AI-driven discovery and molecular design

• Target identification: finding a biological molecule (usually a protein) that, if you nudge it, treats a disease. Example: a kinase enzyme overactive in a cancer.
• Lead optimization: designing a small molecule (a "lead compound") that binds that target well, without harming healthy tissue, and can actually be manufactured.

Historically, these two steps alone could take three to five years and burn through tens of thousands of synthesized compounds. AI is attacking both.

Where AlphaFold changes the game

To design a drug that fits a protein, you need the protein's shape. AlphaFold gives you a credible starting structure in hours instead of years.

Concrete impact:

• Undruggable targets get a look. Many disease-relevant proteins were never crystallized, so drug hunters ignored them. Predicted structures put thousands back on the table.
• Structure-based design starts on day one. Chemists can begin docking candidate molecules into a predicted binding pocket immediately.

AlphaFold3, released in 2024, went further: it predicts how proteins interact with other molecules, including DNA, RNA, and small-molecule drugs (called ligands). That moves the tool from "here is the protein" toward "here is the protein with your drug candidate sitting in it."

The catch: a prediction is not an experiment

A predicted structure is a hypothesis. Two failure modes matter for professionals evaluating claims:

1. Static snapshots. Proteins move. AlphaFold typically gives one dominant conformation, but drug binding often depends on rare "open" states or flexible loops the model may miss.

2. Confidence is uneven. AlphaFold reports a per-residue confidence score (pLDDT). Low-confidence regions, often the floppy parts, are exactly where binding sometimes happens. Trusting a low-confidence pocket is a common rookie error.

Bottom line: AlphaFold accelerates the *start* of structure-based design. It does not replace the wet lab that confirms the structure and the binding.

Generative chemistry: designing the molecule itself

Once you have a target pocket, you need molecules to fill it. Traditional approach: screen giant physical or virtual libraries of existing compounds. Generative approach: have a model *invent* new molecules aimed at your pocket.

These models learn the "grammar" of drug-like chemistry from millions of known molecules, then generate novel candidates conditioned on goals: bind this target, stay soluble, avoid toxicity flags.

A common way to represent molecules to a model is SMILES, a text string that encodes chemical structure. For example, aspirin looks like this:

A generative model can output thousands of such strings, each a proposed molecule, which chemists then filter and rank.

What this compresses

• Idea generation goes from weeks to minutes. A model can propose tens of thousands of candidate structures overnight.
• Multi-objective optimization. Models can be steered toward several properties at once (potency, solubility, metabolic stability), the kind of balancing act that used to take many manual design cycles.

Companies including Insilico Medicine and Recursion have publicly reported AI-originated candidates advancing into clinical trials, which is a genuine milestone. Note the honest caveat: reaching a trial proves the pipelinepipelineAll active sales opportunities across the stages of the sales process, together with their combined potential value and probability of closing.Voir la définition complète → can produce a testable molecule, not that AI shortens the full path to an approved drug. Clinical trials remain long, expensive, and mostly outside AI's control.

The two walls: synthesizability and binding affinity

Here is where the hype meets the bench. Generative models routinely fail on two fronts.

Wall 1: Synthesizability

A model can draw a beautiful molecule that no chemist can actually make, or can only make in 30 costly steps. The model optimized for a shape, not for a route to build it.

The field measures this with synthetic accessibility (SA) scores and, increasingly, with retrosynthesis models that try to plan a step-by-step recipe from purchasable starting materials. Tools like IBM RXN for Chemistry predict reactions and routes and are free to try.

But retrosynthesis prediction is itself imperfect. A proposed route can fail because a reaction that works on paper does not work on your specific molecule, or a reagent is unavailable. Real teams now bake synthesizability constraints directly into generation, penalizing molecules the model cannot plausibly build. This helps but does not fully solve the problem.

Wall 2: Binding affinity

Binding affinity is how tightly a molecule sticks to its target, often reported as a value like IC50 or Kd (lower means tighter binding). Predicting it accurately from structure alone is still hard.

Why it stays hard:

• Water and entropy. Binding happens in a wet, wobbling environment. Displacing water molecules and accounting for flexibility involves physics that fast AI scoring functions approximate crudely.
• Small errors, big consequences. An error of a couple of kilocalories per mole in predicted binding energy can be the difference between a promising drug and a useless one.
• Data bias. Models learn from published binding data, which skews toward compounds that worked. They see fewer clean examples of near-misses, so they overpredict success.

The practical result: AI proposes and ranks, but you still synthesize the top candidates and measure binding in an assay (a lab test). The AI narrows a haystack; it does not hand you the needle pre-verified.

How a modern discovery loop actually runs

The mature 2026 workflow is not "AI replaces scientists." It is a tight loop:

1. Predict the target structure (AlphaFold or an experimental structure).

2. Generate candidate molecules with a generative model, constrained for synthesizability.

3. Score candidates with docking and affinity predictors; rank them.

4. Synthesize a small, smart subset in the lab.

5. Assay them for real binding and early toxicity.

6. Feed results back to retrain or fine-tune the models.

This "design, make, test, analyze" cycle existed before AI. What AI changes is the *size and speed* of the design step and the quality of prioritization, so chemists synthesize fewer, better candidates.

A useful mental model for non-technical leaders: AI is a very fast, very well-read intern who proposes thousands of ideas and ranks them plausibly, but has never touched a beaker. You still need experienced chemists and a wet lab to decide what is real.

What this means for the business

• The bottleneck moves. When idea generation gets cheap, the constraint shifts to lab throughput and assay quality. Firms investing in automated, high-throughput labs (sometimes called "self-driving labs") pair well with generative models.
• Data becomes the moat. Public models are widely available. The advantage comes from proprietary assay data used to fine-tune and validate predictions on your specific targets.
• Timelines compress at the front, not the back. AI can shrink target-to-lead work from years to months. It does little for the clinical trials that dominate cost and time. Be skeptical of any claim that AI slashes total time-to-market.

Key Takeaways

• AlphaFold turned protein structure from a multi-year experiment into a fast prediction, reopening previously "undruggable" targets, but predictions are static hypotheses that need lab confirmation, especially in low-confidence and flexible regions.
• Generative chemistry models invent novel candidate molecules in minutes and optimize for several properties at once, compressing the idea stage of discovery dramatically.
• The two persistent walls are synthesizability (can we actually make it?) and binding affinity (will it stick tightly enough?), both still requiring wet-lab synthesis and assays to verify.
• The real workflow is a fast design-make-test-analyze loop, where AI narrows the haystack and scientists find the needle; proprietary assay data, not the models themselves, is the durable advantage.
• AI compresses the early stages, not clinical trials. Reaching a trial with an AI-designed molecule is a milestone, not proof of a shorter path to an approved drug.