AI-Driven Chemical Design: Transforming the Sustainability of the Pharmaceutical Industry

Abstract

The pharmaceutical industry faces mounting pressure to reduce its environmental impact while maintaining innovation in drug development. Artificial intelligence (AI) has emerged as a transformative tool across healthcare and drug discovery, yet its potential to drive sustainability by improving molecular design remains underexplored. This review critically examines the applications of AI in molecular design that can support in advancing greener pharmaceutical practices across the entire drug life cycle—from design and synthesis to waste management and solvent optimisation. We explore how AI-driven models are being used to personalise dosing, reduce pharmaceutical waste, and design biodegradable drugs with enhanced environmental compatibility. Significant advances have also been made in the predictive modelling of pharmacokinetics, drug–polymer interactions, and polymer biodegradability. AI’s role in the synthesis of active pharmaceutical compounds, including catalysts, enzymes, solvents, and synthesis pathways, is also examined. We highlight recent breakthroughs in protein engineering, biocatalyst stability, and heterogeneous catalyst screening using generative and language models. This review also explores opportunities and limitations in the field. Despite progress, several limitations constrain impact. Many AI models are trained on small or inconsistent datasets or rely on computationally intensive inputs that limit scalability. Moreover, a lack of standardised performance metrics and life cycle assessments prevents the robust evaluation of AI’s true environmental benefits. In particular, the environmental impact of AI-driven molecules and synthesis pathways remains poorly quantified due to limited data on emissions, waste, and energy usage at the compound level. Finally, a summary of challenges and future directions in the field is provided.

1. Introduction

Worldwide markets for active pharmaceutical ingredients are currently growing at unprecedented levels, with the sector’s annual revenue having almost quadrupled over the past two decades [1]. This market was valued at USD 1.86 trillion in 2022 [1] and is expected to increase. This growth is driven by a combination of factors, including rising healthcare demands [2], increased life expectancy [3], and technological advancements in biomedical research [4].

However, healthcare industries produce a significant amount of CO₂ emissions annually. It is estimated that, on average, the healthcare sector accounts for 5% of national CO₂ emissions from OECD countries, China, and India [5]. The carbon footprint from pharmaceuticals can be as high as 20%, as observed in the UK’s National Health Service (NHS) [6]. Globally, the emission intensity from the pharmaceutical industry is about 55% higher than that of the automotive industry, and its CO₂ footprint is forecast to triple by 2050 if left unchecked [7] according to the World Economic Forum. These figures highlight a pressing need to rethink how drugs are discovered, manufactured, and administered.

Wynendaele et al. [8] proposed ten elements that define sustainability within drug discovery. These include (1) ecological–environmental impact (benign-by-design); (2) medical needs; (3) green chemistry; (4) artificial intelligence and big data; (5) the root cause of illness; (6) risk- and decision-taking models; (7) biomarkers and bioinformatics to support precision medicine; (8) cost-effectiveness; (9) the lean discovery process; (10) responsible research and innovation. These principles evidence that incorporating sustainability into drug discovery is a multifaceted and urgent challenge that interfaces with every step on the development journey, from energy use and material inputs to waste management, toxicity, and ecosystem disruption. The interest in sustainability in the drug discovery pipeline has greatly grown within the past few years, and there is a consensus around the need for more sustainable practices in pharmaceutical development. Moreover, recent policy changes show promise to reduce the environmental impact of the industry, especially in climate change, biodiversity loss, wildlife decline, and environmental pollution. This is a consequence of pharmaceutical life cycles being mostly driven by legislative measures [9]. Multiple policy frameworks have also been co-developed for pharmaceutical prescribing, with a focus on preventing pollution and reducing environmental impact [10]. However, public awareness of the impact of the pharmaceutical industry, especially on the presence of residues in drinking water, remains poor in multiple geographies [11,12].

The integration of artificial intelligence (AI) is emerging as a transformative tool for enabling sustainability across the drug development pipeline. AI models can accelerate the discovery of novel compounds, reduce experimental waste, optimise synthesis pathways, and predict properties such as toxicity, biodegradability, or pharmacokinetics [13]. This review summarises the latest advances in molecular design and clinical AI algorithms with a potential impact on the sustainability of the pharmaceutical industry. We cover every aspect of the drug development journey, from the discovery and exploration of appropriate solvents and catalysts to the resulting waste and strategies to mitigate environmental impact. The manuscript is organised as follows: Initially, advances in drug waste reduction are described, either through dosage optimisation or the design of biodegradable drugs. We describe approaches to optimise the materials required in the synthesis of drugs, focusing on biocatalysts (enzymes) and heterogeneous catalysts. We then review advances in synthesis optimisation, including the incorporation of AI in the design of new drugs, solvents, and optimised synthetic pathways. Finally, a summary of limitations, opportunities, and future perspectives of the field is provided.

2. Background Information on AI Models and Key Concepts

Artificial intelligence and machine learning have crucial roles in modern pharmaceutical research. Multiple models are used nowadays to estimate numerical outcomes (regression) or categorical labels (classification) based on patterns from a training set. Some of the most commonly used models include the following:

-

Deep Neural Networks (DNNs) are versatile architectures composed of multiple hidden layers, capable of learning complex, non-linear relationships in data. They are used for both regression (i.e., predicting solubility or melting point) and classification (i.e., toxic, non-toxic substances). DNNs are especially powerful when working with large, high-dimensional datasets such as molecular fingerprints [14].

-

Extreme Gradient Boosting (XGBoost) is mainly used with structured data and is widely applied in regression tasks (i.e., predicting pharmacokinetic parameters), as well as in classification problems (i.e., drug–target interaction) [15].

-

Support Vector Machines (SVMs) are supervised algorithms that can operate as classifiers or regressors. In drug discovery, they are often used in classification tasks (i.e., active vs. inactive compounds), but they are also applicable in regression settings (i.e., estimating binding affinity or partition coefficients). SVMs perform well in high-dimensional spaces and with small-to-medium-sized datasets [16].

-

Random Forests are ensemble models that build multiple decision trees and average their outputs for regression or vote for classification. They are robust to overfitting, handle missing values well, and are commonly used in tasks such as predicting biodegradability, toxicity, or release profiles in drug delivery systems [17].

Among the most transformative AI approaches are generative models, which learn data distributions and can propose novel compounds with desired properties. These include architectures like Generative Adversarial Networks (GANs), variational autoencoders (VAEs), and diffusion models. In pharmaceutical contexts, they are used to propose new molecular structures, reaction pathways, or polymer designs tailored for biodegradability or low environmental persistence. A subset of generative models, language models (LMs), has demonstrated particular success in molecular and protein design [18]. Initially developed for natural language processing, LMs such as GPT or BERT have been adapted to “learn” molecular syntax or protein sequences, enabling the design of synthetic enzymes and prediction of drug-like molecule behaviour. These models, especially when scaled to billions of parameters, show remarkable generalisation, allowing for zero-shot predictions and novel molecule generation without explicit prior examples.

To offer an objective and robust comparison between models, multiple metrics have been developed and are frequently reported. These include regression-based measures, such as Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Mean Relative Error (MRE), and Mean Absolute Percentage Error (MAPE), which quantify deviations between predicted and observed values. These metrics quantify the deviation between predicted and actual observed values:

-

Mean Absolute Error (MAE) measures the average magnitude of errors between predicted and true values. It is robust to outliers and provides a straightforward interpretation in the same units as the output variable. It is calculated as in (Equation (1)):

$$MAE={\displaystyle \frac{1}{n}}\sum _{i=1}^n|y_i-{\widehat{y}}_i|$$

(1)

where ${\widehat{y}}_i$ represent the predicted values, $y_i$ represents the actual values, and $n$ is the number of samples.

-

Root Mean Squared Error (RMSE) is calculated as the square root of the average of the squared differences between the predicted and actual values. This value is more sensitive to outliers than MAE (Equation (2)):

$$RMSE=\sqrt{{\displaystyle \frac{1}{n}}\sum _{i=1}^n(y_i-{\widehat{y}}_i{)}^2}$$

(2)

-

Mean Relative Error (MRE) expresses the absolute error as a proportion of the true value, averaged across all samples (Equation (3)):

$$MRE={\displaystyle \frac{1}{n}}\sum _{i=1}^n{\displaystyle \frac{|y_i-{\widehat{y}}_i|}{y_i}}$$

(3)

-

Mean Absolute Percentage Error (MAPE) represents the average percentage difference between the predicted and actual values (Equation (4)):

$$MAPE={\displaystyle \frac{100}{n}}\sum _{i=1}^n{\displaystyle \frac{|y_i-{\widehat{y}}_i|}{y_i}}$$

(4)

In addition to the above-mentioned absolute and relative error metrics, model performance can be evaluated using statistical and probabilistic measures such as R-squared (R²) for regression algorithms and Receiver Operating Characteristic (ROC) curves and the Area Under the Curve (AUC) for classification tasks. R², a statistical measure of fit, evaluates how well model predictions explain variability in the data. It quantifies the proportion of variance in the dependent variable that is predictable from the independent variables (Equation (5)):

$$R^2=1-{\displaystyle \frac{{\sum }_{i=1}^n(y_i-{\widehat{y}}_i{)}^2}{{\sum }_{i=1}^n(y_i-\overline{y}{)}^2}}$$

(5)

where $\overline{y}$ represents the mean of the observed data. For classification models, metrics such as Sensitivity, Specificity, Receiver Operating Characteristic (ROC) curves, and Area Under the ROC Curve (AUC) are used to assess discriminatory power:

-

ROC Curves (Receiver Operating Characteristic) are used for binary classification problems. They are graphical representations of the True Positive Rate (Sensitivity) against the False Positive Rate (1 − Specificity) at various threshold settings.

-

AUC (Area Under the ROC Curve) summarises the ROC curve into a single number ranging from 0 to 1. It represents the probability that a randomly chosen positive instance is ranked higher than a randomly chosen negative one. Values equal or below 0.5 indicate that the model has no ability to discriminate between categories.

Beyond model evaluation, life cycle assessment (LCA) provides a critical framework for assessing the full environmental impact of pharmaceuticals, from raw material extraction to disposal. It is a systematic methodology used to evaluate the environmental impacts associated with all stages of a product’s life, from the extraction of raw materials to its production, distribution, use, and end-of-life disposal or recycling [19].

3. Optimisation of Synthetic Routes

One of the future applications of AI in organic synthesis, and the development of pharmaceuticals, is the optimisation of chemical reactions, including designing new synthetic pathways and optimising existing reactions. Algorithms have been used, for example, to improve the yield and selectivity of reactions [20], which could reduce waste and energy expenditure. The yields obtained in the development of pharmaceuticals are generally low due to a number of factors:

• Molecular Complexity: Many pharmaceutical compounds have intricate structures with multiple functional groups. While these complex structures have been associated with higher selectivity [21], it makes their synthesis challenging.
• Multi-Step Synthesis: Drug molecules often require multiple synthetic steps, each with its own inefficiencies, side reactions, and purification losses [22,23].
• Chirality and Selectivity: Over half of drugs require specific stereochemistry [24], meaning that reactions must favour the desired enantiomer or diastereomer, which can reduce the effective yield.
• Purification Losses: Rigorous purification steps are needed to meet regulatory purity requirements, leading to additional material loss [25].

The use of catalysts can improve the yield of drug synthesis. It is estimated that around 90% of chemicals, including pharmaceuticals, derive from catalytic processes [26]. These include the use of biocatalysts, such as enzymes, or heterogeneous catalysts. Catalysts can boost the efficiency of synthetic processes and reduce environmental impact by minimising energy requirements. The synthetisation process of drugs contributes significantly to the CO₂ production of pharmaceutical companies. This CO₂ is generally a consequence of the reagents, solvents, and transport required. A life cycle study of a prostate cancer drug found its emissions per declared unit (blisters) to be equal to 34 kg CO_2eq, with intermediate and active pharmaceutical ingredient synthesis being the largest contributors to this footprint (about 96%) [27]. However, the use of AI in the development and optimisation of new catalysts is still in the early stages.

3.1. Biocatalyst Design

Enzymes play a pivotal role in the synthesis of pharmaceuticals, especially in stereoselective reactions [28]. Enzymatic processes help improve reaction efficiency, reduce waste, and offer a greener alternative to traditional chemical synthesis [29]. Some examples of pharmaceuticals developed using enzymatic reactions include ibuprofen, where the use of an enzymatic catalyst has been shown to improve the sustainability of production [30]; L-DOPA, used as cardio stimulant [31]; and statins [32], among others. However, the number of drug synthesis pathways that can be improved through enzymatic catalysis is limited, and our ability to optimise reactions is limited to current methods to evolve enzymes. While some methods have been developed to develop new enzymes, including random mutagenesis [33] and recombination [34], their yields are suboptimal, leading to high resource and time requirements.

New AI models trained on protein sequences and conformation information can speed up the development of new enzymes by predicting their performance in silico. So far, multiple models, including adversarial networks, language models, variational autoencoders, and maximum entropy models, have been used in this field, each offering distinct advantages for modelling sequence–function relationships and exploring protein fitness landscapes [35]. In particular, language models are one of the most widespread algorithms used in the development of generative models for protein design. These models have been successful in the modification of proteins to obtain functional enzymes with desired properties, as well as the de novo design of proteins. Munsamy et al. [36] developed a conditional language model for the design of artificial enzymes. The ZymCTRL model developed by this team could generate active enzymes even when sequence identities were below 40% [37]. This concept was demonstrated by the design of carbonic anhydrases and lactate dehydrogenases. Moreover, Madani et al. [38] developed ProGen, that uses autoregressive LLMs to generate protein sequences with predictable functions. This model was trained on 280 million protein sequences and could produce curated protein sequences to generate lysosome analogues. Beyond language models, diffusion-based generative models have also been explored for protein engineering. These models, originally developed for high-fidelity image synthesis, have been adapted to operate in sequence or structure space, allowing for fine-grained control over protein design. Notably, diffusion models have been applied to the modification of cytochrome P450 enzymes, leading to engineered variants with enhanced biocatalytic activity toward flavonoid substrates [39].

AI in de novo protein design has also become of great interest within recent years, opening up a wealth of possibilities in the development of custom-made enzymes that can catalyse key reactions to produce molecules of interest for the pharmaceutical industry. The ability of these models to generate functional proteins has already been demonstrated. The recently published ESM3 generative model, for example, could generate a new green fluorescent protein, with a distance from existing proteins equivalent to simulating over 500 million years of evolution [40]. Moreover, Lauko et al. [41] designed de novo enzymes to mimic the serine hydrolase function, with catalytic efficiencies (k_cat/k_m) up to 3.8 × 10³ M⁻¹ s⁻¹ (Figure 1a). One of the limitations of these models is their high computational requirements. The ESM3 model was trained on 3.15 billion protein sequences, 236 million protein structures, and 539 million proteins with function annotations [40].

Another application of AI in enzyme design focuses on the enhancement of enzyme properties, particularly catalytic stability. Catalytic stability refers to the ability of an enzyme to retain its catalytic activity over extended periods and under a variety of physicochemical conditions. This trait is of paramount importance in industrial and pharmaceutical settings, where enzymes are often required to function in non-native environments such as elevated temperatures, extreme pH values, the presence of organic solvents, or inhibitory compounds without a significant loss of activity or structural integrity.

Stability is a particularly key parameter in the development of biocatalysts for large-scale chemical synthesis, where prolonged reaction times and harsh reaction conditions are common. Enhancing enzyme lifetime is key to reducing environmental CO₂ emissions. Some cradle-to-gate evaluations of enzyme production estimates that each kg of enzyme produced for synthesis releases between 1 and 10 kg eCO₂ [42]. These emissions can be even higher (between 16 and 25 kg eCO₂) in the case of immobilised enzymes for pharmaceutical applications [43].

Traditional approaches to improving enzyme stability typically involve labour-intensive cycles of mutagenesis and screening, often limited by the size of sequence space that can be experimentally explored [44,45]. AI-based strategies now offer an efficient alternative by enabling the predictive modelling of stabilising mutations and guiding rational protein engineering. Deep learning models trained on datasets of thermostable and non-thermostable enzymes can identify sequence motifs or structural features correlated with enhanced stability. For example, bi-long short-term memory algorithms have been used to predict thermostability by mining hidden features that can be used to build a network classifier and distinguish between thermostable (active beyond 40 °C) and non-thermostable proteins [46]. Bayesian optimisation techniques have also been introduced to explore fitness landscapes for stability in a guided manner and determine the optimal catalytic temperature for enzymes which could be used to find favourable mutations [47]. Finally, large language models have also been incorporated in this field [48] by converting each mutant into a latent space representation. This algorithm allowed the design of a creatinase with a 10.2 °C increase in melting temperature and a 655-fold increase in half-life at 58 °C. These predictions can then be validated experimentally or integrated into directed evolution pipelines. Furthermore, structural AI tools, such as AlphaFold2, are being leveraged to assess the stability of designed sequences by modelling structural conformations and calculating their associated thermodynamic parameters. While some studies point out that AlphaFold excels at protein structure prediction, the ability to determine stability changes upon inducing mutations might be limited [49]. Nevertheless, collectively, these AI-driven approaches are streamlining the development of robust enzymes suitable for diverse applications, ranging from green chemistry to therapeutic enzyme formulations. A summary of the machine learning models trained to design enzymes is provided in

Table 1. A comparison of the AI models used in the design of enzymes and the prediction of protein stability.

Dataset	Model Prediction	Algorithm	Performance	Refs.
280 million protein sequences from >19,000 families	Protein function	1.2-billion-parameter neural network	AUC = 0.85 (experimentally validated)	[38]
BRENDA DATASET (37,624,812 sequences)	Protein activity (model made available as Open Source)	Language model	-	[36,37]
3.15 billion protein sequences, 236 million protein structures, and 539 million proteins with function annotations	Protein activity	Language model	SS3 > 80% pTM > 0.8 pLDDT > 0.8 RMSD < 1.5 Å (experimentally validated)	[40]
Protein–small molecule complexes in PDB	Enzyme structure	Deep Neural Network	RMSD < 1.5 Å (experimentally validated)	[41]
2950 thermophilic protein sequences	Enzyme thermostability	Deep Neural Network and bi-long short-term memory	ACC (%) = 94.34 PR (%) = 93.97 REC (%) = 94.81 REC (%) = 94.36 MCC (%) = 88.73 AUROC (%) = 98.68	[46]
Dataset from 21,498 microorganisms	Enzyme optimal temperature (model made available as Open Source)	Random Forest regressor	R2 = 0.94 RMSE = 4.46	[47]
Optimal growth temperatures for 96 million host bacterial strains	Enzyme thermostability	Language model	Experimentally validated	[48]
16,706 unique sequences	Catalytic activity	Generative Adversarial Network	Experimentally validated	[50]

.

3.2. Heterogeneous Catalyst Design

Heterogeneous catalysis, where the catalysts are used in solid form, represent around 80% of catalytic processes in industrial chemistry [51]. As such, they have been extensively used in the pharmaceutical industry, given their ability to support the synthesis of key active pharmaceutical ingredients. Some of the most used catalysts include platinum, used in the synthesis of certain antifungal medications, among others [52]; palladium, which can catalyse C-N Cross-Coupling Reactions [53]; and ruthenium, used in olefin metathesis [54].

The design of new catalysts has traditionally relied on trial and error, which has hindered the development of the field. While some in silico approaches have been proposed [55,56], computer modelling is still mainly applied to study the mechanism of catalysis after experimental results have been obtained, instead of virtual screening. Following this trend, Li. et al. [57] developed a general theory of metal–support interactions for metal catalyst supports, using a machine learning model trained on experimental results of nanoparticle adhesion to metal oxides. However, our understanding of metal catalysts and how they can be modified to enhance chemical reactions is still limited.

The research into AI models for heterogeneous catalyst discovery is complex, due to the involvement of multiple factors in the reaction pathways, including solvents, temperature, and substrates, among others. Moreover, these models often require molecular descriptors that require expensive computing (i.e., DFT, molecular mechanics). Moreover, there is a growing pool of chemical information, both for ligands tested on different catalysts and materials that can be used as substrates, that increases the demand for computing resources. Multiple public databases have been published, such as the CHGNet [58], M3GNet [59], or the Open Catalyst Project [60], among others. Consequently, it is estimated that the search for catalysts with desired functions is more computationally demanding compared to API discovery [61]. This has also led to most material screening approaches based on modelling-based approaches being limited, with potential artefacts that limit the discovery of catalysts [62].

One of the most prominent uses of AI in this area is the screening of the “chemical space” to identify the most optimal set of parameters and construct the chemical space for model training. An alternative which can reduce the need for complex quantum mechanical calculations is the use of machine learning potentials [63]. These are used as surrogate models to increase computational speed, enabling the exploration of large search spaces and the optimisation of catalyst structures [62].

Additionally, AI has been employed to leverage empirical data from the scientific literature to optimise catalyst synthesis conditions. For example, Lai et al. [64] combined Bayesian optimisation with LLMs to automate information extraction from published studies, subsequently identifying conditions that optimised ammonia synthesis using Co₃Mo₃ (Figure 1b). This approach then provided a set of conditions that could maximise ammonia synthesis using Co₃Mo₃. LLMs have also been used by Wang et al. [65] to provide insights into the design of new electrocatalytic materials by training them on published evidence. However, these approaches are limited by the heterogeneity and unstructured nature of the experimental literature, which often lacks consistent reporting standards. Addressing this, domain-specific languages have been developed to facilitate structured data extraction. As an example, the Chemical Markdown Language, developed by Park et al. [66], provides a simple syntax for experiment documentation. This enabled the training of machine learning models to design catalysts for ring-opening polymerisation and experimental validation (Figure 1c).

Despite these advances, significant challenges remain. The discovery of heterogeneous catalysts using AI is hindered by (1) the computational intensity required to generate accurate molecular descriptors and (2) the vastness and complexity of chemical space. A promising strategy to overcome these limitations is the integration of AI with high-throughput experimentation (HTE), which can accelerate the optimisation of catalytic reactions [67]. However, this integration introduces additional logistical and data management challenges, including experimental reproducibility, the standardisation of protocols, and seamless data flow between experimental and computational platforms.

Figure 1. (a) A schematic representation of the protein design process using an RFdiffusion model. The model uses information from a given active site to generate a protein structure and backbone. Image reprinted with permissions from [41] (Copyright © 2025, American Association for the Advancement of Science, CC-BY 4.0 licence). (b) A schematic representation of an AI-driven workflow for catalyst design and optimisation, using Bayesian optimisation to speed up the discovery of materials, and real-world validation. Image reprinted with permissions from [64] (Copyright © 2023, American Chemical Society). (c) The tree manifold projection and visualisation of the catalysts generated by a fine-tuned regression transformer model trained on historical experimental data. The colours represent the synthesisability score (1–8 scale). Image reprinted with permissions from [66] (Copyright © 2025 Springer Nature, Creative Commons Attribution 4.0 International Licence).

Figure 1. (a) A schematic representation of the protein design process using an RFdiffusion model. The model uses information from a given active site to generate a protein structure and backbone. Image reprinted with permissions from [41] (Copyright © 2025, American Association for the Advancement of Science, CC-BY 4.0 licence). (b) A schematic representation of an AI-driven workflow for catalyst design and optimisation, using Bayesian optimisation to speed up the discovery of materials, and real-world validation. Image reprinted with permissions from [64] (Copyright © 2023, American Chemical Society). (c) The tree manifold projection and visualisation of the catalysts generated by a fine-tuned regression transformer model trained on historical experimental data. The colours represent the synthesisability score (1–8 scale). Image reprinted with permissions from [66] (Copyright © 2025 Springer Nature, Creative Commons Attribution 4.0 International Licence).

AI-enabled catalyst discovery holds considerable promise for the pharmaceutical industry, particularly in the pursuit of more sustainable and efficient drug synthesis. The ability to predict optimal catalysts and reaction conditions can reduce resource consumption, lower emissions, and shorten development timelines, key elements in the industry’s push toward net-zero targets. A summary of machine learning models trained to design heterogeneous catalysts is provided on

Table 2. Comparison of AI models used in design of heterogeneous catalysts.

Dataset	Model Prediction	Algorithm	Performance	Ref.
603 articles	Catalyst synthesis optimisation	Large language model and Bayesian optimisation	Experimentally validated	[64]
22,000 articles from Web of Science	Catalyst synthesis optimisation	Language model	Average predicted faraday efficiency = 64.15%	[65]
549 monomer–catalyst pairs for catalyst model	Monomer conversion, dispersity, and average molecular weight	Regression transformer	Pearson correlation = 0.59 (Experimentally validated)	[66]

.

4. Synthesis Optimisation

4.1. Speeding Up the Discovery of Drugs

One of the most immediate benefits of AI in the pharmaceutical industry is the faster development of drugs. The development of models to predict protein structures, such as AlphaFold 3 [68], has enabled the accurate modelling of proteins and interactions (Figure 2a,b). This has allowed for the development of models that can accurately predict drug interactions with targets and generate drugs with specific properties in terms of synthesisability or target properties (Figure 2c). Current projections estimate that, if the success rates observed between Phase I and II clinical trials from AI-discovered molecules hold, the end-to-end probabilities for a molecule to succeed across all phases would almost double, from 5–10% to 9–18% [69].

The faster development of pharmaceuticals due to accurate AI models could potentially reduce the number of laboratory tests and failed clinical trials. Current estimates may indicate a carbon consumption of 27.5 million tonnes of CO₂ spanning all clinical trials [70], just under a third of the emissions of Bangladesh. AI models could also be used in the repurposing of drugs [71], which could avoid further environmental costs due to pharmaceutical production (Figure 2d). One of the limitations of current AI models in drug discovery is the lack of synthesisability considerations when designing new small molecules. Challenges in the synthesis of algorithmically developed molecules have been described since 1995 [72]. This parameter refers to how easily a given molecule can be synthesised and could lead to target drugs being difficult or impossible to synthesise. In turn, drugs with low synthesisability can be translated into costly processes, which increases their carbon footprint. To circumnavigate these challenges, multiple approaches have been developed. Parrot et al. [73] introduced a Retro Score for molecular generator algorithms, which compute synthetic accessibility scores of molecules through a retrosynthesis analysis. This score correlated well with the assessments performed by chemists and could be implemented as a constraint in complex molecular generation algorithms. Alternatively, Swanson et al. [74] assembled a database containing 132,000 molecular building blocks and 13 well-validated chemical synthesis pathways which could be used to explore a chemical space of nearly 30 billion easy-to-synthesise molecules. The team employed the resulting algorithm to design and experimentally validate 58 molecules with antibacterial activity. Models are expected to improve in accuracy as the size of the available molecule libraries increases. However, accurate measurements of CO₂ derived from the training and use of these model are still required to determine the benefits of AI in the field.

Figure 2. (a) Predicted structures and interaction of designed proteins with three different ligands. Image reprinted with permissions from [68] (Copyright © 2023 American Association for the Advancement of Science, Creative Commons Attribution 4.0 International Licence). (b) Example of protein designed using Alpha Fold 3. Image reprinted with permissions from [68] (Copyright © 2023 American Association for the Advancement of Science, Creative Commons Attribution 4.0 International Licence). (c) Schematic representation of training of generative models for de novo drug design of molecules. Process includes using databases and iterating over model weight adjustments and accessibility scores to design novel drugs to meet specifications. Image reprinted with permissions from [73] (Copyright © 2023 Springer Nature, Creative Commons Attribution 4.0 International Licence). (d) Schematic representation of TxGNN processing of queries to calculate predictions for drug–disease relationships. Approach takes into account interconnectedness between both factors. Predictions from model align with off-label prescriptions from clinicians in large healthcare systems. Image reprinted with permissions from [71] (Copyright © 2024 Springer Nature, American Association for the Advancement of Science, Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International Licence).

Figure 2. (a) Predicted structures and interaction of designed proteins with three different ligands. Image reprinted with permissions from [68] (Copyright © 2023 American Association for the Advancement of Science, Creative Commons Attribution 4.0 International Licence). (b) Example of protein designed using Alpha Fold 3. Image reprinted with permissions from [68] (Copyright © 2023 American Association for the Advancement of Science, Creative Commons Attribution 4.0 International Licence). (c) Schematic representation of training of generative models for de novo drug design of molecules. Process includes using databases and iterating over model weight adjustments and accessibility scores to design novel drugs to meet specifications. Image reprinted with permissions from [73] (Copyright © 2023 Springer Nature, Creative Commons Attribution 4.0 International Licence). (d) Schematic representation of TxGNN processing of queries to calculate predictions for drug–disease relationships. Approach takes into account interconnectedness between both factors. Predictions from model align with off-label prescriptions from clinicians in large healthcare systems. Image reprinted with permissions from [71] (Copyright © 2024 Springer Nature, American Association for the Advancement of Science, Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International Licence).

4.2. Solvent Usage

Solvents represent a major contributor to pharmaceutical synthesis. It is estimated that only around 35% of solvent waste generated by the US pharmaceutical industry is recycled, and about 45% of the energy from the incineration of the remaining 65% is recovered [75]. Solvents are used across the whole pharmaceutical synthesis process, from reactions to the extraction of compounds, and they represent the majority of mass utilisation (between 80 and 90%) [76]. An average of six different solvents is used in any given process, with dichloromethane being the most common compound. Dichloromethane is a largely unregulated ozone-depleting compound, with emissions that have increased over the years. This increase could significantly delay the recovery of the ozone layer [77]. Moreover, it has shown negative environmental and toxicological properties [78]. As such, there is a need for greener solutions to reduce the impact of solvent use in the pharmaceutical industry. AI algorithms have been employed to allow a selection of alternative solvents based on their physical–chemical properties [79]. Some advanced models can also predict reactivity rates in asymmetric catalysis [80]. However, in most cases, AI models do not include environmental considerations to support the selection of solvents.

The incorporation of deep eutectic solvents (DESs) represents a promising approach in this field. DESs are a family of green solvents made by mixing two or more components that interact through hydrogen bonding, resulting in a eutectic mixture with a melting point lower than that of its individual constituents [81]. They are considered an alternative to ionic liquids and conventional organic solvents due to their low toxicity, biodegradability, and ease of preparation [82]. Traditional industrial solvents amount to 80–90% of the total mass used during a reaction in the pharmaceutical industry, and 80–85% of the waste produced [83]. This leads to high exposure rates, either through the volatiles produced or through their presence in the environment [84]. Moreover, solvent production and disposal plays a major role in the CO₂ emissions of the pharmaceutical industry. While the process material intensity for API manufacturing can vary, recent estimates indicate that approximately 40–50% of the industry’s carbon footprint can be attributed to solvent use. This is a consequence of their extensive use (accounting for 50–60% of materials required per kg of API produced) and the energy-intensive processes required for their production and waste management [75].

DESs are emerging as greener alternatives to conventional organic solvents due to their lower environmental impact and sustainability benefits. They can be made from natural sources including polyalcohols, sugars, or organic acids, including amino acids [85]. These solvents have the advantage of displaying low toxicity and having high biodegradability [85], and can potentially be synthesised through sustainable routes. However, the sustainability of these substances needs to be assessed on a case-by-case basis. In the case of citric acid, production is carried out through fermentation, leading to high water consumption (about 153 L kg⁻¹) and releasing CO₂ in the process [86].

While the development of DESs is still in an early stage, multiple applications have been described in the pharmaceutical industry, from helping in the delivery of drugs to supporting their synthesis and helping with the dissolution of reagents [87]. DESs have also been used in the extraction of bioactive compounds from plants [88], the green synthesis of pharmaceutical ingredients [89], and the solubilisation and crystallisation of pharmaceutical ingredients [90] (Figure 3a).

Despite the growing interest in DESs, their adoption in the pharmaceutical industry is limited by several challenges; the range of available DES formulations remains narrow, and their physicochemical properties need further characterisation. Additionally, regulatory approvals and industry-wide standardisation are necessary for their widespread implementation. To overcome these barriers, new methodologies for identifying and designing DES-compatible compounds must be developed. AI-driven solvent design, coupled with experimental validation, could significantly expand the repertoire of available DESs, ensuring they meet the stringent requirements of pharmaceutical applications.

New AI models have been developed to support the discovery of deep eutectic solvents with desired properties for the pharmaceutical industry. However, the limited amount of experimentally validated data represents a challenge in the development of AI algorithms for predicting DES formation. Abbas et al. [94] developed a machine learning approach to predict the formation of DESs which would allow for the discovery of new molecular candidates. This algorithm was based on modelled data on the hydrogen bond formation of compounds using molecular simulations. The library employed only included examples of 186 simulated systems and 34 experimentally verified cases as a validation dataset. Despite the limitations of the dataset, the model achieved a good performance, achieving an average ROC-AUC score of 0.8 in the validation set using the hydrogen bond lifetimes. However, this model required complex simulation data to calculate the hydrogen bonding lifetime, which could limit its scalability. Alternative approaches explored the prediction of melting temperatures for the screening of DES components. Lavrinenko et al. [95] trained a Support Vector Regression model using information from 1648 mixtures’ melting temperatures for 237 experimentally validated DESs. While the final algorithm showed good performance in the prediction of eutectic temperatures (R² = 0.74, RMSE = 22.5), its accuracy in eutectic composition was relatively low (R² = 0.34, RMSE = 0.11). The authors attributed these results to the limitations of the methodology, considering only the molar ratio of the lowest temperature predicted as the eutectic point. Diffusion models have also been employed which use the quantum mechanical properties of compounds (HOMO, LUMO, Free Energy, and heat capacity). These models could be used to generate new structures while achieving a good performance on prediction tasks (R² = 0.93). However, the datasets used for training were small (402 DES compositions).

Besides the design of new DES compounds, AI has been employed to select compositions with optimal characteristics through the prediction of their physical–chemical properties (Figure 3c). Among those, density is a key parameter in the extraction and purification of chemicals, especially liquid–liquid extractions. Common methods for extraction, such as counter-current extraction columns, rely on the use of solvents with different densities for phase flow control [96]. As such, multiple models have been focused on predicting DES densities [97], as well as other relevant properties, such as heat capacity [92], among others. A summary of the machine learning models trained to optimise DES design is provided in

Table 3. Machine learning models developed to optimise design of DESs.

Dataset	Model Prediction	Algorithm	Performance	Ref.
186 (simulated systems) + 34 (experimentally verified)	DES formation	Support Vector Machine	Average ROC-AUC score = 0.8 (experimentally validated)	[94]
237 experimentally validated DESs	Melting temperature	Support Vector Regression	R2 = 0.74, RMSE = 22.5	[95]
530 DESs	Heat capacity	Neural Network Multilayer Perceptron	AARD = 4%	[92]
402 different DES compositions		Diffusion model with self-/cross-attention	R2 = 0.93 (experimentally validated)	[98]
435 records	Hydrogen sulphide (H2S) elimination capacity	Cascade neural network	MAE = 0.02 MSE = 0.0031 AARE = 3.03 R2 = 0.99943 (experimentally validated)	[99]
-	Melting temperature, density, viscosity (model made available as Open Source)	CatBoost	R2 = 0.6–0.91 AARD = 3.14–19.05% (experimentally validated)	[100]
-	Solubility	Non-linear Support Vector Regression	Experimentally validated	[93]
1000 labelled mixtures of DESs and NADESs	DES formation	Transformer-based neural network model	F1-Score = 0.82 (experimentally validated)	[101]
1239 records	DES density	Least-squares Support Vector Regression	R2 = 0.99798 MAPE = 0.26%	[102]

.

4.3. Synthetic Pathway Optimisation

AI can aid in the design of efficient reaction pathways for the pharmaceutical industry. A reduction in the number of synthetic steps can contribute, for example, to reducing the amounts of solvents used during a chemical synthesis process. This has led to some companies to reducing their usage from 94 kg_solvent/kg_API to 75 kg_solvent/kg_API within just five years, with a consequent reduction in environmental impact [76]. AI has been used in organic chemistry for over 50 years. One of the first manuscripts in the field dates back to 1969, when Corey and Wipke developed a computer-aided synthesis planning program [103]. Since then, multiple models have been developed to simplify reaction pathways. However, the field’s evolution has progressed slowly, given the time required to evaluate results experimentally.

One of the key applications of AI in this field involves the simplification of reactions and purification processes. This approach has enabled the one-pot synthesis of trans β-lactams, which typically requires multiple steps [104]. The algorithm created reaction rules using the training set of reactions without the need for human inputs. It proposed over 2000 possible reactions, from which researchers selected a one-pot preparation route. Similarly, Takabatake et al. [105] developed a synthetic route for the common intermediate methyl 2-formylbenzofuran-7-carboxylate. However, so far, the range of reactions that can be optimised using this approach is limited due to the lack of reaction data.

To circumnavigate the challenges of the poor availability of large chemical datasets and improve the interpretability of variables across chemical spaces, multiple frameworks have been developed. King-Smith et al. [106] proposed a high-throughput experimentation method to analyse the impact of variables among different reactions. This method was able to find the most important variables for a given reaction, as well as the impact of the reagents. Thus, from reducing solvent use and synthetic complexity to enabling one-pot reactions and data-driven optimisation, AI has already demonstrated the capacity to transform traditional workflows. However, challenges remain, particularly concerning data availability and experimental validation. Addressing these limitations through collaborative efforts, such as combining high-throughput experimentation, Open Access datasets, and robust modelling frameworks, will be essential for unlocking the full potential of AI in green chemistry. As these tools continue to mature, they are poised to play a central role in creating more sustainable, efficient, and innovative pharmaceutical manufacturing processes.

5. Artificial Intelligence in Drug Waste Reduction

Waste generated from the pharmaceutical industry has a significant footprint in the environment. Some drugs, for example, can act as hormone disruptors for certain species, inducing the feminisation of male fish [107] or collapsing populations [108]. In particular, household pharmaceutical waste is becoming a global challenge given its deleterious effects and persistence in the environment. In recent years, the demand has also increased. It is estimated that the global consumption of antibiotics increased by 65% between 2000 and 2015 [109], and about three quarters of households hold, on average, 138 g of unwanted or unused drugs, of which 60% will be wasted [110]. The main sources of pharmaceuticals in the environment are human waste, including excretion through faeces and urine, and incorrect disposal [111]. As such, new strategies to minimise the release of drug components into the environment, as well as to accelerate their biodegradability, are needed.

5.1. Reduction in Drug Waste Through Dosage Optimisation

A potentially viable route to minimise pharmaceutical waste is the personalisation of patient dosage [112]. Zhu et al. [113] developed a machine learning algorithm to optimise dosage through non-invasive clinical parameters, such as demographics, age, weight, and co-medication status. The algorithm was trained using data from 1141 therapeutic drug-monitoring measurements from 347 patients, demonstrating good precision when determining the dose-adjusted concentration (Mean Absolute Error = 8.7 μg mL⁻¹ g⁻¹ day). However, the potential savings in environmental footprints from this approach have not yet been quantified in real-world settings. Instead, the prescription of pharmaceuticals only when they are required seems to represent a more effective approach to reducing the footprint of pharmaceuticals. AI has extensively been used in this field, incorporating patient data (i.e., age, gender, weight) as well as information from electronic records [114].

The prediction of absorption, distribution, metabolism, and excretion (ADME) represents a key tool for determining drug dosing [115]. AI-driven approaches have also demonstrated significant potential in predicting ADME properties, reducing the reliance on extensive in vitro and in vivo experiments. For instance, Satheeskumar et al. [116] tested multiple machine learning algorithms to predict clearance, volume distribution, half-life, and bioavailability, using a dataset of 10,000 compounds, with high accuracy. Alternative models have been developed recently, incorporating further molecular considerations, such as docking, and physicochemical molecular features [117]. However, there are multiple ethical considerations when using AI in prescriptions [118]. These include privacy, data protection, and the need for informed consent.

Despite the limitations, dosage prediction for particularly high-environmental-footprint compounds such as anaesthetics could lead to the greatest impact. Emissions derived from inhaled anaesthetics account for at least 3% of the total healthcare footprint [6]. While the figures of how much these gases contribute to climate change can vary greatly due to inconsistencies in the methodology used for assessment, it is estimated that they amount to 0.01–0.1% of total global greenhouse gas emissions [119], with annual emissions of 5 × 10⁶ tonnes CO₂e/y [120]. This is a consequence of the high greenhouse effect of common anaesthetic gases, such as nitrous oxide, and their long lifetimes [119]. Current models in this field have been used to predict the effects of anaesthesia in patients or to control their continuous administration [121,122]. Some of these models employ additional electroencephalogram monitoring data to predict the depth of anaesthesia [123]. However, they are focused on monitoring patient consciousness to administer gas anaesthetics, which produce high amounts of waste due to their intensive usage and uncontrolled escape into the surrounding environment, paired with the poor efficiency of scavenging systems [124]. Some studies have explored alternative methods to optimise anaesthetic administration, with the potential to reduce environmental impact. In particular, increasing the use of intravenous anaesthetic can reduce the overall consumption of volatile agents without compromising patient safety. Schamberg et al. [125] developed a reinforcement learning algorithm to improve the administration of propofol anaesthetics, leading to low performance errors (1.1% ± 0.5). This model was trained using simulated data. However, these advancements highlight the potential for reducing the environmental footprint of drug compound usage, especially anaesthesia, through predictive modelling. Moreover, the amount of experimentally validated data is low, which hinders the training of accurate models. A summary of the machine learning algorithms used in drug dosage optimisation is provided in

Table 4. Comparison of AI models used in prediction of pharmacokinetics and dosing of drugs.

Dataset	Model Prediction	Algorithm	Performance	Ref.
1141 therapeutic drug-monitoring measurements from 347 patients	Dose-adjusted concentrations of lamotrigine	Extra-trees regression algorithm	MAE  =  8.7 μg mL−1 g−1 day Mean Relative Error (%) =   3% (experimentally validated)	[113]
1099 patients, each described with 280 attributes	Postoperative analgesic requirements	Decision trees	Prediction accuracies of total analgesic consumption of 80.9% (experimentally validated)	[122]
10,000 bioactive compounds from ChEMBL database	Clearance, volume distribution, half-life, bioavailability	Stacking ensemble model	R2 = 0.92 MAE = 0.062	[116]
Simulated pharmacokinetic and pharmacodynamic data	Anaesthetic dosing	Reinforcement learning algorithm	Median episode median performance error of 1.1% ± 0.5 (experimentally validated with retrospective data)	[125]
210 case data	Increase in flow rate of remifentanil	Long short-term memory	Specificity = 0.73 Sensitivity = 0.66 ROC-AUC = 0.75	[121]
13 processed EEG parameters from 200 patients	Anaesthesia scores	Reinforcement learning algorithm	R2 = 0.94 (experimentally validated)	[123]

.

5.2. Design of Biodegradable Drug Delivery Systems

While the optimisation of drug dosage can lead to a reduced waste and minimise environmental impact, most common drugs have very low biodegradability [126]. In particular, hydrophobic drugs with low water solubility have been shown to bioaccumulate, triggering concern [127]. Recent studies estimate that over 25% of rivers globally might contain active pharmaceutical ingredients at levels higher than those considered safe for aquatic life [128]. As such, the development of biodegradable drugs shows promise in mitigating the pharmaceutical environmental footprint. These are normally developed using biodegradable nanoparticles, encapsulating the drug of interest inside a polymeric matrix [129]. Traditionally, synthetic polymers such as poly (ethylene glycol) [130,131], poly (lactic acid) [132,133] or poly (lactic-co-glycolic acid) [134] have been used alone, or in combination with other matrix elements. Natural polymers such as cellulose [135] or starch [136] have also been employed. However, their use is limited due to their general variability in purity and, in some cases, the need for chemical treatments.

Biodegradable drug delivery systems have been designed using AI to allow for the precise optimisation of drug formulations, enhancing their biodegradability while maintaining therapeutic efficacy. AI-driven computational models can predict the interactions between different polymers and active pharmaceutical ingredients [137]. This interaction allows the study of drug release behaviours [138], including those of stimuli-responsive hydrogels, which ultimately enables the selection of optimal biodegradable carriers. These models could potentially be paired with recent AI algorithms to predict the biodegradability of polymers and allow for the in silico selection of the most optimal materials for drug design with a Mean Absolute Percentage Error (MAPE) between 7.6 and 16.99% [139].

Recent advancements in modelling and AI have also enabled the discovery of new biodegradable polymers by predicting the molecular structures that exhibit desirable properties, such as high drug-loading capacities and tuneable degradation rates [140]. AI can also integrate multi-objective optimisation strategies, balancing biodegradability, drug stability, and manufacturing feasibility to create next-generation eco-friendly pharmaceuticals. However, the application of AI in the drug design space, specifically to reduce the environmental impact of treatments, has been limited so far. A summary of the machine learning algorithms developed to predict drug–polymer interactions and polymer biodegradability is provided in

Table 5. Comparison of AI models used in prediction of biodegradable drug delivery systems.

Dataset	Model Prediction	Algorithm	Performance	Ref.
540 experimental data points	Release kinetics	Support Vector Machine	R2 = 0.998 RMSE = 0.3701 MSE = 0.137 MAPE = 0.944 (experimentally validated)	[137]
76 biodegradation curves	Biodegradation kinetics (%)	Long short-term memory neural network	MAPE = 7.6–16.99% MAE = 0.018–0.027 MSE = 0.0004–0.001 RMSE = 6.71–11.33 (experimentally validated with validation dataset)	[139]
Data from 133 experiments	Polymer properties (Young’s modulus, flammability)	Artificial neural network	MRE = 21% (experimentally validated with molecular dynamic simulations)	[140]
15,000 data samples	Drug release kinetics	Decision Tree Regression	R2 = 0.94652 RMSE = 6.04 × 10−5 MAE = 4.83 × 10−5	[141]
500 datapoints	Drug release kinetics	Random Forest	R2 = 0.99 MAPE = 0.002 (experimentally validated)	[142]

.

6. Current Status of Molecular Design for Sustainability and Future Perspectives

The integration of sustainability into molecular design is gaining momentum across the pharmaceutical industry, driven by regulatory pressure, environmental concerns, and the need for more efficient R&D pipelines to reduce production costs. Traditionally, the discovery and optimisation of active pharmaceutical ingredients prioritised efficacy and safety, with less emphasis on the environmental impact of synthetic routes, degradation products, or raw material sourcing. However, a paradigm shift is underway. There is an increasing number of reported manuscripts focusing on how specific aspects of drug life cycles can be optimised to reduce their environmental impact. The need for this shift has been highlighted by key pharmaceutical stakeholders, including the WHO [143].

Currently, sustainable molecular design leverages advances in cheminformatics, high-throughput screening, and retrosynthetic analysis to identify and optimise molecules with improved biodegradability, lower toxicity, and reduced reliance on rare or hazardous reagents (Figure 4). Frameworks such as the ACS Green Chemistry Institute’s Pharmaceutical Roundtable (GCIPR) [144] have accelerated best practice sharing, and metrics like Process Mass Intensity (PMI) are increasingly being adopted to quantify sustainability performance [145].

AI is beginning to reshape this landscape. Machine learning models trained on large-scale chemical and biological datasets are enabling the generation of novel compound libraries optimised for pharmacological activity, synthetic tractability, and environmental compatibility. Predictive models for aqueous persistence, bioaccumulation, and toxicity are being integrated into molecular design workflows, offering a proactive approach to reducing the environmental footprint of future pharmaceuticals. However, in some cases, there are limitations regarding the size of the available datasets, which limits the accuracy of models.

Besides the need for data to enable the training of accurate models, there is a lack of life cycle assessments, especially regarding the environmental impact and standardisation of methodologies used to study drug compounds. This could reveal further challenges and unique opportunities for improvement [19]. The fusion of AI with system-level sustainability assessments and life cycle analysis holds transformative potential. Generative AI models will be able to design compounds that meet complex multi-objective criteria, including efficacy, safety, manufacturability, and sustainability [13]. Moreover, the emergence of generative models and reinforcement learning could open new frontiers in green-by-design molecules, tailored for minimal environmental impact throughout their life cycle.

To fully realise this potential, cross-sector collaboration will be essential. Interdisciplinary partnerships between computational chemists, environmental scientists, process engineers, and regulators will help ensure that AI-driven molecular design not only accelerates innovation but also aligns with broader sustainability goals. As the field evolves, the pharmaceutical industry stands at the cusp of a more responsible and resource-efficient future, where AI is not just a tool for drug discovery but a cornerstone of sustainable innovation. Moreover, Open Source initiatives making molecular information available to researchers across all fields will be key to allowing the development and training of more accurate AI models. Some examples of Open Source initiatives include the Open Reaction Database [146] or PubChem [147]. These initiatives, alongside trained online AI platforms, are playing increasingly central roles in accelerating innovation. They provide researchers with powerful tools for virtual screening, property prediction, protein structure modelling, and synthesis planning, often without the need for specialised local infrastructure. Notable examples include AlphaFold [148] for protein structure prediction, SwissADME [149] for pharmacokinetic profiling, DeepChem [150] for molecular property prediction, and ProGen [151] for designing novel protein sequences with desired functions. These platforms can lower the barrier to entry for sustainable molecular design and support the rapid prototyping of drug candidates.

7. Conclusions

AI is playing an increasingly pivotal role in enhancing the sustainability of the pharmaceutical industry. From optimising drug dosage and enabling the design of biodegradable compounds to streamlining synthesis pathways and discovering green solvents and catalysts, AI offers transformative potential across the pharmaceutical value chain. However, several systemic limitations hinder the full realisation of AI’s environmental impact. In some areas, such as in the design of deep eutectic solvents, progress is constrained by the lack of large, high-quality databases. This leads to the need to calculate quantum parameters to increase the accuracy of models. In contrast, other domains, such as enzyme engineering and catalyst discovery, involve computationally intensive models, often involving billions of parameters that require vast and complex datasets. This leads to high energy costs for training, which have not yet been fully and accurately quantified. In fact, across most applications, there is a general scarcity of robust environmental data, limiting our ability to quantify sustainability benefits.

A critical gap remains in the integration of life cycle assessments into AI-driven pharmaceutical development. The general lack of drug-specific life cycle assessments makes it challenging to evaluate the true environmental implications of AI interventions. Moreover, inconsistencies in the size, quality, and structure of training datasets, as well as the lack of standardised metrics for evaluating model performance, further complicate cross-study comparisons and model reproducibility.

Despite these limitations, the breadth of potential impact remains wide. AI can enable the multi-objective optimisation of drug candidates, balancing pharmacological efficacy with factors such as biodegradability, synthesisability, and environmental persistence. The ability to incorporate sustainability considerations into early-stage design and development offers a unique opportunity for the pharmaceutical sector to align innovation with broader climate and environmental goals.

Going forward, addressing these challenges will require coordinated efforts: expanding Open Access datasets, standardising performance metrics, incorporating sustainability endpoints into training objectives, and fostering collaboration between AI researchers, chemists, and environmental scientists. More importantly, with the right frameworks, AI can become not just a tool for faster and more effective drug discovery, but a catalyst for a more sustainable pharmaceutical future.