When the way drugs are found changes, the timetable of the pharmaceutical industry changes as well. AI is no longer merely an auxiliary tool that organizes papers and analyzes data. It is entering the pipeline that identifies disease targets, designs molecules, and repeatedly improves candidate compounds.

***

Drug development is one of the highest-cost, highest-risk industries humanity has pursued for the longest time. Before a single drug reaches patients, researchers must understand the cause of a disease, identify a therapeutic target, test countless compounds, and pass through animal studies and clinical trials. In this process, most candidates disappear. A substance that initially seemed promising may be eliminated along the way because of toxicity, absorption rate, metabolic stability, insufficient efficacy, or unexpected side effects. It is no exaggeration to say that drug development is an industry that grows not only on one successful medicine, but also on the many failed candidates behind it.

That is why the core question of the pharmaceutical industry has always been the same. How can we find drug candidates faster? How can we identify candidates that are likely to fail earlier? How can we reduce the time and cost wasted in laboratories and clinical stages? AI drug discovery emerged as an answer to these questions. At first, AI was closer to a tool that supported the eyes and hands of researchers. It read vast amounts of academic papers, organized genetic and protein data, and selected substances from existing compound databases that might match disease targets. Its role was to quickly organize complex biological information and detect patterns that humans might miss.

But the recent shift goes one step deeper than that. AI is now moving beyond a tool that simply analyzes data and is entering the very process of producing candidate molecules. It is going beyond the level of choosing promising substances from existing compounds. It proposes new molecular structures with desired properties, predicts how those structures will bind to target proteins, and redesigns them while simultaneously considering efficacy, toxicity, stability, and synthesizability. In other words, part of the exploration and design work that researchers used to perform at the earliest stage of drug development is moving into AI models and automated experimental systems.

This change is not merely a technological fad. It is a signal that the production method of the pharmaceutical industry itself is changing. Drug development has long been a laboratory-centered industry. Experienced researchers interpreted disease mechanisms, imagined candidate compounds, designed experiments, and read the results again. Of course, computational chemistry and bioinformatics have been used for a long time. But recent AI goes beyond the scope of auxiliary computation. As generative models, protein-structure prediction, multimodal biological data analysis, automated synthesis, and robotic laboratories converge, the speed and method of candidate-molecule discovery are changing.

In the past, drug discovery was like searching for useful items in a vast warehouse. Researchers selected substances from existing compound libraries that might bind to a specific disease target. They verified some of them through experiments and gradually modified substances that showed potential in order to improve efficacy. AI played the role of narrowing the search range in this process. Among millions of candidates, it helped identify the substances worth testing first and excluded those with a high probability of failure.

Now the question itself is changing. It is moving from “What substances already exist?” to “What substances should we create?” Generative AI proposes new molecular structures based on target conditions. A candidate must bind strongly to a specific protein, bind less to other proteins, not break down too quickly inside the body, have low toxicity, and be a structure that can actually be synthesized. A drug candidate cannot succeed by satisfying only one condition. Even if efficacy is strong, it fails if toxicity is high. Even if it binds well to the target, it has little meaning if it is not absorbed in the body. Even if stability is good, development feasibility becomes low if synthesis is difficult or the cost is excessively high.

The difficulty of AI-based drug design lies precisely in the need to handle these multiple conditions at the same time. A good drug is not simply a strong molecule, but a well-balanced molecule. Binding affinity to the target protein, selectivity, solubility, cellular permeability, metabolic stability, toxicity risk, and manufacturability must all be considered together. AI models calculate these complex conditions simultaneously and propose candidates. They rapidly explore chemical spaces that are difficult for humans to imagine one by one and present structures with potential.

What matters here is that this does not mean AI replaces researchers. Molecules proposed by AI still need to be verified. Researchers must confirm whether they can actually be synthesized, whether they work in cells and animal models, whether they are toxic, and whether they produce meaningful effects in humans during clinical trials. What AI changes, however, is the starting point. In the past, researchers imagined candidates within the limits of their experience and data. Now, AI first opens a much broader space of possibilities. Researchers then take on the role of judging which of those possibilities are scientifically valid and developmentally feasible.

This change transforms drug development from a linear process into a circular process. In the past, target discovery, candidate exploration, synthesis, experimentation, and optimization proceeded as relatively separate stages. One stage had to be completed before moving on to the next. But when AI and automated experimentation are combined, the process gains a faster circular structure. AI designs candidates, automated laboratories synthesize them, experimental results are fed back into the model, and the model proposes better candidates again. Design and experimentation, failure and revision, are repeated in short cycles.

The reason this closed loop is important is that the essence of drug development is ultimately iterative learning. The first candidate rarely becomes a drug immediately. Usually, efficacy is insufficient, toxicity appears, the substance disappears too quickly in the body, or it works only under certain conditions. What matters is how quickly failure is read and reflected in the next design. The statement that AI is entering the pipeline means precisely that this learning speed is becoming faster. Laboratory results become data, data change the judgment of the model, and the judgment of the model leads the next experiment.

The integration of AI drug discovery into the pipeline is also visible in the changing collaboration patterns between global pharmaceutical companies and AI biotech firms. Large pharmaceutical companies have strengths in disease understanding, clinical development, regulatory response, manufacturing, and sales. AI biotech companies, on the other hand, emphasize their strengths in generative models, protein-structure prediction, large-scale biological data analysis, and automated experimental platforms. The combination of the two is not simply the adoption of a technology, but rather a reorganization of roles.

In the past, when pharmaceutical companies adopted external technologies, they often acquired a specific candidate compound or platform. Now, more companies are forming long-term partnerships with AI firms to jointly discover candidate compounds in specific disease areas. The reason AI’s role is growing in areas such as cancer, immune diseases, inflammatory diseases, and rare diseases is that these fields have high biological complexity and large unmet needs. In areas where it is difficult to identify targets through conventional methods, where candidate optimization takes a long time, or where patient groups are divided into complex subgroups, AI’s analytical and design capabilities may hold greater value.

AI companies are also undergoing change. If they remain merely software vendors, it is difficult for them to enter deeply into the core value of drug development. The value of their technology can be proven only when they actually produce candidate molecules and place them in clinical pipelines. That is why AI drug-development companies are not stopping at promoting model performance, but are building their own candidate-molecule pipelines or pursuing joint development with large pharmaceutical companies. Only when AI-created candidates enter clinical trials and prove efficacy and safety in humans can industrial trust be established.

At this point, the expression “AI creates drugs” must still be handled cautiously. AI can propose and optimize candidate molecules, but the success of a drug is still proven in biology and clinical trials. The human body cannot be explained by a single protein structure. Disease is a complex phenomenon involving genes, immunity, metabolism, environment, lifestyle, and differences among patient groups. A candidate that showed effects in cell experiments may fail in animal studies, and a candidate that passed animal studies may fail to show efficacy in clinical trials. A candidate discovered by AI cannot bypass this process.

Even so, the significance of AI is clear. AI is not a technology that eliminates failure, but a technology that reduces the cost of failure. If a major loss in drug development is recognizing bad candidates too late, AI can help filter them out earlier. It can exclude structures suspected of toxicity at the initial stage, adjust candidates with poor drug-like properties, and guide experiments toward more promising directions. It may not completely eliminate failure, but it can move the point of failure earlier. That alone has great industrial value.

In drug development, time is cost, and cost ultimately leads to the issue of accessibility. When development costs rise too high, drug prices become expensive, and rare diseases or diseases with small market sizes may be pushed down the list of research priorities. If AI can lower the cost of early discovery and candidate optimization, new possibilities may open even in areas that have not been sufficiently studied because they were considered economically unattractive. Of course, this does not automatically guarantee public benefit. But the lower the entry cost of drug development becomes, the more room there is to test a wider range of diseases and targets.

The core asset of AI drug development is not the model alone. What matters even more is data. AI grows by feeding on good data. Protein structures, gene expression, disease targets, compound responses, toxicity, clinical outcomes, and patient-group information must be sufficiently connected for predictions to become precise. If the data are inaccurate or biased, AI may produce candidates that look plausible but are useless. In drug development, plausibility is dangerous. Even if a molecular structure looks elegant and receives a high prediction score, it has no meaning if it does not work in real biology.

That is why future competition will be competition in AI models and, at the same time, competition in data quality. More important than who has more data is who has more reliable data. Even the same experiment can be interpreted differently depending on its conditions. Even under the same disease name, patient groups divide into many branches. It is also necessary to distinguish whether a target protein is the cause of a disease, the result of it, or merely an accompanying phenomenon. No matter how quickly AI calculates, if it misunderstands the biological meaning, the direction of the candidate molecule can go wrong.

For this reason, AI drug development becomes not simply an IT technology, but a life-science infrastructure in which experimentation and computation are combined. Building a good model alone is not enough. Experiments must be well designed, experimental results must be standardized, and failed data must also be accumulated. If only the data from successful candidates remain, AI has difficulty learning the patterns of failure. In drug development, failure data are extremely valuable. Only by knowing why a certain structure showed toxicity, under what conditions an effect disappeared, and which targets did not lead to actual disease improvement can the next candidate be designed more accurately.

The role of automated laboratories is also growing. Even if AI proposes countless candidate structures in a single day, the bottleneck remains if actual experimentation is slow. Conversely, if synthesis, analysis, cell experiments, and result recording are automated, the design speed of AI and the verification speed of experiments can be aligned. Ultimately, the competitiveness of AI drug development comes not from a single model, but from a system in which models, data, robotic experimentation, and researchers’ judgment are connected. How tightly this system can be built becomes a new standard of pharmaceutical competitiveness.

Regulation and trust are also important issues. Medicines are directly connected to human life. Verification standards cannot be lowered simply because AI designed a molecule. Rather, clearer explanations are needed. Companies must be able to explain what data were used for training, by what criteria candidates were selected, how predicted results were confirmed in experiments, and how the possibility of failure was assessed. In drug development, explainability is not merely a technical decoration, but a condition for passing through regulation and trust.

Patent issues surrounding candidate molecules proposed by AI may also become complicated. It will be necessary to examine how different they are from existing compounds, how the inventiveness of structures created by AI should be evaluated, and whether the data used to train the model raise rights-related issues. Drug development is science and, at the same time, an intellectual-property industry. In a field where the value of a single candidate molecule can be assessed in the billions of dollars, the rights and responsibilities associated with molecules created by AI will inevitably become more important.

The integration of AI drug discovery into the pipeline also raises important questions for Korea’s pharmaceutical and biotech industries. Korea has strengths in clinical-trial capabilities, hospital data, biomanufacturing, and certain platform technologies. However, there are only a limited number of companies with the vast in-house pipelines and large-scale R&D budgets of global big pharma. In this environment, AI drug development is both an opportunity and a pressure. If the barriers to early discovery and candidate-molecule design fall, smaller companies can also present possibilities more quickly. Conversely, if they fail to secure sufficient data and platforms, they may fall behind in global competition more quickly.

What Korean companies need is not a simple declaration that they are adopting AI. They need concrete strategies regarding which disease areas they will build strengths in, what data they will secure, which hospitals, research institutes, and pharmaceutical companies they will connect with, and how they will acquire the capability to bring AI-created candidates into actual clinical development. AI drug development is not completed by models alone. The entire process must be connected: synthesizing candidates, verifying them, securing patents, designing clinical trials, and communicating with regulatory agencies.

The choice of disease area is especially important. A strategy of applying AI to every disease lacks strength. Selection and concentration are needed in areas where Korea has strengths, such as cancer, immune diseases, metabolic diseases, rare diseases, and disease groups based on digital hospital data. It is also important to improve the quality of patient data and biological data. In the AI era, biotech competitiveness depends on how well hospitals, laboratories, companies, and data infrastructure are connected.

The global collaboration structure of drug development may also change significantly. In the past, the main route was to create a promising candidate molecule and then license it out to a large pharmaceutical company. Going forward, collaborations that combine AI platforms, disease data, candidate-molecule design capabilities, and clinical-development capabilities may increase. The model may shift from selling a single candidate to building a joint pipeline that continuously produces candidates in a specific disease area. In that case, the value of a company will come not from one candidate molecule, but from a system capable of continuously producing candidates.

AI drug development is also changing the language of the pharmaceutical industry. In the past, the word discovery was more familiar. There was a strong image of accidentally discovering possibilities in nature, in compound libraries, or in experimental results. Now, the language of design and production is becoming more important. A disease target is defined, desired molecular conditions are set, the model creates candidates, experiments verify them, and the model revises them again. Drug candidates are being discovered and, at the same time, produced.

Expectations surrounding AI drug development are high, but there is also no shortage of hype. It is dangerous to speak as though AI will soon solve every problem in drug development. The failure rate of clinical trials remains high, biology is complex, and the human body often betrays predictions. More verified cases are still needed before candidates created by AI can become approved medicines. The market wants quick success stories, but medicine demands safety and reproducibility more than speed.

But the existence of hype does not mean the change itself should be taken lightly. What matters is how deeply AI has entered drug development. AI is no longer an assistant organizing materials on the periphery. It participates in candidate-discovery strategies, engages in molecular design, determines experimental priorities, and influences decisions on whether to stop or continue a pipeline. It is becoming a technology that sets direction at the very front end of drug development.

The meaning of this change is clear. Competition in the pharmaceutical industry is no longer determined only by laboratory scale. Data quality, model accuracy, the speed of automated experimentation, judgment in clinical development, and regulatory-response capability must be bound together into one pipeline. AI becomes part of candidate-molecule production at the point where these elements connect. The gap between companies that use AI merely as a tool and those that integrate it as a pipeline may grow wider over time.

The future of drug development is not a confrontation between human researchers and AI. The more appropriate picture is collaboration. AI expands possibilities that humans have not seen, and humans interpret the meaning of the possibilities proposed by AI. AI quickly creates candidates, and experiments verify those candidates in reality. AI calculates the possibility of failure, and researchers judge what that failure means scientifically. In the end, a medicine is not completed inside a model; it is proven inside the human body.

That is why the true meaning of AI drug discovery is not found in the sentence “AI makes drugs instead of humans.” A more accurate expression is this: AI is becoming part of the pipeline that produces new drug candidate molecules. This is a change that alters the starting point of drug development. It creates more candidates faster, discards worse candidates earlier, and allows experiments to be repeated in more promising directions.

For a long time, drug development has been a slow, expensive industry with many failures. Making medicines will not become easy in the future either. But as AI enters the flow of discovery and design, experimentation and learning, the timetable of drug development is changing. The important question now is no longer whether to use AI or not. It is whether AI will remain a convenient tool outside the laboratory, or whether it will be integrated as a core component of a pipeline that continuously produces candidate molecules. Depending on the answer, the speed and competitiveness of the future pharmaceutical industry will differ greatly.

Reuters, May 2026, Google-backed Isomorphic raises $2.1 billion to scale AI-driven drug discovery

Isomorphic Labs, May 2026, Isomorphic Labs announces Series B investment round

Isomorphic Labs, February 2026, The Isomorphic Labs Drug Design Engine unlocks a new frontier

Iambic Therapeutics, February 2026, Iambic announces collaboration with Takeda to advance AI-driven design of small molecules

Nature Biotechnology, July 2025, Clinical trials gain intelligence

Insilico Medicine, July 2023, First generative AI drug begins Phase II trials with patients

arXiv, April 2025, AI-guided antibiotic discovery pipeline from target selection to compound identification

arXiv, July 2025, Accelerating drug discovery through agentic AI: a multi-agent approach to laboratory automation in the DMTA cycle