (see 2.8 Documentation). Achieving this level of predictive synthesis will require concerted efforts to generate datasets, calling for chemists and organisations alike to treat data stewardship as a central pillar of digital chemistry innovation. This would represent a major leap forward in digitalising chemistry.

2.2. Sourcing of starting materials

The speed of compound synthesis in medicinal chemistry fundamentally relies on the quick access to diverse monomers and building blocks (BBs).³¹ The rapid availability and structural diversity of these BBs are paramount, as they directly dictate the chemical space that can be explored for potential drug candidates. Many pharmaceutical companies use a sophisticated Chemical Inventory Management System to manage their vast and diverse chemical inventory. This system is crucial for real-time tracking, secure storage, and ensuring regulatory compliance (e.g., narcotics). Computational tools, increasingly enhanced by AI, are also integrated to efficiently explore chemical space and generate new target compound ideas.^31–36

The array of functionalities available in BBs is extensive, spanning from carboxylic acids and amines to diverse aromatic and heterocyclic boronic acids and halides, with each providing handles for synthesis. To empower individual research chemists, many companies developed an in-house interface for searching BBs and screening compounds. This desktop-based tool provides frequently updated punch-out catalogues from all major global BB providers (e.g., Enamine, eMolecules, Chemspace, WuXi LabNetwork, Sigma-Aldrich, MCule), next to smaller vendors providing access to specific and rare compound collections (e.g., unnatural amino acids, fluorinated BBs, etc.). It offers comprehensive metadata-based (e.g., supplier, compound identifiers, lead time, packaging size, price) and structure-based filtering options, allowing chemists to quickly identify project-relevant BBs. Some vendors also offer pre-weighted BB support, which allows cherry-picking compounds from their stock collection to create a custom library tailored to their exact specifications that is shipped within a few days. This option reduces the overhead of managing a large internal compound archive. It also eliminates the need for labour-intensive and error-prone in-house weighing, dissolution, and reformatting, thereby freeing up valuable internal resources. Beyond physically stocked compounds, virtual catalogues are expanding accessible chemical space even further. The Enamine MADE (MAke-on-DEmand) building block collection, for instance, represents a vast virtual catalogue with currently over a billion compounds not held in physical stock but that can be synthesised upon request.³⁷ This model relies on pre-validated synthetic protocols and readily available starting materials enabling delivery within a few weeks with a high success rate. We anticipate that the option to select from virtual BBs, not just physically available stock, will be integrated as a basic feature into future enumeration tools.

In the past, pharmaceutical companies spent millions establishing internal chemical stores to support parallel synthesis.³⁸ This resulted in large, yet structurally biased, BB collections. While some sought diversity through commercial or custom BBs, the recent explosion in commercially available options has rendered owning vast internal collections costly and unsustainable. To better utilise BBs, Pfizer and Janssen established the Building Block Exchange Consortium (BBXC) in 2018.³⁹ This innovative, pre-competitive model allows member companies to swap BBs, thereby increasing diversity and maximising returns on existing assets. Roche joined in 2019, and the consortium now comprises seven large pharmaceutical companies and one agrochemical company. While establishing such a consortium certainly presents challenges, such collaborative frameworks enable the pooling of diverse resources, expertise, and infrastructure, facilitating ambitious research projects beyond individual institutional capacities. At Roche we have seen success when utilising consortium-derived BBs for internal DNA-Encoded Library production and parallel synthesis campaigns.

The increasing complexity of modern drug targets necessitates BBs that can give rise to final molecules with specific physicochemical properties (e.g., solubility, permeability, metabolic stability). The advent of AI and digital chemistry is rapidly reshaping BB sourcing and utilisation by accelerating molecular design, optimising synthesis via advanced retrosynthesis, and streamlining inventory management. These innovations promise to reduce drug discovery timelines and costs, fostering a more efficient and predictive development ecosystem.⁴⁰

2.3. Reaction set-up

The process of setting up a chemical reaction lies at the heart of small molecule synthesis in medicinal chemistry. Historically, and still to a large extent today, this has been a highly skilled but manual process.⁴¹ A chemist, trained in the art, weighs out solid reagents on a balance, measures liquid reagents, and assembles the necessary reaction vessel (round-bottomed flask or vial) and apparatus (stirrer, hotplate, condenser, etc.).^42,43 While this traditional approach allows for a high degree of flexibility and is even well-suited for the synthesis of single examples of complex, late-stage molecules, it is also labour-intensive, time-consuming, and susceptible to human error, which can impact reproducibility. For the rapid synthesis of libraries of final analogues, automation and parallelisation are essential.

The relentless pressure to accelerate the drug discovery process has driven an evolution in how reactions are set up. While the automated synthesis of peptides on solid supports has long been established and has certainly inspired progress in other areas, its repetitive nature makes it an outlier rather than a template for the more diverse and complex world of small molecule synthesis.^44,45

Across the pharmaceutical industry, automation is being steadily used for reaction setup and downstream applications.^46,47 Certain transformations are more amenable to automation, such as amidation, click chemistry, reductive amination and some cross-couplings.^45,48–50 Yet, initiatives like Eli Lilly's Life Sciences Studio, which aimed to revolutionise compound synthesis through remote, robotic management, ultimately faced challenges.^51,52 Indeed, achieving fully autonomous and universally applicable reaction setups remains a formidable task as automation struggles with the nuances of chemical transformations, adaptive troubleshooting, and the inherent inflexibility of programmed workflows for non-routine chemical reactions.^53,54 Combined with substantial upfront investment costs and the intricate integration required for diverse laboratory equipment, widespread adoption is further limited.⁵⁵

A pivotal development in the automation of reaction setup for medicinal chemistry has been the advent of HTE, which has found great adoption across the pharmaceutical industry.^46,56–58 HTE embraces the principles of miniaturisation and parallelisation, allowing for hundreds or even thousands of reaction conditions for one chemical reaction to be screened simultaneously in plate format.^59–61 At the core of HTE are liquid handlers and robotic solid dispensers, which can accurately dispense small quantities of reaction components. This automation increases the number of reactions that can be set up in a given time while also allowing immediate detection of potential errors. Nanomole-scale synthesis using HTE is technologically feasible.⁵⁸ Yet its practical use so far in medicinal chemistry is currently limited. Most HTE applications operate on a micromole to low-millimole scale to ensure robustness of sensitive catalytic reactions and allow sufficient material to be produced for initial biological testing and characterisation. Further miniaturisation is a potential future direction but would require parallel advances in ultrahigh-throughput analytical and purification devices, bioassay technologies and specialised equipment, i.e. nanodroplet dispensers for 1536 plates.

We view HTE with a high degree of automation as a go-to solution to conduct reaction condition screening. Similarly, this holds true for library production where a validated reaction is used to generate an array of final analogues. One of the persistent challenges for automation lies in handling solid reagents, as many are not soluble in appropriate solvents, and their diverse physical properties make the accurate robotic dispensing of sub-milligram amounts a significant hurdle. Researchers at AbbVie pioneered the ChemBeads technology to address this critical bottleneck.^62–65 Uniformly coating glass beads with the desired reagent can circumvent dispensing issues. By augmenting the total mass of the reagent as a bead construct with predictable physical characteristics, a robot or another device can achieve precise and consistent dispensing of minuscule amounts of material. Selected building blocks (see 2.2 Sourcing of compounds) need to be readily available for dispensing as solutions or solids to improve Make process times.

For library production, high value can be derived from the interplay between reaction screening and machine learning to select a single set of robust conditions to achieve high success for the entire array. In future, these models could predict optimal conditions for the individual chemical targets within an array; however, this would be complex to practically realise. A more realistic and expedient philosophy would be to bin or categorise chemical classes, with similar reactivities, together and predict the best conditions for each chemical class. The steering of automation solutions requires appropriate software that is interconnected with internal (compound) databases to operate these systems efficiently and allow for seamless data analysis.⁶⁶ Data from HTE reaction condition screening and from the reaction outcome of parallel synthesis libraries can be used to further train AI models, especially as negative data points are collected. However, use of isolated yields from productive library synthesis does not necessarily capture the true result of the reaction as the purification step introduces variability.

From our point of view, the starting material synthesis required for HTE and parallel synthesis campaigns, for the foreseeable future, will continue to require the flexibility and adaptability offered by the more manual synthesis approach. The automation of such syntheses, encompassing different reaction methodologies, demands a novel optimisation skillset incorporating chemistry, robotics and data science. The same holds true for late-stage lead optimisation, where, for example, targeted C–H functionalisation or molecular editing methods can be expeditious.^67–70 For such transformations, with limited material available, miniaturised HTE-based reaction condition screening using the ChemBeads technology can also help to boost reaction success.²⁴ Academic and industrial exchange that allows sharing of innovative technologies, such as new synthesis methodology or advances in practical techniques, should be strongly encouraged.^71–75

Looking ahead, we foresee that the development of customised automation software and hardware solutions will be needed.^66,76 In general, however, approaches should be centred around ‘the right tool for the right job’ philosophy, applying automation where it reduces manual, repetitive and error-prone steps.

2.4. Reaction monitoring

In monitoring a chemical reaction, chemists seek to understand the transformation taking place in the reaction vessel. Traditionally, this process is characterised by manual sampling. At discrete intervals, an aliquot of the reaction mixture is analysed, most commonly by Liquid Chromatography-Mass Spectrometry (LC-MS). The objective is to obtain a static snapshot in time, to confirm the consumption of starting materials, verify the formation of the desired product, and identify any emerging side-products.

Discovery and early process research mainly utilises manual LC-MS sampling for in-process control (IPC). HTE and array chemistry can utilise automated solutions for sampling, including robotics, to minimise manual manipulations. For the synthesis of larger scales, manual sampling can be coupled with in-process-control technologies such as in-line Nuclear Magnetic Resonance (NMR) or Infrared (IR) spectroscopy.

Conventionally, chromatography data is treated as ephemeral, used for go/no-go decisions, while the rich underlying data is discarded. This information is typically lost in disparate silos, disconnected from Electronic Lab Notebooks (ELNs) and inaccessible for future analysis or machine learning.⁷⁷ This paradigm of disconnected data substantially bottlenecks chemistry research. The solution is a robust digital infrastructure built on FAIR principles (Findable, Accessible, Interoperable, Reusable).⁷⁸ By adopting standardised sample formats, data from any instrument becomes seamlessly integrated and analysable. This process turns structured, contextualised data into a permanent and valuable asset, ready for daily use by chemistry teams and for training advanced algorithms.⁷⁹ However, harmonising data output is challenging, given the plethora of software applications and the diverse output formats generated by various analytical devices from different vendors. This is an opportunity for suppliers to differentiate their products by offering simple access to the raw and processed data their devices are producing.

Samples for single batch experiments are generally analysed manually through visual inspection of PDFs. Correlation of these analytical samples to the ELN has not been established broadly, which can lead to connection gaps and the loss of important data points. For HTE and parallel synthesis, a nomenclature to allow seamless connection to the actual experiment is critical, which Roche has addressed with HTE OS.⁶⁶ In this system, internally developed scripts enable the parsing of the LC-MS raw data into a structured format readily available for automated sample analysis. This reduces the time taken to analyse the vast amount of reaction data.

We see this FAIR data foundation as the key to enabling digitalised chemical synthesis. Connecting ELN information of the reaction with a unified (raw) analytical dataset will become an invaluable asset. To make the data readily available for machine learning, it should be structured in a hierarchical way, starting at the sample level, followed by peak and mass levels. Combining these extra data layers with reaction information (see section 2.8), the prediction of reaction outcomes can be further improved. Going forward, this will create a loop where every monitored reaction enriches the analytical database, thereby enabling better predictions.

2.5. Work-up and purification

Upon verification of product formation and reaction completion, the often challenging and laborious process of work-up and purification begins. This stage involves isolating the desired compound from the crude reaction mixture, which may contain unreacted starting materials, by-products (direct result of the desired reaction), side-products (result of a competing side reaction; less predictable), reagents, catalysts or residual solvents. Chemists rely on their considerable experience to devise efficient work-up and purification protocols, based on the properties of the target molecule or the utilised reagents. This remains a manual step with possibly non-obvious parameters.^80–84 Subsequently, it is typically followed by chromatographic purification, where selection of the right separation method, machine, column and mobile phase is important to achieve the desired separation.⁸⁵ The choice of purification technology is highly dependent on the nature of the synthesis. For gram-scale intermediates, classic techniques like crystallisation or manual flash chromatography are common. In contrast, the milligram-scale products generated from parallel synthesis of final analogues are more suited for automated, high-throughput purification systems like mass-directed High-Performance Liquid Chromatography (HPLC) and Supercritical Fluid Chromatography (SFC). Due to its time, variability and resource intensity, purification can be a bottleneck, and a failure at this advanced stage necessitates restarting the synthesis process.

Systems for automated flash chromatography, mass-directed preparative HPLC and SFC have become commonplace in many medicinal chemistry labs. These instruments are already an improvement over traditional manual flash chromatography. More advanced automated purification platforms have been established at a number of pharmaceutical companies. Novartis has developed a fully automated platform that integrates robotic liquid handlers and purification systems to process entire compound libraries with minimal human intervention.^86,87 AbbVie pioneered an automated high-throughput system for both purification and subsequent post-purification handling, such as desalting and compound distribution, which improved their capacity to support discovery programs.⁸⁸ Merck Sharp & Dohme (MSD) describes different parallel purification strategies using solid phase extraction to generate high-purity final analogue libraries via a more efficient and sustainable workflow.^84,89 These applications demonstrate a unified strategy across the pharmaceutical industry: integrating synthesis, purification, and analytics to increase throughput, yielding high-quality final analogues for biological testing. However, these are usually systems run by dedicated, specialised teams, which require substantial investment; therefore, in a lot of cases, purifications are still carried out manually.

Going forward, we see an opportunity to leverage the advancements discussed in sections 2.1 and 2.3, which should enable optimised work-up and purification, by, for example, reducing side products. Further, the data from the isolation step itself could be utilised to improve the process. Usually, numerous fractions from extractions or chromatographic runs are analysed by LC-MS. However, automatically accessing the underlying raw data to allow for improved data capturing, analysis, reporting and storing remains a bottleneck.⁹⁰ Already, the last IPC (see section 2.4) of the reaction mixture contains a wealth of information that delivers the basis for the purification strategy. Accessing, cleaning and curating the raw LC-MS data of all purification steps, obtaining the purification method information and connecting the data with the reaction will render a more streamlined purification approach possible. Building on the analytical data structure (see section 2.4), an additional purification layer can be added before an associated sample layer, as each purification run can have several associated samples.

We are confident that this can enable a more data-driven purification workflow based on ML and predictive modelling. As an example, for liquid–liquid extraction, digital tools are emerging that can predict the optimal solvent system based on the physicochemical properties (e.g., log P, pK_α) of the target molecule.⁹¹ First models that can predict retention times of final analogues based on their structure, which can help in selecting the appropriate chromatographic method and optimising the gradient, have also been disclosed.^92–94 However, these models often rely on calculated properties or data from pure molecules, rather than on the complex reality of a crude reaction mixture. Incorporating and interconnecting all relevant data along the purification process could lead to a breakthrough. Solving this complex data puzzle to obtain large, high-quality, and well-annotated datasets will be key to increasing purification efficiency.

2.6. Analytical characterisation

Analytical characterisation represents a critical quality checkpoint in the synthesis process. It is here that purified molecules undergo final characterisation, traditionally using MS and NMR spectroscopy. This analytical data serves multiple essential purposes: (1) structure confirmation and characterisation; (2) purity determination for subsequent biological testing; (3) providing the necessary evidence for patent applications; and (4) creating a permanent, searchable record for future research and re-synthesis efforts. Historically, this process involves chemists manually preparing samples, setting up the analytical runs on MS and NMR devices, and then individually analysing the resulting spectra. This represents a time-consuming, manual workflow with multiple data interfaces.

Today, multiple different characterisation workflows are in place, generally following open-access, centralised, or hybrid models. In the open-access approach, chemists perform their own analyses, a model pioneered by Pfizer with “walk-up” LC-MS systems for rapid purity and identity checks.⁹⁵ More recently, Eli Lilly developed a workflow for automated structure identification using the measured analytical data, markedly reducing the amount of uncharacterised samples.⁹⁶ Many companies now use hybrid models, such as MSD's integrated system, where automated purification is coupled with manual final characterisation, delivering a complete LC/MS and NMR data package for each final analogue and balancing speed with expert oversight.⁸⁹

With automation reducing manual sample preparation and measurements, the advent of advanced data analysis and ML has shifted interest towards the subsequent steps: data analysis, interpretation, and knowledge extraction. The raw data, often stored in proprietary vendor formats originating from different devices, remains difficult to access, compare, and reuse. Many vendors offer software that allows analysis based on structural input by the user. Yet, there remains potential to transform analytical data from a static, siloed report into a dynamic, queryable digital asset. These efforts towards standardising data formats (e.g., nmrML, JCAMP-DX) should also include enriching the data with persistent identifiers and structured information (e.g., InChI keys, reaction IDs) to establish the important connection between experiment, analytical measurements and obtained molecule. This final analytical characterisation data can be stored similarly to the method described in section 2.4. However, it is critical to store the structural information of the product. For NMR data, sample information and chemical shifts can be documented in a similar hierarchical structure.

With structured data available, ML can be leveraged to further automate the cognitive tasks of data interpretation. Instead of a chemist manually interpreting raw NMR and MS spectra, the first studies covering advanced algorithms performing this task have been disclosed.⁹⁷ Further advances in the field will allow enriched analytical data to be automatically verified for quality and correctness (including prediction accuracy), flowing from the instruments into a centralised data warehouse. This will reduce manual, laborious NMR spectra interpretation and also follow FAIR principles. As a consequence, a high-quality digital asset can be established, which automatically generates the experimental sections for laboratory notebooks, reports, publications and even patent applications. While the automated drafting of full patents and publications is a complex legal and scientific task, the generation of the detailed, data-rich experimental characterisation section is a tangible near-term goal. By further reducing the documentation, analysis and sample handling burden for scientists, we envisage increasing efficiency for the compound synthesis process.

2.7. Compound registration and handover

Upon confirmation of a final analogue's purity and structure, it is usually registered in the compound management or logistics platform, where it receives a unique identifier. This process can typically be initiated directly from the ELN by validating the structure through analytical data interpretation (see 2.6 Analytical characterisation of compounds).

Transferring a pure final analogue into vials for compound logistics, while seemingly a minor step, has historically been a labour-intensive bottleneck. Many scientists still manually transfer compounds from various containers into storage or assay vials. Some companies have addressed this by implementing redesigned workflows and automation. For instance, AstraZeneca shifted from a solid to a liquid compound delivery platform for assays, significantly reducing the upstream processing time.⁹⁸ Final analogues are weighed into robot-friendly vials for local solubilisation, and an integrated shaker efficiently mixes the compounds with the appropriate solvent. Concentrations are assessed in each tube, and they are immediately uploaded to the database, making final analogues available for ordering. This move towards liquid distribution and assay-ready plates has drastically reduced the time from final analogue submission to global availability for testing to within one week. Many companies use dedicated software platforms allowing for seamless management of submissions to enable tailored project-specific assay cascading (e.g., potency, ADMET, safety, etc.). These innovations and investments in compound logistics, i.e. through the use of stock solutions, demonstrate how automation, in even apparently simple manual steps, can be used to accelerate cycle times and improve overall data generation speed.

2.8. Documentation

The challenges of unfeasible routes and unreliable reaction conditions predicted by current CASP tools are closely related to the dependency on high-quality data. The inability to predict purification conditions has a similar root cause. Most pharmaceutical companies simply do not have their data in a structured, machine-readable format, much less connected to the actual experimental information. To create reliable ML tools that truly accelerate compound synthesis, we must rethink how we document experiments. Importantly, this needs to happen in close collaboration with the documenting scientists. Finding a fine balance between the time spent on the actual experiment and its associated documentation is crucial.

Historically, documentation meant creating a human-readable record in a paper notebook, essential for IP protection, but creating isolated silos of unstructured knowledge.⁹⁹ The ELN promised a leap forward, but for years, it has functioned as little more than “digital paper”, creating a machine-readable facsimile of the old process.¹⁰⁰ This approach, which captures critical experimental parameters in unstructured prose, is directly responsible for the “data incompleteness” that plagues modern synthesis planning tools. It fails to capture data in a format that can be considered as fully FAIR,⁷⁸ making it nearly impossible for a computer to understand and to learn from decades of accumulated knowledge and perpetuating a loss of trust in CASP systems.

At Roche, we recognised that to build the envisioned trustworthy, next-generation planning tools, we first had to fix the data input problem. We are taking this approach step by step. Firstly, we have delved into reaction data without including detailed analytical data, only adding a quantitative reaction outcome assessment. This led to the development of the Simple User-Friendly Reaction Format (SURF), our approach to standardising chemical reaction data (Fig. 3).⁷⁹ SURF provides a structured, machine-readable schema for capturing all aspects of a reaction, including the precise conditions and, critically, the negative results that are missing from public datasets but are essential for training robust reaction prediction models.^{23,66,101–103} The SURF data format is flexible and allows the addition of further data layers including data from analysis and purification.

Fig. 3. The simple user-friendly reaction format (SURF) bridges the gap between the laboratory and data science worlds to make experimental data seamlessly available for data-driven applications.

However, this foundational reaction data is only the beginning. The true power of this ecosystem will be fully realised as the rich streams of data from adjacent processes are being connected. In the future, a seamless digital thread that links reaction setup and monitoring with downstream work-up, purification, and final analytics could be in place. By automatically capturing data from analytical instruments and directly linking it to the corresponding experiment in a centralised data warehouse, collecting all experimental reaction data, a digital representation of every reaction can be created. This enriched, interconnected dataset will enable far more sophisticated machine learning applications, moving beyond yield prediction to forecasting impurity profiles, suggesting optimal purification methods, and even flagging potential downstream challenges before a reaction is ever run. As a consequence, currently used ELN templates need to be reassessed. From our perspective, it will be crucial to keep the scientist at the center and any adjustments should make the documentation process easier. Of course, this shift towards data-driven chemistry will not happen overnight, but we have demonstrated first successes, which deliver a solid basis for the next steps.

Ultimately, we hope that we can make every experiment count twice. On the one hand, we obtain chemical matter, and on the other hand, we make use of all the valuable data that the Make process delivers. Moving forward, this will lead to more efficient synthesis processes.

3. Future prospects, challenges and conclusions

The preceding sections have deconstructed the compound Make process, illustrating that this remains a rate-limiting step in the DMTA cycle and the discovery of novel medicines. We have highlighted how each stage – from initial synthesis planning and material sourcing to reaction setup, monitoring, purification, and final documentation – is ripe for transformation. The central argument is clear: moving beyond traditional, manual-intensive practices toward a more integrated and supported workflow is not just an opportunity but a necessity. The journey from disconnected data silos and bespoke techniques to a cohesive, data-driven synthesis engine is a central challenge we must address to accelerate drug discovery.

To achieve this acceleration, the strategic pillars of digitalisation, automation, and collaboration are of high importance. Digitalisation, through the rigorous application of FAIR data principles, is the bedrock upon which this progress can be built; it transforms ephemeral experimental observations into permanent, machine-readable assets that fuel predictive models. Automation, spanning from HTE robotics to purification platforms, provides the physical means to execute complex experimental designs with speed and precision, freeing scientists from repetitive, error-prone tasks. Finally, collaboration, exemplified by pre-competitive consortia like the BBXC, demonstrates that sharing resources, data, and risk is a powerful strategy to expand our collective capabilities and overcome challenges that are too large for any single entity to solve alone.

Looking forward, the ongoing evolution of these synthesis technologies can be a true enhancer of efficiency. AI-driven systems could not only propose viable retrosynthetic routes but could simultaneously predict optimal reaction conditions, forecast impurity profiles, and recommend purification strategies, all before a reaction is physically run. Intuitive interfaces, akin to a “ChatGPT for Chemists”, could allow scientists to design, simulate, and troubleshoot entire synthesis campaigns in a digital environment while also making it easier for chemists to engage with AI. Our forward-looking opinion sees a future where every experiment has a digital representation, contributing to a constantly learning system that makes each subsequent synthesis faster, more efficient, and more likely to succeed.

It is important to acknowledge the difficulties in implementing such an ambition. Machine learning models trained primarily on decades of well-established reactions may even limit chemical innovation. Such models will struggle to propose or accurately predict the outcomes of truly novel transformations that lie outside their training data. Furthermore, achieving the vision of data-driven synthesis requires more than just a standardised data format like SURF, it necessitates robust data governance. This involves establishing clear institutional policies for data quality, ownership, stewardship, and long-term maintenance. This is a significant organisational challenge that extends beyond the medicinal chemist.

Such difficulties need to be overcome before considering the ‘lab-in-the-loop’¹⁰⁴ for the entire small molecule DMTA cycle. In such a loop, AI would design novel molecules, a robotic platform would autonomously execute the multi-step synthesis, purify the products, and submit them for testing, with the resulting data feeding directly back to the AI to inform the next design cycle. Such a fully autonomous, universal platform for small molecule chemistry remains a distant goal, mainly because of the complex chemical problems that need to be solved across multiple chemical series during early-phase medicinal chemistry campaigns. However, more focused, semi-automated loops for late-stage lead optimisation using selected methodologies offer significant advantages, including being more flexible, agile and robust. As we have discussed throughout this article, these more limited loops are becoming increasingly feasible and offer small steps towards the greater ambition of ‘lab-in-the-loop’.

In the highly competitive landscape of drug discovery, the ability to rapidly design, make, and test novel chemical matter remains a key competitive advantage. Maintaining that edge will depend on how effectively an organisation can strategically implement these advanced synthesis capabilities to make smarter decisions faster. Adopting these technologies is not merely a matter of acquiring new hardware or software; it requires a deep-seated organisational commitment to fostering a culture of innovation, investing in a robust and interconnected digital infrastructure, and re-engineering legacy workflows.

This technological shift culminates in the evolution of the medicinal chemist's role. The chemist of the future will be a multidisciplinary architect of molecular synthesis, whose value lies not just in manual execution but in strategic orchestration. A strong medicinal chemistry skill set will need to be complemented by a new trinity of skills: proficiency in computational tools for synthetic accessibility prediction; mastery of automation systems to execute complex experiments; and a commitment to continuous education to leverage data science and AI. By embracing this evolution, we empower our scientists to focus on what they do best – innovate and solve complex chemical challenges – thereby accelerating the delivery of transformative medicines to patients.

Conflicts of interest

D. F. N., A. J. B., K. A., U. G., H. B., R. E. M. are full employees of F. Hoffmann-La Roche Ltd.

Data availability

No primary research results, software or code have been included and no new data were generated or analysed as part of this review.