Choose which designed molecules to develop on measured toxicity

AI is getting good at drug design first, at predicting toxicity next and at predicting clinical efficacy last, so the decision that counts in a discovery lab is moving from making a molecule to choosing which molecules deserve development, and the strongest evidence a lab can produce for that choice within weeks is toxicity measured in its own cells. Elnora is the AI agent for lab work: it plans experiments, writes and optimises protocols, analyses results and remembers what a lab has decided, and for this choice it plans the toxicity screen beside the potency assay, fits both from the plate data to rank compounds by cytotoxicity and potency, and keeps the results where the next decision reads them.

Take a medicinal chemistry team with fourteen designed compounds that inhibit their target in the primary assay and bench capacity to take three of them forward. What separates the fourteen is how close the concentration that harms liver cells sits to the concentration where each compound works, and a 72-hour viability assay answers that.

Prediction improves first where an experiment answers in days

Molecule design is the part of drug development AI handles best because its feedback is quick and its benchmarks are clear. Binding, potency and the pharmacokinetic behaviour of a compound each have an established assay that returns in hours or days, the space of candidate structures can be searched on a computer, and a model can propose, be tested and improve over many rounds.

Toxicity is harder biology and still a structured problem. A liver toxicity signal has a recognisable shape, the same failures recur across molecules, the assays are established, and testing runs in stages from cells to organoids to animal models over weeks. Clinical efficacy sits furthest away, because whether a drug helps patients depends on disease heterogeneity, patient selection, dose, endpoints and trial design, it takes years to read out, and the long-term human data a model would need to predict it is still mostly missing.

Cheap molecules make the choice between them the scarce skill

Over the past decade the number of drug candidates in development has roughly doubled while new drug approvals have stayed at around 50 a year. Over two decades the typical number of drugs pursued against an attractive target has risen from two or three to seven or eight. As design tools improve, a potent molecule against a well-studied target becomes something several groups hold at once, and what sets one apart is the evidence beyond potency: whether it is safe at the dose it needs, whether the biology carries over into people, and whether there is a credible route into the clinic.

Of those, safety is the evidence a discovery lab can generate itself before a compound leaves the building, and it decides which of the fourteen deserve the next round of synthesis and testing.

A lab's own toxicity measurements are the data the next models learn from

Toxicity prediction improves as structured measurements build up across molecules, doses and cell systems, and a lab running viability and liver-cell assays on its own compound series is producing that kind of record. The record is useful when each run uses a consistent plate design and controls, sits with the protocol that produced it and keeps the compounds that failed, because a toxic analogue tells the next design round which structural feature to avoid.

The figure that ranks a candidate is its selectivity index: the CC50, the concentration that halves cell viability, divided by the IC50 against the target. A compound with a tenfold gap and one with a thousandfold gap can look identical in the primary screen.

Elnora plans the screen and ranks the fourteen by selectivity index

The team asks Elnora to plan a 72-hour viability assay in a liver-derived cell line for the fourteen compounds, across the concentrations used in the potency assay, and to rank them by the ratio between the two.

Elnora first searches the team's Knowledge Base with that question, and the earlier viability runs and the potency protocol come back as passages with their source files named, so the new assay matches the old conditions. It resolves each compound from its identifier in public chemistry databases and returns the measured bioactivity and toxicology on record for close analogues, so a structure with a known liability is flagged before it takes up wells.

It lays out the plates with a dilution series for each compound, vehicle controls and a reference cytotoxic compound on each plate. A critic agent reviews that design, grades each problem it finds and proposes a concrete change, such as moving the reference control out of the edge wells, while the plate can still be changed. Elnora then writes the run file for the lab's connected liquid handler and rehearses the run, and the scientist leading the screen signs off the rehearsed deck before anyone thaws the cells.

When the plate reader export comes back, it binds to the plate map, so the analysis runs on one table keyed by compound and concentration. Each compound is compared against the vehicle control on the same plate and gets a fitted dose-response curve with its potency value and the points behind it, and any well left out of a fit is listed with the reason. Elnora puts each compound's CC50 next to its IC50 from the primary assay, ranks the fourteen by selectivity index, and returns an editable workbook for the group meeting.

The results land in the shared Knowledge Base beside the protocol that produced them, toxic compounds included, so the next design round and the next team working on this target start from what this screen measured. The chemists choose which three molecules go forward and what the next design round should change.