Rank designed antibodies on what your own lab measured

The useful job for AI early in biologics discovery is putting a panel of candidate molecules in order on several measured properties at once, and that ordering is worth only as much as the lab's own measurement record behind it. A variant that binds its target tightly and then comes out of the expression run at a low titre, or forms aggregates after a week at 4 degrees, has already cost the team a round of cloning, expression and purification to learn something the previous round's data could have flagged.

A candidate has to clear binding, expression and stability together

Finding a molecule worth developing means finding one that binds the right target at the right site with an affinity the biology calls for, behaves in the body as intended, and survives being made at scale. Those properties are measured in separate assays and they pull against each other, so a sequence change that buys affinity can cost thermostability or push a variant towards sticking to unrelated proteins.

The formats a biologics team works in now make that harder, because a multispecific antibody, an antibody drug conjugate and a cell therapy each add properties to balance on top of the ones a plain monoclonal already had. Scoring a candidate on one of them and treating the score as a verdict on the molecule is how a lab ends up synthesising the wrong four variants out of forty.

So the question a predictive model has to answer is the joint one, and what settles whether it answered correctly is a measurement from a bench. A model's affinity prediction is checked against a measured binding constant, its expression prediction against a titre, its aggregation prediction against what the size exclusion trace shows after storage, and each of those numbers comes out of a run somebody did.

Prediction gets better on data recorded the same way each round

Connecting design, testing and analysis so that one round's results arrive in the shape the next round reads is what makes a model improve across rounds. A lab that runs the thermal shift assay three times with the plate laid out differently each time, exports each reader file into its own folder, and writes the titres into a spreadsheet whose columns changed between rounds, has the measurements and still cannot compare them without a person reconciling the files by hand.

The record that trains and checks a prediction is a table of variants with the same properties in the same units, next to the protocol each number came from and the run that produced it. Keeping the failures in that table matters as much as the successes, because a variant that expressed at a tenth of the expected titre is the clearest signal a model has about what to avoid, and it is the one most often left out when a round gets written up.

One table across four rounds of expression and stability data

Elnora is the AI agent for lab work, and it reads a lab's protocols, results and instrument exports before it writes anything, then files what it produces back where the rest of the team looks. A protein engineering team working through rounds of candidate variants can hand it the raw files and get the comparison the rounds were run for.

A team with four rounds of variants uploads the plate reader exports from the expression titre and the thermal shift assay and asks Elnora for one table keyed by variant, with the variants that are missing a property marked, and a layout for the next round's plate. Elnora decodes each reader export into tidy data keyed by condition, runs the Z prime quality check before anything is concluded from a plate, fits the curves, and resolves each target and gene identifier through UniProt so one naming scheme holds across all four rounds. The finished table is filed into the shared Knowledge Base the team already keeps its protocols in, on the day the rounds are compared, and each save keeps the version before it, so a colleague who asks about a titre three weeks later gets the passage that holds it together with the file it sits in.

For the next round, Elnora writes the plate map down to the well, placing the variants, the reference standard and enough replicates for the comparison the team wants to make, and its critic specialist grades each weakness in that layout Critical, Major or Minor and proposes a fix while the plate can still be changed. A team that trained its own sequence design model can have Elnora call that model as a tool, so the generation stays where they built it and the ranking is done against the measurements they hold. Where the plate goes onto a connected liquid handler, Elnora writes the run file that machine accepts, rehearses it in simulation, and holds until a scientist has confirmed the deck.

Which variants are worth the bench time, and what a weak correlation between a prediction and a measured titre means for the next design round, stay with the scientists. The evidence for both of those calls sits in one table the whole team can open, with the run and the protocol behind each number still attached to it.