From Medical AI to AI Drug Discovery to Medical World Models: The Formation of Next-Generation Biomedical AI Platforms

Over the past decade, artificial intelligence has entered medicine following a clear migration path.

The first phase was Medical AI. Its representative applications include medical imaging interpretation, clinical note structuring, risk prediction, medical question answering, triage, and clinical documentation. The core problem it addresses: how to faster and more reliably understand the medical facts that have already occurred.

The second phase is AI Drug Discovery. Its representative applications include target discovery, molecular generation, protein structure prediction, drug screening, toxicity prediction, and preclinical optimization. The core problem it addresses: how to faster discover potentially effective intervention tools.

But the ultimate problem in medicine goes beyond "understanding disease" and beyond "finding candidate drugs." The hardest problem in medicine is:

> After an intervention enters a real human body, can it safely, interpretably, and verifiably alter the trajectory of disease or health?

This is precisely why medical world models may become the direction of the future. They are not simply scaled-up versions of ordinary prediction models, nor are they chatbots that can write medical texts. Rather, they represent a new form of medical computation: treating the human body as a continuously changing system, treating drugs, nutrition, exercise, surgery, and behavioral changes as interventions, treating follow-up data as feedback, and reasoning about the consequences of different choices within safe boundaries.

The "Plausible Mechanism Framework" proposed by the U.S. FDA in 2026 around individualized therapies raises similar questions from the perspective of regulatory science: when traditional large-scale clinical trials are difficult to conduct, how can medicine establish a credible chain of evidence around pathogenic mechanisms, therapeutic targets, natural disease course, biomarker changes, and clinical outcomes? This does not mean lowering evidentiary standards or bypassing clinical validation; rather, it signals that future medicine increasingly needs a new infrastructure capable of connecting mechanisms, data, interventions, and verification.

The core thesis of this article is:

> Medical AI solves the problem of "understanding medicine," AI drug discovery solves the problem of "discovering tools," and medical world models solve the problem of "verifying intervention outcomes." The first two primarily change efficiency; the latter may change the medical paradigm.

But it must be emphasized: medical world models are still in their early stages. Their outputs can only be regarded as testable hypotheses, research aids, or decision support — not as clinical facts, proof of efficacy, or regulatory endorsements.

1. Why is Medical AI the "Past"?

Calling it the "past" does not mean Medical AI has lost its value. It means that it has already completed the first migration from a novel concept to infrastructure.

Over the past decade, the most mature applications of Medical AI have largely concentrated on recognition, extraction, and workflow assistance:

- Detecting abnormalities in medical images;

- Extracting structured information from clinical notes;

- Predicting the risk of future complications in certain patient groups;

- Assisting physicians with documentation;

- Performing medical search, question answering, triage, and quality control.

These capabilities are important and will continue to generate value. But from an investment narrative perspective, they increasingly resemble the "tool layer" and "efficiency layer" within the healthcare system.

They address the question:

> What has already happened in the medical data?

In other words, the first generation of Medical AI primarily trained machines to become faster, cheaper, and more stable medical recognition systems. It can help doctors see faster, write faster, search faster, and triage more accurately.

But it has not yet truly entered the deepest core problems of medicine.

What is truly difficult in medicine is often not recognizing what has already happened, but judging what will happen after an intervention.

Whether a patient will deteriorate in the future is a prediction problem. But whether switching to a different drug, intervening earlier, adjusting the dose, or changing lifestyle habits would alter the body's trajectory — that is an intervention problem. The former can be estimated based on historical similarities, while the latter must simultaneously contend with disease mechanisms, patient state, therapeutic actions, temporal changes, and real-world feedback.

This is the boundary of Medical AI: it has made artificial intelligence the eyes and assistant of medicine, but it has not yet made artificial intelligence a system for verifying intervention outcomes.

2. Why is AI Drug Discovery the "Present"?

If the first phase of Medical AI was about "understanding medicine," then the second phase of AI drug discovery is about "discovering intervention tools."

This is also one of the areas where capital, pharmaceutical companies, and AI firms are most concentrated today. The reason is not difficult to understand: drug development cycles are long, costs are high, and failure rates are staggering, while AI appears capable of directly addressing the most expensive and uncertain links in the R&D chain.

The problems AI drug discovery is attempting to solve include:

- Identifying new disease targets;

- Generating new candidate molecules;

- Predicting interactions between molecules and proteins or pathways;

- Optimizing drug-like properties;

- Predicting toxicity and side effects;

- Improving preclinical screening efficiency;

- Helping design more rational experimental roadmaps.

All of these tasks are critically important and far from simple. They may significantly alter the cost structure and speed of drug development.

But from the perspective of medical world models, AI drug discovery still primarily addresses an upstream problem:

> Can we find a potentially useful "key" faster?

The problem is that what medicine truly cares about is not the key itself, but:

> After this key is inserted into a real living system, will the door open? Which door will it open? Will it trigger other doors? And will the long-term outcome truly benefit the patient?

This is precisely one of the root causes of the persistently high failure rate in drug development: a target that looks reasonable does not guarantee that real patients will benefit; a biomarker improvement does not guarantee that long-term outcomes will improve; and a molecule that works in models does not guarantee it will remain effective in the human body.

Therefore, AI drug discovery is the present because it is actively reshaping drug discovery and preclinical R&D. But it is not the endgame, because it primarily answers "how to find intervention tools" rather than comprehensively answering "how intervention tools alter human trajectories."

3. Why are World Models the Future?

What medical world models aim to enter is precisely the gap left by both Medical AI and AI drug discovery.

They are designed to answer not a single question, but a set of questions closer to the essence of medicine:

- If this drug is administered now, what will happen to the patient in three months?

- Without treatment, how will the disease naturally progress?

- If the dose is halved, how will efficacy and risk change?

- If drugs, diet, exercise, and sleep are simultaneously modified, will the outcome be better, worse, or indeterminate?

- If a certain biomarker improves, does it truly mean the patient has benefited?

- If the model's judgment is wrong, does the error stem from mechanistic assumptions, state measurements, dosing choices, patient stratification, or the human body's systemic feedback?

These questions are not simply about "predicting the future." They involve a deeper judgment:

> Whether an intervention truly alters the trajectory of how the body operates.

Ordinary prediction models can tell us "how similar people have typically fared in the past," but they cannot reliably answer "whether taking a different approach with this person would change the outcome."

The goal of medical world models is to place state, intervention, time, mechanism, counterfactuals, and feedback into the same closed loop. This involves not only prediction capability but also causal inference and counterfactual comparison — and therefore cannot be accomplished by correlation-based models alone.

It is not about building yet another risk score, nor about generating a plausible-sounding medical note, but about conducting verifiable reasoning about post-intervention bodily changes under medical mechanism constraints.

4. Three Generations of Biomedical AI: From Efficiency Tools to a Medical Operating System

From an investment narrative perspective, biomedical AI can be understood as three generational opportunities.

| Phase | Representative Directions | Core Question | Industry Implications |

|-------|--------------------------|---------------|----------------------|

| Medical AI | Imaging, clinical notes, consultations, triage, documentation, risk prediction | How to understand the medical facts that have already occurred? | Toolification, workflow efficiency, infrastructure |

| AI Drug Discovery | Targets, molecules, screening, toxicity, preclinical optimization | How to find potentially effective intervention tools faster? | Reshaping R&D efficiency and candidate discovery processes |

| Medical World Models | Intervention simulation, trajectory reasoning, counterfactual comparison, evidence chain auditing | After an intervention enters the body, does it truly alter the life trajectory? | A new platform connecting detection, intervention, follow-up, R&D, and regulation |

To summarize in one sentence:

> Medical AI is an efficiency tool, AI drug discovery is an R&D engine, and medical world models may become the operating system for next-generation biomedical AI.

The term "operating system" here is not a marketing metaphor — it refers to the potential to connect multiple previously siloed components: diagnostic data, patient state, intervention plans, follow-up feedback, mechanistic evidence, real-world data, and regulatory auditing.

Medical AI is more like the perception layer. AI drug discovery is more like the tool production layer. Medical world models aspire to become the middleware for intervention decisions and evidence generation.

This is why they may possess platform properties.

5. What is a Medical World Model?

A simple analogy may help.

An ordinary medical model is like a risk map. It tells you that a certain area may be more dangerous, or that a certain person may be at higher risk.

A medical world model is more like a medical simulator. It not only reads the map but also attempts to understand: what might happen next if we change the route, speed, weather conditions, and vehicle condition.

In the context of medicine, a medical world model is a computational system that attempts to simulate changes in human body states. It cares about five things:

1. What state is this person currently in — not just a diagnostic label, but a holistic state encompassing genetics, molecular profiles, immune function, metabolism, organ function, lifestyle, symptoms, and test results.

2. What intervention is about to be applied — interventions can be drugs, dosages, surgery, nutrition, exercise, sleep, behavioral changes, or combinations of multiple measures.

3. How the body might change after the intervention — which biomarkers will change first, which functions will change later, which changes are merely short-term fluctuations, and which may represent genuine benefit.

4. What the alternative path would look like — for example, no treatment, delayed treatment, or an alternative regimen: would the outcomes differ?

5. Whether real-world follow-up can verify the model's judgment — if the model predicts improvement, do subsequent biomarkers and symptoms actually improve? If not, what is the reason?

Therefore, a medical world model is not "a bigger chatbot," nor is it "a more complex risk score." Its core is:

> Verifiable reasoning about post-intervention bodily changes under medical mechanism constraints.

6. Why Does the FDA's Individualized Therapy Framework Matter?

The Plausible Mechanism Framework proposed by the U.S. FDA was originally designed for ultra-rare genetic diseases and highly individualized therapies.

These treatments face a practical challenge: there are too few patients to conduct large-scale randomized controlled trials in the manner of conventional drug approvals. But this does not mean that efficacy can be judged solely on the basis of being "theoretically plausible." Regulation still requires a credible chain of evidence.

This evidence chain roughly includes:

- Is there a clear cause for the disease?

- Does the treatment target this cause or a closely related mechanism?

- Without treatment, how would the disease typically progress?

- Does the treatment actually alter the expected mechanism?

- Are biomarker or symptom changes consistent with the mechanism?

- Are these changes sufficiently credible to warrant continued observation or evaluation?

This has important implications for medical world models.

Because what medical world models aim to do is also not to conjure a future out of thin air, but to organize the problem into an evidence chain:

> Why this intervention? What is it expected to change? How is it verified? How is it corrected if it fails?

This is also its fundamental distinction from ordinary generative models.

In other words, regulatory science is reminding us: future medicine cannot rely solely on what "sounds mechanistically plausible," nor solely on what "appears effective in a model." What truly matters is connecting mechanisms, natural disease course, intervention response, follow-up biomarkers, and clinical outcomes into an auditable evidence chain.

7. Why is This Especially Important for Longevity Medicine?

Chronic diseases and aging do not develop overnight. Blood glucose, inflammation, metabolism, muscle mass, sleep, cognition, and immune function often gradually deviate over many years.

Therefore, what chronic disease management and longevity medicine need most is not a one-time assessment, but long-term trajectory management.

For example, a decrease in a biological age biomarker sounds promising, but it does not automatically equate to improved bodily function, let alone extended lifespan. What truly matters is:

- Are indicators at multiple levels changing in the same direction?

- Is function actually improving?

- Is risk actually decreasing?

- Can these changes be sustained?

- Is there evidence of over-intervention?

The value of medical world models lies in placing these scattered indicators within the same temporal trajectory for interpretation.

This is also what distinguishes longevity medicine from ordinary health management. It cannot merely chase aesthetically pleasing changes in individual biomarkers — it must answer whether an intervention truly makes a person's health trajectory more stable, more resilient, and more sustainable.

Therefore, longevity medicine may be one of the earliest domains where "weak world models" emerge. It does not need to simulate the entire human body from the outset; instead, it can begin within well-defined boundaries by placing methylation, proteomics, metabolomics, functional performance, lifestyle, and follow-up feedback onto the same trajectory for observation.

8. Recent Representative Advances: Medical World Models Are Taking Shape

If "medical world models" were primarily a theoretical imagination a few years ago, a recent wave of research and frameworks suggests that this direction has begun transitioning from concept to early prototype.

But it must be stated upfront: most of these systems remain at the stage of research, preprints, conference papers, or conceptual frameworks. They should not be understood as mature clinical products, much less as direct replacements for physicians. More accurately, they collectively point to a trend: medical AI is moving from "recognition and prediction" toward "simulating and verifying interventions."

8.1 Tumor Treatment Simulation Models

One representative example is a tumor treatment simulation medical world model released in 2025. Using liver cancer interventional therapy as the clinical scenario, it attempts to simulate post-treatment tumor status based on pre-treatment imaging and the treatment plan, combined with survival risk assessment to support comparison across different treatment strategies.

The significance of this direction lies in the fact that it does not merely judge "what the tumor looks like now" but attempts to answer "how the tumor might change if a particular treatment is administered."

Such models still require larger-scale, multi-center, prospective validation, but they already demonstrate a typical form of medical world models: connecting imaging, treatment actions, disease changes, and outcome assessment.

8.2 Clinical Trajectory World Models

Another new direction is patient trajectory models based on electronic health records. A related study published in 2026 attempts to learn patient state changes during hospitalization from large volumes of longitudinal records and simulate longer-term clinical trajectories.

The value of such models lies not in generating a plausible-looking medical record but in learning the temporal relationships among disease, tests, medications, treatments, and outcomes. The question it attempts to answer is: how will patient states change step by step? Which events may signal rising risk? Could the trajectory shift after an intervention?

These models remain some distance from true clinical intervention planning, because EHR data inherently suffer from recording bias, treatment selection bias, and missing data. But they demonstrate that medical world models are not confined to imaging — they are also entering real clinical workflow data.

8.3 Long-Term Risk Foundation Models

Another adjacent line of work involves temporal risk models for multiple chronic conditions. They learn long-term risk changes in multimorbidity populations from large-scale primary care records, aiming to understand the temporal relationships between different disease events.

Strictly speaking, these models are not yet full-fledged medical world models, because they primarily predict risk rather than directly simulating intervention outcomes. But they provide an important foundation: without long-term disease trajectory modeling, it is difficult to develop medical world models that can truly compare intervention pathways.

8.4 Steerable Medical World Model (SteeroMed)

The Steerable Medical World Model (SteeroMed) proposed by DeepoMe emphasizes a different question: the goal of a world model should not be merely prediction — it should be steerable. The underlying SEWO framework asserts that a medical world model must be directionally set by physicians, constrained by mechanisms, calibrated by follow-up, and amenable to auditing and course correction. It should be noted that SteeroMed is presented here as a framework direction and research proposal, and does not represent completed clinical validation or regulatory endorsement.

In other words, a medical world model cannot simply produce a future outcome — it must also be able to answer:

- How is the current state characterized?

- Is the intervention direction clear?

- Are the expected changes consistent with mechanisms?

- If expectations are not met, does the problem lie in state measurement, mechanistic assumptions, intervention execution, or model extrapolation?

- Can the entire reasoning process be audited by physicians, researchers, and regulators?

SteeroMed's core proposition is that a medical world model must possess directional input, process constraints, counterfactual reasoning, quality control, and failure diagnosis capabilities. It is not claiming to have completed a clinical decision machine, but rather proposing a framework for evaluating and building trustworthy medical world models.

This also explains why "steerability" is more important than mere "prediction accuracy." Medicine is not only concerned with what the future will be — medicine cares even more about:

> Whether we can alter the future within safe boundaries, and prove that this alteration is credible.

9. From an Investment Perspective: Why Might This Be the Next-Generation Platform?

Determining whether a direction may become a platform depends not only on whether it can accomplish individual tasks, but on whether it can connect multiple value links and form a positive feedback loop among data, workflows, verification systems, and regulatory trust.

Medical AI connects physicians and data. AI drug discovery connects R&D teams and candidate molecules. If medical world models succeed, they would connect a much longer chain:

- Diagnostics and multi-omics data;

- Patient state characterization;

- Interventions including drugs, nutrition, exercise, sleep, and surgery;

- Follow-up biomarkers and real-world feedback;

- Mechanistic hypotheses and counterfactual reasoning;

- Clinical validation and regulatory auditing;

- Drug development and individualized health management.

This means that the opportunity of medical world models may not manifest as a single-point product, but rather as a new type of infrastructure.

It may correspond to several key themes:

- Workflow entry point: it does not simply provide answers — it embeds into a continuous process of detection, intervention, follow-up, review, and re-decision;

- Data flywheel: each real follow-up and intervention feedback may in turn improve patient state characterization, mechanistic hypotheses, and model calibration;

- Model capability stack: it needs to simultaneously possess capabilities in characterization, prediction, counterfactual reasoning, planning, calibration, uncertainty expression, and failure diagnosis;

- Multi-modal infrastructure: it must connect imaging, clinical records, multi-omics, wearables, experimental data, and real-world data, rather than relying on a single data source;

- Verification closed loop: its value lies not in generating more conclusions, but in transforming conclusions into traceable, falsifiable, reviewable evidence chains;

- Regulatory credibility: in medical scenarios, the moat comes not only from algorithmic performance, but also from clinical validation, compliance systems, auditing capability, and safety boundaries;

- Platform extensibility: once the underlying state characterization and intervention verification framework is established, it can extend to drug development, chronic disease management, longevity medicine, rare diseases, and individualized therapy.

Of course, this platform property remains an early-stage judgment. Whether it materializes depends on several critical questions:

- Whether sufficiently high-quality, longitudinal, multi-layer human data can be obtained;

- Whether biases, missing data, and confounders in real medical data can be handled;

- Whether mechanistic knowledge, causal inference, and statistical learning can be integrated;

- Whether uncertainty can be clearly expressed rather than manufacturing false certainty;

- Whether clinical, ethical, and regulatory auditing can be accommodated;

- Whether continuous correction through real-world feedback is possible.

If these problems are progressively resolved, medical world models may evolve from a research concept into the next-generation biomedical AI platform. But if high-quality data, real-world validation, and regulatory credibility are lacking, it may remain a compelling narrative without becoming a sustainable industry opportunity.

10. How Does It Differ from Systems Biology, Virtual Cells, and Molecular Models?

"World model" may sound like a new term, but it did not emerge from nowhere. It exists on the same scientific continuum as decades of systems biology, biological modeling, digital twins, virtual cells, and the structural biology AI represented by DeepMind: using computation to understand living systems.

But they operate at different levels and serve different ultimate tasks.

Systems biology and traditional biological modeling are primarily answering: what networks exist within living systems? How do genes, proteins, metabolites, and signaling pathways interact? Their strengths lie in mechanistic clarity and interpretability, but they typically depend on extensive manually specified equations, parameters, and prior structures, and model boundaries are often confined to a single pathway, cellular process, or disease mechanism.

Virtual cells and whole-cell models push the objective further: they aim to build executable computational simulations at the cellular level, integrating gene expression, metabolism, signal transduction, and cell growth. Recent AI-driven virtual cell initiatives are also attempting to use large-scale data and models to simulate cellular state changes.

The AlphaFold series and subsequent molecular models advanced by DeepMind, Isomorphic Labs, and others resolved another critical bottleneck: they enabled the scientific community to predict protein structures and molecular interactions at unprecedented speed. AlphaFold 2 achieved a breakthrough in protein structure prediction, and AlphaFold 3 further extended this to structure and interaction prediction across proteins, nucleic acids, and small molecules. These are of enormous significance for fundamental biology and drug development, but their core remains oriented toward the molecular structure and molecular interaction level, rather than patient-level long-term intervention trajectory simulation.

The relationship between medical world models and these directions can be understood as follows:

- Systems biology is more like mapping the mechanistic networks of living systems;

- Traditional biological modeling is more like building computable equations around a specific local mechanism;

- Virtual cells are more like constructing an executable simulator at the cellular level;

- Molecular models like AlphaFold are more like providing high-precision maps at the molecular structure and interaction level;

- Medical world models are more concerned with patient-level states, interventions, temporal trajectories, and verifiable outcomes.

Therefore, medical world models are not intended to replace systems biology, virtual cells, or molecular AI. Rather, they aim to connect the mechanistic knowledge, molecular structures, cellular responses, and clinical data produced by these fields to a problem closer to medical decision-making:

> If a particular intervention is applied to this person, will the body's trajectory change? Is this change consistent with mechanisms? Can it be verified by follow-up? If it fails, where does the failure lie?

Of course, this also means that the bar is higher. Model reasoning is not equivalent to fact; mechanistic plausibility is not equivalent to therapeutic efficacy; biomarker improvement is not equivalent to health improvement. Medical world models can only transition from research concepts to trustworthy tools through continuous calibration within real-world follow-up, experimental validation, clinical review, and regulatory frameworks.

11. Conclusion: AI Drug Discovery Finds the Key; World Models Verify Whether the Key Can Open the Living System

To summarize the thesis of this article in one sentence:

> Medical AI is the past, AI drug discovery is the present, medical world models are the future.

But this statement should not be misinterpreted as dismissing the first two.

Medical AI makes the healthcare system more efficient. AI drug discovery makes drug discovery faster. Both are important, and both will continue to evolve.

The real question is: will the next leap in medical AI move from "recognizing medical facts" and "discovering intervention tools" toward "reasoning about and verifying intervention outcomes"?

This is precisely the core value of medical world models.

It enables AI to no longer merely "see disease" or "discover drugs," but to begin helping medicine understand and verify:

> How to safely, interpretably, and verifiably alter the trajectory of disease and health.

Thus, AI drug discovery is more like searching for keys; medical world models aim to answer: after this key is inserted into a real living system, will the door open, which door will open, will other doors be affected, and will the long-term outcome truly benefit the patient.

When medicine shifts from "predicting who will get sick" to "verifying how to safely alter health trajectories," world models will become not merely a technical concept but potentially one of the foundational methodologies of next-generation medicine.

References

1. U.S. Food and Drug Administration. Considerations for the Use of the Plausible Mechanism Framework to Develop Individualized Therapies that Target Specific Genetic Conditions with Known Biological Cause. Draft Guidance, February 2026.

2. U.S. Food and Drug Administration. FDA Launches Framework for Accelerating Development of Individualized Therapies for Ultra-Rare Diseases. FDA News Release, February 23, 2026.

3. Prasad V, Makary MA. FDA's New Plausible Mechanism Pathway. New England Journal of Medicine. 2025;393(23):2365–2367. doi:10.1056/NEJMsb2512695.

4. Qazi MA, Nadeem M, Yaqub M. Beyond Generative AI: World Models for Clinical Prediction, Counterfactuals, and Planning. arXiv:2511.16333, 2025.

5. Yang Y, Wang ZY, Liu Q, et al. Medical World Model: Generative Simulation of Tumor Evolution for Treatment Planning. arXiv:2506.02327, 2025.

6. Mu L, Huang X, et al. EHRWorld: A Patient-Centric Medical World Model for Long-Horizon Clinical Trajectories. arXiv:2602.03569, 2026.

7. Gallier S, Azcoaga-Lorenzo A, McCowan C, et al. SurvivEHR: a competing risks, time-to-event foundation model for multiple long-term conditions from primary care electronic health records. npj Digital Medicine. 2026.

8. Augustine EF, Yu TW, Finkel RS. N-of-1 Studies in an Era of Precision Medicine. JAMA. 2024;332(16):1386–1387. doi:10.1001/jama.2024.14637.

9. National Academy of Medicine. Expanding the Role of N-of-1 Trials in the Precision Medicine Era: Action Priorities and Practical Considerations. NAM Perspectives, 2024.

10. Schölkopf B, Locatello F, Bauer S, et al. Toward Causal Representation Learning. Proceedings of the IEEE. 2021;109(5):612–634.

11. Pearl J. Causality: Models, Reasoning, and Inference. Cambridge University Press.

12. Hernán MA, Robins JM. Causal Inference: What If. Chapman & Hall/CRC.

13. LeCun Y. A Path Towards Autonomous Machine Intelligence. OpenReview, 2022.

14. U.S. Food and Drug Administration. Artificial Intelligence and Machine Learning in Software as a Medical Device and related regulatory science resources.

15. SteeraMed: A Steerable Biomedical World Model. https://steerable.world

16. Xiong J, et al. World Models for Biomedicine: A Steerability Framework. Preprints.org, 2026. doi:10.20944/preprints202605.0366.v1.

17. Stanford University School of Engineering. "Virtual cell" would bring benefits of computer simulation to biology. 2010.

18. Tang L. The virtual cell. Nature Methods. 2025. doi:10.1038/s41592-025-02951-5.

19. Jumper J, Evans R, Pritzel A, et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596:583–589. doi:10.1038/s41586-021-03819-2.

20. Abramson J, Adler J, Dunger J, et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature. 2024;630:493–500. doi:10.1038/s41586-024-07487-w.

21. Google DeepMind. AlphaFold: Five years of impact. 2025.