Most molecule design systems behave as if chemistry starts from a blank page.

You give them a target, a set of actives, maybe some assay data, maybe a synthetic accessibility score, and they propose molecules as though the project exists in a vacuum.

Real medicinal chemistry projects do not exist in a vacuum. They operate inside a live synthetic context. Routes the team already trusts. Intermediates on the shelf. Reactions that work on this scaffold. Building blocks that can arrive tomorrow. Protecting group strategies already debugged. Failed disconnections everyone now avoids. A CRO that is fast at one chemistry class and terrible at another. An analogue series expandable in a week if you stay within the existing route, and a “slightly better” idea that needs a month of route invention before the first compound exists.

That synthetic context is not peripheral metadata. It is part of the design problem.

A design algorithm that ignores it is blind to one of the most important facts about the project: which ideas can actually be turned into molecules quickly enough to matter.

The wrong abstraction

A lot of modelling work frames synthesis as a property of the molecule. Is this compound synthetically accessible? Can retrosynthesis find a route? Is the SA score low enough?

That framing is too static. It treats synthesis as an intrinsic, context-free attribute, like molecular weight or cLogP. But the practical question in a live project is not “can a chemist in principle make this molecule?” It is: how practical is this molecule for this team, on this project, right now?

A compound might be perfectly “accessible” in theory and still be a terrible suggestion if it requires a bespoke six-step route from scratch while the team has a validated intermediate enabling twenty close analogues in two steps. Another compound might look awkward on paper and yet be the obvious next move because it falls straight out of the route already running.

Synthetic accessibility asks whether the molecule can exist. Synthetic practicality asks what it costs the project to make it exist. That is the quantity the design system should care about.

The joint problem

Good medicinal chemists do not rank molecules by predicted potency and then ask someone else to worry about the route. They weigh potency against turnaround time, SAR value against route risk, novelty against tractability, exploration against throughput. A beautiful idea against a route that will stall the programme for two weeks.

A compound is not just a structure. It is a proposal for work. Can it be made from an existing advanced intermediate? Does it preserve the same late-stage diversification handle? Is the chemistry already validated on closely related analogues? Will the route branch cleanly into a series, or is it a one-off? Is the required building block commercially available, or is it itself a mini-project?

This is one of the quiet skills of strong med chem: choosing ideas that are scientifically informative and operationally plausible. Med chem teams often look “conservative” to outsiders. They are not avoiding creativity. They are managing an experimental economy.

A project only learns from molecules that get made. And molecules only get made if somebody can put matter through glassware on a realistic timescale.

Bits do not become molecules until they hit the bench

There is a recurring habit in AI for chemistry to act as though the informational object is the molecule and synthesis is a downstream inconvenience. It is the other way around. The wet lab bench is where bits become molecules that shape the world. Until then, a design is just a typed suggestion.

Route awareness is not an implementation detail to bolt on later. It is the core interface between intelligence and reality. An algorithm that proposes compounds disconnected from the project’s existing synthetic state is producing chemically legible text. An algorithm that knows the routes, the intermediates, the diversification handles, and the likely next experiments is participating in the actual project.

What the model needs to see

A design system that behaves like a project scientist needs more than structures and labels as inputs. It needs the routes the team has already executed or trusts — not just whether a retrosynthesis engine can propose one. It needs to know what advanced intermediates exist, in what quantity, with what stability, how easily they branch into further analogues. A Suzuki that works on this scaffold is different from a Suzuki that works in a textbook, so validated reaction templates matter. Building block lead times matter: a molecule accessible from tomorrow’s order is not equivalent to one requiring custom synthesis. And the model should know the project’s synthetic scars — the reactions that repeatedly failed, the protecting group choices that caused trouble, the substitution patterns that killed the route.

Frame the problem this way and the output changes too. Instead of a ranked list of structures, the system should be able to say something like: “This analogue is modestly less exciting on paper, but it can be made from Intermediate 14 in two days using the same late-stage coupling already proven on the last series.” Or: “This idea may only be worth it if the potency upside is very large, because it requires route invention and breaks the current analogue-making cadence.”

Low-hanging fruit is not an embarrassment

There is a cultural bias in AI toward novelty. The system gets rewarded for proposing the surprising structure, the non-obvious scaffold hop, the move no human would have thought of.

In live drug discovery, a lot of value comes from low-hanging fruit. Not because teams lack imagination — because low-hanging fruit is often exactly where the best information-per-week lives. If you already have a route and an intermediate that let you probe a subtle steric pocket, tune basicity, shift lipophilicity, or test vector growth at low cost, that is not boring. That is excellent project economics.

The best next compound is often not the most imaginative one. It is the one that moves the project frontier fastest given the chemistry already in hand. A route-aware design model should understand this, and one of its core jobs should be identifying these local, tractable, high-information moves before reaching for heroic chemistry.

This changes the benchmark too

Once you take synthetic practicality seriously, standard molecule-design evaluations look thin. Generating potent-looking molecules is not enough. Can the system prefer designs reachable from existing intermediates? Can it distinguish between “make tomorrow” and “maybe next quarter”? Can it trade a small predicted gain in potency against a large increase in synthetic burden?

A design system that ignores these questions will routinely over-recommend impressive but mistimed ideas and under-recommend the compounds that actually accelerate learning.

The future molecule design stack should look less like target + actives + model = molecules and more like project state + assay data + route graph + intermediate inventory + procurement reality + model = actionable next compounds.

Medicinal chemistry is not molecular optimisation. It is decision-making under synthetic constraint. Routes and intermediate availability belong in the input to the design algorithm, because a small molecule project is not judged by how many elegant structures were proposed. It is judged by how quickly the right molecules get made, tested, and turned into the next round of better decisions.

That loop runs through the bench.