Design the molecules in software and test a few at the bench

A computer can search an enormous set of possible drug molecules against a written specification and hand back a few worth making, so the compounds that reach a bench should be the survivors of that search, and the experiment that follows decides whether the search was right. An assay development AI assistant belongs at exactly that step, where the list has come down to a handful of molecules and somebody has to plan the measurement that settles them.

Searching a design space

The method comes from engineering. Computer-aided design spread through aerospace and car manufacturing from the 1960s onward, and it was built for a practical reason, which is that a physical prototype costs weeks and money while a simulated one costs an afternoon. You write down what you want from a design, a program generates candidates and scores each against that description, and the few that score best get built.

Route planning on a phone works this way, since you type in a destination and the software searches a very large number of road combinations against travel time and distance before returning two or three, and you drive exactly one of them. Crash testing works this way too, because a carmaker runs the collision on a computer model of the vehicle hundreds of times, changing the thickness of a beam or the shape of a bracket between runs, and builds a physical car to drive into a wall once the simulated design has stopped improving.

A specification usually asks for qualities that pull against each other, and that is where a search of this kind does its work. Choosing a laptop, you want it cheap, light and long-lasting on battery, and pushing any one of those costs you one of the others, so what you pick from is a set of reasonable compromises. A drug molecule is specified the same way, because it has to bind its target strongly, stay selective for that target, dissolve, survive the liver and cross into a cell, and the chemistry that buys one of those often gives up another.

Why it matters in biology

The count of possible drug-like small molecules is usually put at around 10 to the power of 60, which is far past anything a laboratory can make and test. A physical screening library holds somewhere between a hundred thousand and a billion compounds, running one against a target takes months and sometimes years, and turning the hits that come out of it into a clinical candidate is several more years of chemists making molecules one batch at a time.

Predictive models change that arithmetic, because a model that scores a molecule for potency, selectivity and drug-like behaviour can work through millions of candidates on a computer in days, and reported savings in lead optimisation run to roughly 30 to 50% of candidate screening time. The clinical record so far splits in a way worth knowing about. Programmes built on AI-designed compounds report Phase I completion of roughly 80 to 90%, against a historical 40 to 65%, and Phase I measures tolerability, dose and how the body handles a compound, so that gap fits models that are good at picking molecules with sound chemical behaviour. Phase II, where a drug has to show it helps patients, sits near 40% for those same programmes, which is about where the industry has been for years.

Prediction has become strong on the chemistry and stayed ordinary on the biology, because whether blocking a chosen protein changes a disease is a question about a living system and it is answered by measurement. There is a real clinical signal from the new approach: a compound designed by a model against a target a model proposed improved forced vital capacity by about 98 mL against placebo at its best dose in a Phase IIa trial in idiopathic pulmonary fibrosis. One such result shows the route works, and the assay that a lab runs after its own computational search is what tells that lab whether its own prediction holds.

How Elnora fits

A chemistry team has twelve candidate inhibitors from a model they trained themselves to propose new structures, and bench time to synthesise four of them. They ask Elnora to rank the twelve on the evidence that already exists and to design the cell assay that will test whichever four they pick.

Elnora calls the team's own model as a tool, so the molecule design stays where the team built it, and then it works on each candidate with the public record: measured bioactivity for close structural neighbours from ChEMBL, medicinal chemistry filters that flag a structure likely to interfere with the assay readout, a patent search by chemical structure that returns the actual claims covering the nearest matches, and the target's disease evidence and tissue expression from Open Targets and the Human Protein Atlas. The ranking comes back with the evidence for each compound attached, and every cited paper is verified against the literature databases, so a chemist can open the record behind any line of it.

For the assay itself, Elnora picks a cell model on evidence, checking STR authentication through Cellosaurus and the target gene's dependency in DepMap, then lays out a 96-well plate with the dilution series, the vehicle and reference controls and as many replicates as the statistics support. A critic agent reads that layout, grades each weakness it finds Critical, Major or Minor and proposes a concrete fix while the plate can still be changed, and the materials list comes back with catalogue numbers resolved against live product pages. After the run, Elnora decodes the plate reader export into one table keyed by condition, runs the Z prime quality check and fits the four-parameter curve that gives each compound its potency value.

Choosing which four molecules are worth the bench time, making them and judging what the curves mean stay with the chemists.

The point

A computational search cuts an enormous set of possible molecules down to the few worth synthesising, and it is strong on chemistry and unproven on disease biology. The lab that gains from one is the lab whose next assay is designed well enough to settle whether the prediction held.