Artificial intelligence can now propose molecular structures, rank targets, predict binding, estimate physicochemical properties and generate candidate compounds at a speed that would have seemed unrealistic only a few years ago. The challenge is no longer generating possibilities. It is deciding which of those possibilities deserve chemistry, assay time and development resources.
That distinction is becoming increasingly important as generative AI moves deeper into pharmaceutical discovery. Recent 2026 literature is becoming noticeably more critical of workflows that judge AI-generated molecules mainly by novelty, docking scores or predicted affinity. A September 2026 paper in Drug Discovery Today argues that generative molecular design should be evaluated against something much more practical: whether the chemistry is actually executable and whether the proposed compounds can be synthesized and tested. (sciencedirect.com)
For medicinal chemistry and discovery teams, this is an important shift. AI is becoming very good at producing options, but the real value lies in narrowing those options to molecules that are chemically feasible, biologically plausible and worth taking into the next experimental cycle.
Generating a Molecule Is Easier Than Generating a Useful Molecule
A generative model can explore chemical space far more broadly than a medicinal chemist could manually enumerate. That makes it extremely useful for proposing new scaffolds, analogues and structural combinations, especially when the model is constrained by known target information or an existing chemical series.
The difficulty is that chemical novelty alone has little value if the resulting molecule creates problems elsewhere. A generated compound may show a strong predicted interaction with a target while being synthetically unrealistic. Another may look drug-like but contain structural motifs associated with metabolic instability. Improving potency may increase lipophilicity, reduce solubility or introduce new toxicity concerns.
This is why current drug-design research is increasingly treating molecular generation as a multi-objective optimization problem rather than a search for a single maximum score. A September 2026 Nature Communications perspective emphasizes that drug design requires balancing several competing objectives simultaneously, including biological activity, pharmacokinetic behavior and safety. (nature.com)
That reflects the reality medicinal chemists already deal with. The most attractive compound is rarely the one with the highest predicted affinity. It is usually the molecule that offers the best workable balance of potency, selectivity, developability, synthetic feasibility and safety.
Synthetic Accessibility Is Becoming a Bigger Part of AI-Based Design
One of the most practical criticisms of generative AI is that models can propose structures without fully accounting for the difficulty of making them.
A molecule may look attractive computationally while requiring an impractical synthetic route, unstable intermediates, unusual reagents or chemistry that is difficult to reproduce at scale. This becomes especially problematic when a model is rewarded mainly for novelty or target-related scores.
The September Drug Discovery Today framework specifically argues that AI-generated molecules should be assessed through “synthetic reality,” connecting molecular generation more directly with laboratory execution. Proxy metrics alone cannot establish whether a proposed structure is genuinely actionable. (sciencedirect.com)
For pharmaceutical teams, this means retrosynthetic feasibility, route complexity, reagent availability and chemical robustness should influence molecular prioritization earlier in the process. The useful output from an AI system is not a long list of theoretically interesting structures. It is a much smaller group of molecules that can realistically enter a design-make-test-analyze cycle.
ADMET Should Influence Design Earlier
Absorption, distribution, metabolism, excretion and toxicity, or ADMET, has traditionally been treated as a later-stage filter in many discovery workflows. AI makes these predictions much faster, which creates an opportunity to consider developability earlier rather than waiting until promising compounds have already been prioritized.
The challenge is that prediction speed should not be confused with prediction certainty. Recent 2026 work on AI-powered ADMET modelling highlights its value for early-stage compound assessment while also noting continuing limitations around inconsistent datasets, model transparency and the need for experimental confirmation. (sciencedirect.com)
In practice, one model may flag a metabolic liability while another does not. A compound may show acceptable predicted permeability but poor solubility. Improving one molecular property may worsen clearance, plasma-protein binding or off-target risk.
This is why ADMET becomes more useful when treated as part of molecular design rather than as a final pass-or-fail screen. If developability constraints are introduced earlier, generative chemistry can search for structures that balance potency with solubility, permeability, metabolic stability and safety instead of producing large numbers of compounds that are filtered out later.
Applicability Domain Still Matters
One of the easiest mistakes in AI-assisted discovery is assuming that a precise-looking prediction is automatically reliable.
Machine-learning models generally perform best when the molecule being evaluated sits within or reasonably close to the chemical and biological space represented in the training data. When a generated structure moves far outside that space, the model may still return a numerical prediction, but confidence in that prediction can be much weaker.
This is particularly important for generative AI because novelty is often treated as desirable. The model may deliberately explore unfamiliar chemical space, yet the further it moves from known chemistry, the more carefully property predictions should be interpreted.
An August 2026 Nature Reviews Drug Discovery perspective makes this broader point. Despite major progress in AI methods, the authors argue that clinically meaningful impact remains more limited than the overall volume of AI activity might suggest, partly because biological data are conditional and real discovery problems are often underspecified. (nature.com)
For discovery teams, the implication is straightforward: prediction quality cannot be separated from understanding where the molecule sits relative to the model’s reliable domain.
Docking Scores Should Support Decisions, Not Replace Them
AI-assisted docking and structure-based modelling can help prioritize very large compound sets, but a high score should not automatically become a synthesis decision.
Protein flexibility, protonation state, water networks, ligand conformations and scoring-function limitations can all influence computational ranking. Even sophisticated models can disagree, particularly when the binding site is flexible or the proposed ligand introduces chemistry that is poorly represented in the model.
The more useful question is whether the predicted binding mode is chemically plausible, consistent with known structure-activity relationships and supported by other evidence. This is where medicinal chemistry judgment remains essential.
Recent work on AI-assisted fragment-to-lead optimization reflects this balance. AI and in silico methods can accelerate fragment prioritization and analogue design, but structural and biophysical evidence remains central to reliable progression. (sciencedirect.com)
AI can rank thousands of molecules quickly. The difficult part is deciding which rankings deserve to influence real chemistry.
Lead Optimization May Be Where AI Delivers More Practical Value
Generative AI attracts attention because it can create new molecules, but lead optimization may ultimately prove to be one of its more commercially useful applications.
Once a viable chemical series exists, the problem becomes more constrained. Teams already understand parts of the structure-activity relationship, know which regions of the molecule can be modified and have some visibility into emerging ADMET liabilities. Synthetic chemistry is also usually better defined.
AI can then help explore analogues and trade-offs within a more meaningful chemical context. The task is no longer to search unrestricted chemical space but to navigate a narrower series while balancing potency, selectivity, permeability, solubility, metabolic stability and synthetic practicality.
This is exactly the kind of problem where multi-objective optimization becomes useful. Improving potency may worsen permeability. Reducing lipophilicity may weaken binding. Removing a metabolic soft spot may introduce new synthetic difficulty. Increasing solubility may affect selectivity.
The value of AI is therefore not in identifying one mathematically highest score. It is in helping chemists examine those trade-offs more systematically and decide which compromises are acceptable.
AI Still Needs Experimental Reality
The strongest emerging message in 2026 is that drug-discovery AI becomes more valuable when prediction is connected tightly to experimental feedback.
A generated molecule still needs to be synthesized. A predicted binding mode still needs biological evidence. An ADMET prediction still needs appropriate experimental confirmation, and a lead-optimization recommendation still has to survive the next design-make-test-analyze cycle.
This is why the discussion around AI in pharmaceutical discovery is gradually shifting from asking what the model can generate to asking what evidence should make a team trust the recommendation.
That is a much more demanding standard, but it is also where AI begins to become useful in real discovery programmes. Prediction without experimental feedback can produce increasingly sophisticated computational stories without necessarily improving the quality of development decisions.
The Competitive Advantage Will Come From Better Prioritization
Pharmaceutical R&D teams now have access to more computational predictions than ever before. That does not automatically mean better discovery.
The advantage will increasingly come from knowing which model is appropriate, where its reliability limits lie, how to interpret conflicting predictions, how synthetic feasibility should influence prioritization and which experiment should be run next.
This is the central challenge explored in the OnlyTRAININGS advanced session Advanced AI for Pharmaceutical Discovery: Molecular Design, ADMET & Lead Optimization.
The training covers target and compound prioritization, predictive modelling, generative molecular design, ADMET interpretation, lead optimization and prediction-to-experiment workflows, with emphasis on scientific interpretation rather than simply producing AI outputs. It also examines practical AI and computational tools used across discovery so participants can judge them according to the problem being solved rather than treating every platform as interchangeable. (onlytrainings.com)
As AI makes molecule generation faster and cheaper, the more valuable capability will be deciding which proposed molecule deserves the next experiment, which prediction deserves confidence and which apparently attractive result should be challenged before more discovery resources are committed.
Explore Advanced AI for Pharmaceutical Discovery: Molecular Design, ADMET & Lead Optimization:
https://www.onlytrainings.com/course/ai-drug-discovery-molecular-design-admet-training/
AI for drug discovery and development, AI molecular design for drug discovery, generative AI for molecular design, AI-driven drug discovery workflow, AI for ADMET prediction, artificial intelligence in medicinal chemistry, AI for lead optimization, generative molecular design in pharmaceuticals, AI-generated molecules in drug discovery, how AI is used in drug discovery, AI drug discovery molecular design ADMET, AI tools for medicinal chemistry, machine learning for ADMET prediction, AI for pharmaceutical discovery, AI-assisted lead optimization, generative AI in medicinal chemistry, synthetic accessibility of AI-generated molecules, AI drug discovery experimental validation, AI molecular generation and ADMET, AI for compound prioritization, AI for structure-based drug design, AI-assisted molecular docking, AI for hit-to-lead optimization, AI for drug candidate selection, AI drug discovery training
