Antibiotic resistance is one of the great problems of the 21st century. Since their mass use began around the 1950s, these drugs have succeeded in turning deadly diseases of necessity into mere inconveniences that last only a few days. However, their indiscriminate use has provoked a reaction from pathogens. Today, an increasing number of micro-organisms have become resistant to these same antibiotics, and they seem to evolve faster than the time in which mankind develops new formulations.
In this worrying scenario, artificial intelligence can give humans an advantage in both creating and searching for new molecules with antibiotic potential. Thanks to massive data analysis and a deep understanding of chemical interactions between biological systems, the future of antibiotic development looks set to take place largely within data servers. While there is much promise in generating new molecules, this innovation must be mindful of its limitations, as one of the three great pillars of humanity - health - depends on them.
INSIDE. Algorithms to fight against superbugs
Antibiotic compounds have been known for thousands of years. Already in the Ebers papyrus, dated 1550 BC, remedies and poultices containing substances with antibiotic powers are recorded. However, in these ancient texts the medicinal powers of herbs and mushrooms are attributed more to the power of the healers and the rituals used than to the compounds present in the mixtures. At that time diseases were demonic mysteries, caused by malevolent beings that entered the body and lodged in various organs and tissues, causing pain and illness.
This thinking continued to influence medicine for centuries. However, with the invention of the microscope by Dutch merchant Anton van Leeuwenhoek in 1676, microbiology was born - a window into a new world of creatures beyond the human sense of sight. After two centuries of microbiological research and countless experiments, scientists began to realise that these microorganisms could have a major impact on health. Diseases were no longer the stuff of spells and spirits, but took on a more earthly dimension. This important change in mentality was made possible by well-known names such as Louis Pasteur, his wife Marie Anne Laurent, Joseph Lister and Robert Koch. Thanks to them, the foundations of microparasitology were laid and it was established that certain micro-organisms were capable of producing diseases. In other words, the enemies that had to be eliminated in order to cure an infectious disease were given names and surnames.
Following this shocking discovery, experiments with chemical and biological compounds to attack these micro-organisms began. One of the important breakthroughs occurred on 17 December 1897, when the young researcher Ernest Duchesne presented his thesis with the title «Contribution to the study of life competition between micro-organisms: antagonism between moulds and microbes».». In this thesis, supervised by Gabriel Roux, a disciple of Pasteur, Duchesne was able to observe how the fungus Penicillium glaucum prevented bacterial growth.
But Duchesne went a step further. During his thesis, he inoculated lethal bacteria for guinea pigs such as E.coli y Staphylococcus typhy (the cause of typhoid fever) together with the fungus P.glaucum. His surprise was that those animals in which he had performed the double inoculation survived the disease, while those infected only with the bacteria died within a few days. Unfortunately, Duchesne was unable to identify the antibiotic substance and decided to abandon his studies for a military career. Like other discoveries, his promising thesis was forgotten in a drawer.
Until 1928, when Dr Alexander Fleming had his moment of serendipity on his return from holiday. When he returned to the laboratory, he noticed that some of the bacterial cultures from Staphylococcus he had left behind had become contaminated with fungi. Instead of immediately throwing away the plates, Fleming noticed that, around the places where these fungi had grown, there was a substance that Fleming called «mould juice», which prevented bacterial growth. That substance was penicillin, and it would go on to become one of the most important medical discoveries in history.[i]. Then, in 1942, researchers Howard Walter Florey and Ernst Boris Chain devised a system that allowed penicillin to be extracted in sufficient quantity to be applied to patients. The novel drug played a key role in the wartime context of World War II, treating infections on the battlefield and in field hospitals that would otherwise have been fatal to soldiers. Such was the success of penicillin that, once the war was over, it became part of the arsenal of doctors in hospitals around the world, ushering in the antibiotic era.
The researchers quickly realised that penicillin was not a panacea, but that there are species of microorganisms against which penicillin has no effect. Worse still, they were also able to observe how penicillin-resistant strains of previously sensitive species were appearing.[ii]. Therefore, a search for new antibiotics began, which could either be found naturally or could be synthesised from chemical knowledge. But this solution was only temporary, as soon micro-organisms resistant to the new antibiotics also began to appear.[iii].
These resistant populations arise due to chance in the genetics intrinsic to life. In a population of bacteria there are colonies that are slightly different from each other. Some of them may have in their DNA the ability to utilise other food sources, some may have the ability to better attach to surfaces, and, as for the topic at hand, there is the possibility that some may contain the information to expel or degrade, and thus inactivate, certain antibiotics.
When an antibiotic is used to treat a disease, all populations of pathogenic bacteria sensitive to the antibiotic are killed, leaving only resistant colonies. These, left without competition for food or space, have a free hand to grow. They begin to occupy all the space previously colonised by a diverse population, resulting in a second infection. However, the bacteria present in this second wave are descended from antibiotic-resistant colonies and therefore known treatments will not work, complicating the cure.
But this is not the only method of acquiring resistance. It has also been observed that certain bacteria can transmit information horizontally, i.e. one bacterium can pass it on to another, even if they are not of the same species. Therefore, a bacterium that is not even pathogenic for humans could pass this resistance gene to one that is, giving rise to a resistant bacterium that is dangerous for humans.
Thus, the overuse of antibiotics, both in animal husbandry and health, has allowed the spread of antibiotic resistance genes in several human pathogenic species. The most worrying of these are listed in the report “Antibiotic Resistance in the Human Pathogenic Species".“WHO bacterial pathogen list”issued by the World Health Organisation (WHO) and updated to 2024. In it, 24 pathogens covering 15 bacterial families and different levels of resistance can be found. These include Salmonella, which causes salmonellosis; the Neisseria gonorrhoeae, which causes the well-known sexually transmitted disease; or the Pseudomonas aeruginosa, But particularly worrying is the case of Mycobacterium tuberculosis, which causes tuberculosis, a lung disease whose cases have increased in recent years, some of them with strains resistant to all known antibiotics.[iv].
As the report also indicates, antibiotic resistance directly caused an estimated 1.27 million deaths in 2019. In addition, it was involved in some way in around five million deaths worldwide, a number that exceeds the total number of deaths from other epidemics such as AIDS or malaria. Critically, an estimated 20% of these deaths occurred in children under the age of five, an indication of the severity of the problem. If this trend continues, computational modelling warns that by 2050, around 10 million people a year may die from antibiotic resistance.
Since their discovery around a century ago, antibiotics have therefore become one of the cornerstones of healthcare. Proof of this is the large number of antibiotics that have been developed over the years. Therefore, the fact that there are micro-organisms against which they do not work may pose a challenge for the future.
There are currently several dozen different antibiotics that can be classified according to their mechanism of action (how they affect bacteria), according to their spectrum of activity (whether they target specific types of bacteria or are broad-spectrum), or according to their chemical structure.
Classification by chemical structure, in turn, allows the identification of certain families of antibiotics that start from the same base, but have certain modifications that slightly change their applications. These families are considered to be rather static, since, although new members have emerged in most of them, no new antibiotic family has been discovered since 1987.[v]. This is attributable to bottlenecks in traditional forms of discovery. As mentioned above, most clinical antibiotics originate from micro-organisms that produce them naturally in their life cycle.
For this reason, scientists “in boots” have established synergies with those “in coats” and have gone to the ends of the earth to extract samples in order to find new substances. Once obtained, by characterising their chemical structure and understanding their mechanism of action, they have managed to create modified antibiotics that maintain their bactericidal or bacteriostatic power, but are less likely to generate resistance because they are not natural. However, there is also resistance against entire families of antibiotics, which is a much more complex challenge to tackle. For this reason, in recent years, these techniques have fallen into disuse and have been replaced by screening of hundreds of molecules based on the prediction of bonds between molecules.
This method is also called reverse pharmacology, because scientists turn the usual path of knowledge on its head. Instead of finding a specific compound or enzyme and then searching the genome of the fungus or bacterium for the specific gene, they do it the other way around. That is, the genome of the micro-organism is studied in detail, and possible targets are selected to develop compounds that can kill it. This depends on previous knowledge of the micro-organisms.[vi].
According to the WHO, there are currently only 77 antibiotics in development, an insufficient number to address the problem of resistance in microorganisms. Of these 77, most are derived from existing antibiotics, so bacteria with resistance could quickly adapt mechanisms. Therefore, of the 77, very few are expected to reach the market.[vii]. This, coupled with the fact that experts believe there is no clear market for novel antibiotics, explains why large companies have no incentive to develop them, and why it is too big a risk for small biotechs. Fortunately, increasing computational power coupled with artificial intelligence (AI) techniques may turn the tide and allow antibiotics with novel properties to be created.[viii]. Figure 1 explains what their major contributions can be.
One of the methods of combating the antibiotic crisis has been to study in depth the defence mechanisms of resistant bacteria. To this end, bacterial genomes have been analysed, antibiotic susceptibility tests and biochemical activity panels have been applied. All these methods have collected a huge amount of data that can be used to train an AI. Specifically for a type of AI called machine learning, which uses statistical algorithms to identify sophisticated relationships between data and extrapolates these results each time new data are added. Within the machine learning, the deep learning uses neural networks to process data using interconnected nodes at different layers. These models are typically trained on existing task-specific data and are used to draw conclusions from new data. That is, the model is first trained to learn to perform a task with known data, and then set to perform that task with unknown data. This data is then tested in the traditional way to check its success rate. In this way, AI can be used for a variety of methods, such as virtual screening of molecules, molecular generation of compounds and the discovery of more naturally occurring antibiotics.[ix].
Virtual screening can be a less costly and time-consuming alternative to conventional HTS, as it takes advantage of machine learning techniques to predict novel molecules with a specific molecular property by searching vast chemical repositories. in silico. While conventional HTS campaigns have an upper limit of a few million compounds, molecular property prediction models only require the screening of thousands or tens of thousands of compounds and can then be applied to virtually screen much larger chemical libraries, in the order of tens of millions of compounds to find the optimal ones. Databases of these molecules are becoming increasingly large and sometimes open access, so that chemists, biotechnologists and pharmacists can synthesise the most promising compounds.[x].
Although molecular screening is a more powerful technique than traditional ones, it is subject to the limitation of known molecules registered in databases. However, by training an AI with notions of chemistry, it can eventually generate molecules with a possible structure that is not currently contained in any living being. In this way a model can test the efficacy of these novel molecules for which resistant micro-organisms have no response, as they have never encountered them before. Among the major breakthroughs in the field is AlphaFold3[xi], Google DeepMind, which models the structure of complexes containing combinations of biomolecules, including proteins, nucleic acids, small molecules and ions. In other words, it can understand how antibiotics will react to the different structures of a microorganism. The biggest problem with this technique is often its synthesis. Sometimes, we do not know the techniques needed to synthesise the organic molecules proposed by AI. Or it may be impossible to synthesise them at all. For this reason, there is a need for guardrails for this type of AI, with which to stick to the molecules that are possible.
Finally, AI can also help in the search for new natural sources of antibiotics. With the huge genetic databases available on virtual platforms, gene mining is a very good possibility to explore antibiotic molecules that have gone unnoticed. This method is particularly relevant for genomes of species that cannot be cultured in the laboratory. Since they cannot be grown in culture, it is necessary to introduce the antibiotic production genes into another, more manageable species by genetic engineering. Once cultured, the efficacy of the particular metabolite can be extracted and tested. However, this technique works well for sequences that produce known antibiotics, but less well for those yet to be discovered, as it is difficult to get them to “imagine” the properties of a promising metabolite. Therefore, its implementation is complex.
Antibiotic resistance puts today's healthcare system at serious risk. To combat it, AI is not a panacea, but it is a set of powerful tools that humans with expertise in the field can harness to defeat the rise of multidrug-resistant bacteria. Today, with increasingly sophisticated models, AI is still limited by the data on which the model is trained. Fortunately, this limitation is fading and has enabled the creation of systems that outperform humans in the analysis of molecular structures.
However, it is important to bear in mind that, as Dr Ian Malcolm, the beloved Jurassic Park character played by Jeff Goldblum in the film of the same name, would say, “life is making its own way”. In other words, micro-organisms will eventually develop resistance to new antibiotics, so it is important to develop them, yes, but also to use them intelligently in order to delay this as much as possible.
IN ACTION. A field with a lot of research and little industry
There are long-established and proven methods for discovering new antibiotics. Unfortunately, most of them ran out of steam decades ago, so we have spent years trying to find new ways to kill these bacteria.[xii]. Artificial intelligence (AI) is not new to the search for antibiotics. Before the advent of generative AI, researchers often used algorithms to scan libraries of existing drugs and identify those most likely to act against a given pathogen. Screening up to 100 million known compounds[xiii], can yield very interesting results, but it falls far short of tracking all possible options: it is estimated that there are about 10¹⁴ drug-like molecules. There are also many generative algorithms for molecular design. ex novo, but they tend to build molecules, atom by atom, synthetically intractable in a laboratory[xiv]. The innovation race is still open. Recent advances in protein structure prediction and machine learning-driven molecule generation offer compelling reasons to revisit rational drug design.[xv].
Among the pioneers of the new AI cycle to solve that immeasurable gap is a group of researchers from Stanford Medicine and McMaster University. They pioneered a new model, called SyntheMol (molecule synthesiser).[xvi], to create the structures and chemical formulas for new drugs to eliminate resistant strains of Acinetobacter baumannii, one of the main pathogens responsible for antibacterial resistance-related deaths. The aim was to use AI to design completely new molecules, never before seen in nature.[xvii], on the condition that they could be synthesised in a laboratory. SyntheMol was trained on a library of more than 130,000 molecular building blocks and generated, in less than nine hours, around 25,000 possible antibiotics and the recipes for making them. Of these, only those that differed from existing compounds were selected to make it harder for bacteria to resist. Seventy compounds were obtained. The Ukrainian chemical company Enamine efficiently generated 58, six of which eliminated a resistant strain of A. baumannii when tested in the laboratory. However, four did not dissolve in water, leaving only two. After toxicity testing in mice, both proved to be safe.
This is one of the research models currently being developed that is helping to establish many of the general principles relevant to the discovery of new antibiotics. Although generative AI is still in its infancy, its impact is beginning to be devastating. Professor José R. Penadés and his team at Imperial College London have spent years researching and demonstrating why some superbugs are immune to antibiotics. When given the chance to try out the Google co-scientist tool, the scientist provided a brief instruction to solve the central problem he had been researching. The AI returned an answer within 48 hours.[xviii], had reached a similar conclusion to his own. Penadés could not help but contact Google to make sure that his computer had not been hacked.
From the University of Pennsylvania, the team led by the Spaniard César de la Fuente searches for new compounds in nature itself with the help of AI. Their work is based on the computational identification of more than 2,603 peptide antibiotics within the human proteome, plus 323 additional peptide antibiotics encoded in small open reading frames within the metagenomes of the human gut. There's a hidden world of peptide-based immunity[xix].
In fact, one million new antibiotic molecules have already been located, after using an algorithm to scan the Earth's entire microbial diversity, from the ocean to the human gut. It is the largest antibiotic discovery effort to date. A hundred DNA sequences were synthesised in the lab to assess their effectiveness against bacteria and 79% could kill at least one microbe, indicating their potential as antibiotics.[xx]. The result is AMPSphere, an open and accessible resource that provides information on antimicrobial peptides, including their sequences, original genes and biochemical properties. But De la Fuente has gone further and applied a machine learning model to search for similar molecules in Neanderthals and Denisovans, allowing him to launch the field of molecular de-extinction.[xxi]. From that work, a new deep learning model called APEX has been developed that can extract antibiotics from all known extinct organisms, including pre-clinical candidates from the woolly mammoth and other organisms. More than 37,000 encrypted sequences have already been discovered, some of them with potent anti-infective activity.
Even poisons are a huge and largely untapped reservoir of bioactive molecules with antimicrobial potential. Through deep learning, a research[xxii] has obtained more than 40 million venom-encrypted peptides, from which 386 candidates structurally and functionally distinct from known compounds were selected. He finally chose 58 for experimental validation and confirmed that 53 of them exhibit potent antimicrobial activity including in the case of confrontation with the Acinetobacter baumannii.
The new possibilities of AI are, however, trying to shake up an environment that is moving forward with a very rigid inertia that is difficult to redirect, as Figure 3 shows. Pharmaceuticals and private investors are operating far away from the development of new antibiotics and innovation has slowed down over the last four decades. Since the 1980s[xxiii], no new family of antibiotics is on the market, despite the fact that available compounds are steadily losing efficacy due to antimicrobial resistant bacteria (AMR). By mid-2025, only 32 antimicrobial drugs were in clinical development, and of these, only 12 could be considered innovative according to the WHO.[xxiv]. The fate for most of them will be failure in clinical trials and lack of regulatory approval.
Private investment in antimicrobial development is scarce because of the dysfunctionality of the market itself, which has meant that companies that have secured permits for new compounds in recent years have had to declare bankruptcy almost immediately. In fact, small companies account for 80% of proposed new antibiotic therapies, while 8% emerge from non-profit institutes and universities, and only 12% originate from large companies.[xxv]. Each formulation requires on average around $1 billion in development, and recouping that investment is extremely difficult for drugs that, unlike those that treat lifelong chronic diseases, are used for a short period of time - the duration of the infection. Between 2011 and 2020, venture capital invested $1.6 billion in antibacterial research companies and $26.5 billion in oncology companies.[xxvi]. As a result, it could happen that AI generates a large number of promising therapeutic candidates, but there is no funding necessary for them to complete clinical trials and reach patients. The G7 has issued a number of statements in support of so-called «pull incentives» for companies that successfully develop effective antibiotics, but only the UK[xxvii] and Italy[xxviii] have enacted significant measures in this regard.
The prognosis for patients with dangerous bacterial infections has worsened in recent years due to the spread of antibiotic-resistant bacterial strains and the stalled development of new treatment options.[xxix]. Antimicrobial resistance is projected to cause up to 1.91 million attributable deaths and 8.22 million associated deaths by 2050. Without further interventions, it will be unlikely to achieve the WHO's proposed 10% reduction in AMR mortality by 2030 and meet the requirements of the European Parliament, which in 2023 recognised AMR as one of the top three priority health threats in the EU.[xxx]. This is an issue that directly impacts the fight to reduce inequality gaps in the world, as low- and middle-income countries bear most of the burden of antimicrobial resistance.[xxxi] and have limited capacity to implement action plans due to insufficient staff and financial resources.
The failure of antimicrobial treatments cannot be attributed to antimicrobial resistant genes alone. Multiple factors play a role, including the complex mechanisms underlying critical clinical conditions such as sepsis. In fact, there is a significant and alarming increase in resistant infections occurring during hospitalisations.[xxxii]. Annually, about 1.7 million people develop sepsis in the US alone, and approximately 350,000 die during hospitalisation.[xxxiii]. The efficacy of antibiotics is then limited by high mortality rates (23-35%) and clinical complications. In addition to its benefits in combating pathogens, the integration of AI into healthcare systems could lead to more accurate antimicrobial prescribing.[xxxiv].
Research on antibiotics discovered through deep learning-assisted processes is still in the early stages. None of the drugs identified have yet begun clinical trials, let alone achieved regulatory approval. As mentioned above, there is a need to fully understand the context and ensure the quality of the training data. AI systems are trained in clearly defined environments, whereas the physical world often has complex and volatile underlying phenomena. This can call into question the validity of the generated solutions.
On the technology side, data quality, model interpretability and real-world implementation of the results obtained from the computation remain the big challenges ahead. And it is not strictly a question of quantity: large datasets can help train advanced modelling, but smaller, well-annotated and accurate datasets can provide more valuable information by avoiding excessive noise. International collaboration is needed to share information and knowledge between institutions. In addition, machine learning algorithms must be transparent to ensure the safety and efficacy of new drugs. Those based on deep neural networks, which are often described as black boxes, generate mistrust among health professionals and patients. Directly interpretable models are preferred in clinical settings as they are easily understood by all stakeholders.[xxxv].
In response to these concerns, authorities are working to create clear guidelines for the safe and ethical implementation of AI. Regulatory bodies such as the FDA in the US and the Medicines and Healthcare products Regulatory Agency (MHRA) in the UK have taken on the role of monitoring the use of AI in healthcare. The EU's General Data Protection Regulation (GDPR) sets out strict data privacy obligations, while the European Data Act, in force since January 2024, regulates access and use, including health data.
The AI Law is comprehensive[xxxvi] and puts a special emphasis on high-risk scenarios. In addition, the introduction of the European Health Data Space (EHDS)[xxxvii] will play a key role in supporting these regulations by ensuring the availability of health data on a large scale. Beyond the EU's borders, the AI Act could also impact the global market by setting rigorous standards for the development and use of AI.[xxxviii] that can be followed and adopted by other regulators and national or international authorities. The rapid development of AI technologies will, in any case, require a continuous process of reassessment and refinement of all regulations so that they do not create unequal market opportunities or risks for consumers.
On the fringes of the antibiotics race, AI has established itself as a mainstream technology for the pharmaceutical sector. Exscientia, a spin-off from the University of Dundee in 2012, was one of the first to apply AI to drug discovery. In 2024, the company took its first AI-designed drug candidate into human clinical trials, after a selection process that took just 12 months. US rival Recursion acquired the Oxford-based company for $688 million. By early 2025, Google DeepMind subsidiary Isomorphic Labs was confident that trials of its first AI-designed drugs would begin that year, as tech start-ups race to get real treatments off the ground.[xxxix]. There are more than 460 AI startups currently working on drug discovery.[xl] worldwide, a quarter of them in Europe.
As a way to facilitate a strategic and integrative vision, some experts advocate a multidisciplinary approach that integrates AI with other emerging technologies such as synthetic biology and nanomedicine to preserve the efficacy of antibiotics in the future.[xli]. The race to discover new antibiotics would thus benefit most from massive improvements in the accuracy and speed of machine learning algorithms, as well as cost reductions due to the affordability of GPUs (graphics processing units), increased data availability, and advances in research into the underlying algorithms.
AI can thus increase the usefulness of a data source such as the Drug Repurposing Hub.[xlii], which contains a hand-curated collection of thousands of existing medicines, including those that never gained regulatory approval despite passing clinical trials. It allows existing medicines to be applied to new diseases other than those for which they were originally developed, a pragmatic approach that can have a rapid and lower-cost impact. Another useful tool that could receive a boost is the ZINC database15 , which includes data on small molecules, including their biological activity and chemical properties. Alongside the former, it has been used in neural network approaches.[xliii]. Finally, RDKit[xliv] is an open source library for calculating molecular features.
Regarding currently available AI models, a study has analysed the usability, chemical validity and biological relevance of the six most prominent ones with 3D structure recognition. They are built on different technologies: diffusion, autoregressive, graph neural networks and language models.[xlv]. DeepBlock and TamGen could be considered the best on several criteria. The former decomposes molecules into synthetically accessible and chemically reactive building blocks. In a different methodological direction, the latter applies a pre-trained generative chemical language (GPT) model, combined with protein coders based on transformers, to generate molecules de novo, which improves chemical validity, synthetic accessibility and generation speed.
These include DiffSBDD, which ensures chemically valid and geometrically consistent results, and is able to adapt generative capabilities to various design scenarios, without the need for retraining. Pocket2Mol uniquely balances structural accuracy, generative efficiency and chemical validity, and to further improve structural fidelity, ResGen significantly accelerates the generation of drug-like molecules. Finally, TargetDiff ensures realistic molecule generation, conditional on protein binding sites. Overall, the six models constitute complementary approaches, varying mainly in their reliance on structural data or sequence data, the use of different generative methodologies and specific design goals.
In a comparative study, significant differences were observed in key metrics such as molecular validity, structural diversity, pharmacological similarity and target specificity. DeepBlock consistently outperformed its peers and had the highest proportion of chemically valid, structurally diverse and drug-like compounds. Surprisingly, after rigorous post-generation screening, it was found that 61% of all commercially available candidate molecules came from DeepBlock, TamGen contributed 34% and the remaining models contributed only 5%.
Scientists are particularly concerned about six highly virulent and drug-resistant bacterial species: Enterococcus faecium, Staphylococcus aureus, Klebsiella pneumoniae, Acinetobacter baumannii, Pseudomonas aeruginosa y Enterobacter spp. known as ESKAPE pathogens. The case of A. baumannii is of particular concern: it resists desiccation and disinfectants and is responsible for life-threatening, hospital-acquired infections of the skin, lungs, urinary tract, brain, bloodstream and soft tissues.[xlvi]. The World Health Organisation has classified it as[xlvii] as a critical priority for the development of new treatments and diagnostic tools.
The potential security impact has mobilised the Defence Advanced Research Projects Agency (Darpa) in the United States. Leading the TARGET (Transforming Antibiotic R&D with Generative AI to Stop Emerging Threats) project.[xlviii] has placed the social enterprise Phare Bio, together with the Collins Laboratory at the Massachusetts Institute of Technology (MIT) and the Wyss Institute at Harvard. It aims to expand the number of molecules currently screened for antibiotic activity, including the Broad Institute's Drug Reuse Center and the aforementioned ZINC15 library, which together hold 107 million candidate molecules. Will use deep learning to develop screening tools in silico to assess the efficacy of each candidate molecule as an antibiotic and, ultimately, as a drug. Finally, TARGET will validate each promising new discovery in terms of antibiotic activity and pharmacological properties, with the aim of identifying 15 promising new lines of research for new antibiotics, which would contribute to replenishing the global pipeline.
AI can benefit from developments in the field of virtual human twins, which simulate real-world scenarios while preserving the individual characteristics of each patient. TWIN-GPT is an extended language model (LLM)-based approach designed to create digital twins useful for clinical trials.[xlix]. AI company Unlearn.AI has simulated control groups in clinical trials, reducing the number of participants needed and speeding up timelines, and Sanofi is also testing compounds on digital patients before moving on to clinical trials. The European Commission has funded the European Virtual Human Twin (EDITH) project, in parallel to the launch of a call for proposals to advance the uptake of AI in healthcare under the EU4Health Programme.).
While the excitement around LLMs is justified, achieving their full potential will require a strategic approach to balance innovation with rigorous evaluation, ensuring that these tools are safely and effectively integrated into clinical practice.[l]. Moreover, as the deployment of generative AI progresses, access to the advanced graphics processing units (GPUs) with which large language models are trained and the AI applications derived from them are processed, while increasingly affordable, may become another barrier for academic researchers to overcome, compared to large industry teams. In some ways, the odyssey of antibiotic discovery with artificial intelligence can be said to reflect both the good and the less desirable aspects of the digital revolution.
IN SPAIN. AI use in pharma, still far from antibiotics, is a powerhouse
Although there are no specific examples in Spain of antibiotic research using artificial intelligence, there is a growing trend for companies and research centres to establish new ways of collaborating with intelligent technologies. Grifols has partnered with Google Cloud to use large language models (LLMs) as a lever for accelerating new biopharmaceutical therapies.[li]. This includes both the identification of potential therapeutic candidates and the management of clinical data. Almirall, for its part, is working on the development of new treatments for dermatological diseases with AI in collaboration with US-based Absci, which has a drug discovery platform.
In the field of research centres, the Foundation for Biomedical Research of the Ramón y Cajal University Hospital (IRYCIS) is collaborating in Spain with the French biotech company Owkin to optimise cancer treatment using AI. In a first phase, they are focusing on optimising therapeutic strategies for prostate cancer detected early, in order to identify which patients would benefit from specific therapies. By working with Owkin, the IRYCIS database is ready for the needs of the technology sector to be applied in future research using AI.
In the case of the Cima University of Navarra, Engineering-Tecnun and the Data Science and Artificial Intelligence Institute (DATAI), both also at the University of Navarra, their collaboration with Stanford University's Center for Research in Biomedical Informatics has given rise to a proprietary technology that could revolutionise the discovery of new drugs. This artificial intelligence system, called GENNIUS, uses graph neural networks (GNN) and deep learning to predict drug-drug interactions.[lii].
The model uses nodes, which represent drugs and proteins and their connections, and generates graphical representations with which key biological information can be worked with agility. GENNIUS has made it possible to identify hitherto hidden patterns and discover new drug-protein interactions, a decisive step in the development and repositioning of drugs. The model driven by researchers at the University of Navarra outperforms others by evaluating different databases with unprecedented accuracy and speed, enabling it to make more accurate and generalisable predictions. All of this could translate in the future into a reduction of time and costs in the development of therapies.
Finally, a group of researchers at the Polytechnic University of Madrid (UPM) has just developed a method for drug repurposing that can justify the use of a particular compound in the treatment of a disease. The new algorithm is called XG4REPO (eXplainable Graphs for Repurposing).[liii] and one of its main contributions is its effort for explainability: the results are presented in a comprehensible way, alluding to the biological mechanisms used, and their prognoses can thus be validated by medical experts, who can assess whether the explanation is sufficient and legitimate.
The magazine Scientific Reports published the results of a test in which the researchers sought to demonstrate the effectiveness of the XG4REPO algorithm. In it, they made a series of predictions about the outcomes of the application of three known cancer drugs. They discovered that much of what was anticipated by artificial intelligence was already present in the initial clinical trial phase.
[i] Discovery and Development of Penicillin (No date) American Chemical Society. Accessed: 20/07/2025.
[ii] Pellicer Roig, D. (2025) Antibiotics are failing in Africa and the Centre for Disease Control is calling for the problem not to be ignored., National Geographic Spain. Accessed: 28/07/2025.
[iii] Michael, C.A., Dominey-Howes, D. and Labbate, M. (2014) ‘The antimicrobial resistance crisis: Causes, consequences, and management’, Frontiers in Public Health, 2. doi:10.3389/fpubh.2014.00145.
[iv] Who bacterial priority pathogens list, 2024: Bacterial pathogens of public health importance to guide research, development and strategies to prevent and control antimicrobial resistance (No date) World Health Organization. (Accessed: 20/07/2025).
[v] Brown, E., Wright, G. Antibacterial drug discovery in the resistance era. Nature 529, 336-343 (2016). doi: 10.1038/nature17042
[vi] Parameswaran, P. (No date) Target based screening, NC State University Highthroughput Discovery. (Accessed: 20/07/2025).
[vii] Cameron, A. (No date) Incentivising-development-of-new-antibacterial-treatments- ..., World Health Organization. Accessed: 20/07/2025.
[viii] Cesaro, A., Hoffman, S.C., Das, P. et al. Challenges and applications of artificial intelligence in infectious diseases and antimicrobial resistance. npj Antimicrob Resist 3, 2 (2025). doi: 10.1038/s44259-024-00068-x
[ix] Branda, F. and Scarpa, F. (2024) ‘Implications of artificial intelligence in addressing antimicrobial resistance: Innovations, global challenges, and healthcare's future’, Antibiotics, 13(6), p. 502. doi:10.3390/antibiotics13060502.
[x] Arnold, A., McLellan, S. & Stokes, J.M. How AI can help us beat AMR. npj Antimicrob Resist 3, 18 (2025). doi: 10.1038/s44259-025-00085-4
[xi] Abramson, J., Adler, J., Dunger, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493-500 (2024). doi: 10.1038/s41586-024-07487-w
[xii] https://www.gene.com/stories/ai-and-the-quest-for-new-antibiotics
[xiii] Jonathan M Stokes et al. A Deep Learning Approach to Antibiotic Discovery, Cell, 20 February 2020, doi: 10.1016/j.cell.2020.04.001
[xiv] Charlene Lancaster, “Harnessing the Power of AI to Design Novel Antibiotics”, The Scientist, 27 June 2024.
[xv] Abramson, J. et al. Accurate Structure Prediction of Biomolecular Interactions with AlphaFold 3, Nature, June 2024, doi.org/10.1038/s41586-024-07487-w
[xvi] Kyle Swanson et al. Generative AI for designing and validating easily synthesizable and structurally novel antibiotics, Nature Machine Intelligence, 22 March 2024, doi.org/10.1038/s42256-024-00809-7
[xvii] Rachel Tompa, “Generative AI develops potential new drugs for antibiotic-resistant bacteria”, Stanford Medicine, 28 March 2024, accessed on 16/07/2025.
[xviii] Tom Gerken, “AI cracks superbug problem in two days that took scientists years”, BBC, 20 February 2025, accessed 16/07/2025
[xix] Torres, MD, Cesaro, A. and de la Fuente-Nunez, C. “Non-immune protein peptides target infections through antimicrobial and immunomodulatory properties”. Trends Biotechnology, January 2025, doi.org/10.1016/j.tibtech.2024.09.008.
[xx] Mathias Sundin, “AI discovers nearly 1 million new antibiotics - progress in the fight against antibiotic resistance”, WARP, 9 June 2024.
[xxi] Fangping Wan et al. Deep-learning-enabled antibiotic discovery through molecular de-extinction, Nature Biomedical Engineering, 11 June 2024, doi.org/10.1038/s41551-024-01201-x
[xxii] Changge Guan et al. Computational exploration of global venoms for antimicrobial discovery with Venomics artificial Intelligence, Nature Communications, 12 July 2025, doi.org/10.1038/s41467-025-60051-6
[xxiii] César de la Fuente-Núñez, Henry Skinner, Christina Yen, “How AI can slow the rise of antibiotic-resistant “superbugs””, Bulletin of the Atomic Scientists, 13 May 2025, accessed 17/07/2025
[xxiv] “Antibacterial agents in clinical and preclinical development: overview and analysis”, WHO, 14 June 2023.
[xxv] Thomas, D., and Wessel, C. 2022. “The state of innovation in antibacterial therapeutics”, BIO, 14 February 2022.
[xxvi] Thomas, D., Wessel, C., “The State of Innovation in Antibacterial Therapies”, Biotechnology Innovation Organisation, 14 February 2022.
[xxvii] «Netflix for Antimicrobials: The Antimicrobial Subscription Model», House of Commons Library, UK Parliament, 18 September 2024.
[xxviii] Rex, John, Damiano de Felice, «Italy offers a national attraction incentive», AMR Solutions, 9 January 2025
[xxix] Justin R. Randall, Deep mutational scanning and machine learning for the analysis of antimicrobial-peptide features driving membrane selectivity, Nature Biomedical Engineering, 31 July 2024, doi.org/10.1038/s41551-024-01243-1
[xxx] “DRAIGON: pioneering global initiative using AI to predict antimicrobial resistance from genomic data”, European Vaccine Initiative, 18 December 2023.
[xxxi] Hamid Harandi et al. Artificial intelligence-driven approaches in antibiotic stewardship programs and optimising prescription practices: A systematic review, Artificial Intelligence in Medicine, April 2025, doi.org/10.1016/j.artmed.2025.103089
[xxxii] “COVID-19 : U.S. impact on antimicrobial resistance, special report 2022”, National Center for Emerging and Zoonotic Infectious Diseases (U.S.). Division of Healthcare Quality Promotion. Division of Healthcare Quality Promotion, June 2022, dx.doi.org/10.15620/cdc:117915.
[xxxiii] cdc.gov/sepsis/about/?CDC_AAref_Val=https://www.cdc.gov/sepsis/what-is-sepsis.html
[xxxiv] Flavia Pennisi et al. Artificial intelligence in antimicrobial stewardship: a systematic review and meta-analysis of predictive performance and diagnostic accuracy, Eur J Clin Microbiol Infect Dis, 6 January 2025, doi.org/10.1007/s10096-024-05027-y
[xxxv] Hajar Hakkoum, Ibtissam Abnane, Ali Idri, Interpretability in the medical field: A systematic mapping and review study, Applied Soft Computing, March 2022, doi.org/10.1016/j.asoc.2021.108391
[xxxvi] Emmanouil P Vardas , Maria Marketou , Panos E Vardas, Medicine, healthcare and the AI act: gaps, challenges and future implications, European Heart Journal - Digital Health, 23 April 2025, doi.org/10.1093/ehjdh/ztaf041
[xxxvii] health.ec.europa.eu/ehealth-digital-health-and-care/european-health-data-space-regulation-ehds_en
[xxxviii] Felix Busch et al. Navigating the European Union Artificial Intelligence Act for Healthcare, npj Digital Medicine, 12 August 2024, doi.org/10.1038/s41746-024-01213-6
[xxxix] Siôn Geschwindt, “Europe accelerates AI drug discovery as DeepMind spinoff targets trials this year”, TNW, 21 January 2025, accessed 17/07/2025
[xl] Iryna Bursuk, “Meet 20 AI Startups advancing Drug Discovery in 2025”, StartUs Insights, April 2023.
[xli] Francesco Branda, Fabio Scarpa, Implications of Artificial Intelligence in Addressing Antimicrobial Resistance: Innovations, Global Challenges, and Healthcare's Future, Antibiotics (Basel), 29 May 2024, doi: 10.3390/antibiotics13060502
[xlii] broadinstitute.org/drug-repurposing-hub
[xliii] Gary Liu et al. Deep learning-guided discovery of an antibiotic targeting Acinetobacter baumannii, Nat Chem Biol, 25 May 2023, doi.org/10.1038/s41589-023-01349-8
[xliv] rdkit.org/
[xlv] Maximilian G. Schuh, Joshua Hesse, Stephan A. Sieber, AI-guided Antibiotic Discovery Pipeline from Target Selection to Compound Identification, 21 May 2025, doi.org/10.48550/arXiv.2504.11091
[xlvi] Alexandra Maure, Etienne Robino, Charles Van der Henst, The intracellular life of Acinetobacter baumannii, Cell, December 2023, doi.org/10.1016/j.tim.2023.06.007
[xlvii] “WHO bacterial priority pathogens list, 2024: bacterial pathogens of public health importance, to guide research, development and strategies to prevent and control antimicrobial resistance”, World Health Organization, 2024.
[xlviii] “ARPA-H project to accelerate the discovery of new antibiotics using AI”, DARPA, press release, 26 September 2024.
[xlix] “Review of AI/ML applications in medicines lifecycle”, EMA, 9 July 2025,
[l] Daniele Roberto Giacobbe et al. “Antibiotics and Artificial Intelligence: Clinical Considerations on a Rapidly Evolving Landscape”, Infect Dis Ther, 15 February 2025, doi.org/10.1007/s40121-025-01114-5.
[li] “El panorama de la inteligencia artificial en el sector salud en España”, Asebio, 4 April 2024.
[lii] “A new artificial intelligence method predicts drug interactions more accurately”, University of Navarra, 5 March 2025.
[liii] Ana Jiménez, “Explainable drug repurposing via path based knowledge graph completion”, Sci. Rep., 18 July 2024, doi.org/10.1038/s41598-024-67163-x