AI-powered Drug Design: How AIDDISON™ Delivers Makeable Molecules

When a pharmaceutical company sends around 10,000 molecules into preclinical development, statistically only one will make it to patients. That low success rate is not driven by biology alone. Many candidates fail late because they turn out to be difficult or uneconomical to manufacture, or because safety and toxicity issues only emerge after years of work.

That raises a simple but uncomfortable question: how many of those failures were “baked in” at the very moment the molecule was designed?

In recent years, generative AI in healthcare has radically expanded what is possible in early discovery. Platforms like AIDDISON™  exemplify how generative AI can propose thousands of molecules in minutes and optimize them toward specific targets. But this power brings a risk: if design focuses mainly on in-silico activity and ignores how a compound behaves in the body – or whether it can be made at scale – we risk generating elegant dead ends, just faster.

A new generation of tools powered by AI brings three worlds together from the start: real-world chemistry, safety and biology.

Deep learning models, learn from large collections of known drug molecules. They can then generate entirely new structures, tune them toward a protein target and explore parts of chemical space that was out of reach.

In practice, a team might start with the 3D structure of an oncology target or a known reference compound. A generative model then proposes a large library – ranging from tens of thousands to a million molecules – and structure‑based methods such as docking, pharmacophore searches or shape comparisons help estimate which candidates are likely to bind.

At this stage, the output often excites scientists as they sort through novel, structurally diverse and promising molecules on screen. Many molecules are novel, structurally diverse and look promising on screen. But without further filters, that is only half the story.

Ultimately, a drug has to do more than bind its target. It needs to reach therapeutic concentrations in the right tissues, move through absorption, distribution, metabolism and excretion   (ADMET) in a predictable way, and avoid critical off‑targets.

Traditionally, many of these questions were answered stepwise, with in vitro and in vivo studies following initial hits. Today, machine‑learning models trained on decades of experimental ADMET data can flag many potential liabilities much earlier. Well‑calibrated models can, for example, estimate likely oral exposure and bioavailability, metabolic stability and clearance, and indicators of cardiotoxicity or hepatotoxicity. They can also highlight interactions with key transporters or enzymes.

This does not replace experiments, but it changes the balance. Instead of discovering “unmakeable” or unsafe molecules after months or years, teams can deselect high‑risk candidates during virtual screening and focus their resources on those with a better chance of meeting patients. 

Even if a molecule looks attractive from a biological and ADMET perspective, it still has to pass one more test: chemists must be able to make it – robustly, at scale and at a reasonable cost.

Synthetic feasibility depends on the availability and price of starting materials, the number and type of reaction steps, the robustness and scalability of those reactions and, increasingly, regulatory and sustainability considerations such as avoiding particular reagents or solvents. Historically, medicinal chemists relied heavily on intuition and experience to judge whether a structure “felt” synthesizable. 

When AI starts to generate hundreds of thousands of options, gut feeling alone does not scale.

Here, the integration of AI‑driven design with computer‑aided synthesis planning becomes critical. Retrosynthesis engines – such as our SYNTHIA™ Retrosynthesis Software – can propose multi‑step routes, suggest alternative disconnections, and highlight which building blocks are commercially available. Scientists can easily apply this retrosynthetic analysis to AI‑generated molecules, filtering out chemically unrealistic structures before heading to the lab.   Project experience shows that a large part of otherwise “barely makeable” molecules can be eliminated at this virtual stage, allowing chemists to concentrate on candidates that are both interesting and realistically accessible.

The real opportunity lies less in any single model and more in how they are orchestrated. In a modern workflow, generative design proposes structurally diverse, target‑focused molecules. Structure‑based and ligand‑based scoring then estimates their binding to the protein or similarity to known actives. ADMET prediction adds a view on safety and pharmacokinetics. Retrosynthesis and feasibility scoring finally test whether synthetic routes are realistic and scalable.

Platforms such as AIDDISON™ – our cloud environment for AI‑driven molecular design – aim to connect these pieces. In such systems, promising virtual hits are not only scored for activity, but are also passed directly to tools like SYNTHIA™ for automatic route planning and synthesizability assessment.

For researchers, this changes what they see on screen. When they browse through a gallery of candidates, they no longer see just an activity score. They see predicted potency, key ADMET and liability indicators, and at least one viable synthetic route with a feasibility rating.

Instead of optimizing first for activity, then for safety, and only at the very end for manufacturability, all three dimensions are considered in parallel from the outset.

Early users across biotech, industry collaborations and academia report similar patterns when they work this way. Hit‑finding and early optimization cycles can compress from months to weeks because design–make–test loops increasingly begin on the screen. Even if the absolute number of synthesized molecules falls, the share of candidates that meet biological, safety and feasibility criteria tends to rise. Smaller organizations and academic labs are also able to explore sophisticated design spaces that were previously the domain of large R&D organizations with extensive infrastructure.

Generative AI will not replace medicinal chemists, pharmacologists or process chemists. But by weaving together machine learning, ADMET prediction and synthesis planning, it can change where and how their expertise is applied – shifting effort away from late‑stage rescue of problematic molecules toward better decisions at the very beginning.

In that sense, the most powerful promise of AI in drug discovery may not simply be “more ideas faster,” but “better ideas that we can actually turn into medicines.”