What Should Your AI Model Learn From the Next Plate of Experiments?
Imagine your model proposes 96 protein designs. The lab tests them and returns 12 hits. What happens to the other 84? Some may bind weakly. Others may bind the controls, produce inconsistent results, or fail to yield a
Imagine your model proposes 96 protein designs. The lab tests them and returns 12 hits.
What happens to the other 84?
Some may bind weakly. Others may bind the controls, produce inconsistent results, or fail to yield a usable measurement. Grouping all of them under βunsuccessfulβ leaves out information that could shape the next design round.
For teams running repeated designβtestβlearn cycles, deciding what to send back from the lab is part of designing the experiment.
Google Researchβs October 2026 aptamer work offers a useful example. Across three design rounds, researchers returned results from bead-based screening and sequencing to AI agents. Those results helped the agents investigate cross-reactivity and adjust their subsequent designs.
The feedback gave them a way to understand why promising candidates were falling short and choose what to try next.
For teams designing proteins, the same idea raises a practical question: what would your model need to know to make a better decision?
In my earlier post on AI-designed binders, I explored how experimental validation and automation can support candidate evaluation. The next step is deciding which parts of that evaluation should inform the next round.
A single βgood proteinβ label can hide several different properties:
| Measurement | What it can tell the team | Context to keep with the result |
|---|---|---|
| Binding affinity, KD | How tightly a candidate binds under the tested conditions | Concentration range, assay conditions and fit quality |
| Association and dissociation rates | How quickly binding forms and comes apart | Binding curves and uncertainty in the fitted rates |
| Specificity controls | Whether a candidate also binds selected off-targets or assay surfaces | Control identities and their measured responses |
| Thermal behaviour | How candidates or formulations respond to heating | Full thermal curves, buffer conditions and interpretable transitions |
| Measurement uncertainty | Whether an apparent difference is reliable enough to guide selection | Replicates, variability and quality-control flags |
Consider a candidate that binds tightly but also produces a strong response against an unrelated control. Another might show weaker target binding but a cleaner control profile. Which one deserves further development depends on the goal of the programme.
Keeping those measurements separate makes the trade-off visible. If the team combines them into one ranking, retaining the underlying values and their weighting makes that ranking easier to understand and revise.
The same care matters when an experiment does not produce a clear answer.
βNo detectable binding under the tested conditions,β βbinding detected but rates could not be fitted reliably,β and βsample failed quality controlβ should remain distinct outcomes. Each points towards a different next step.
I would keep those records linked to the original design, sample batch, assay conditions and underlying curves. That gives the team a way to decide whether to redesign a candidate, repeat a measurement or investigate the assay before using the result as feedback.
It also makes uncertainty useful. A borderline result can help identify where another experiment would be most informative.
This is where complementary biophysical measurements can contribute.
For suitable protein samples, SUPR-DSF measures thermal behaviour through intrinsic fluorescence in standard 384-well plates. That format supports comparisons across candidates and buffer conditions, with space allocated for controls and repeats.
Alto or Revo can add binding kinetics and affinity measurements, with automated serial dilutions reducing manual preparation. Together, these measurements could help teams compare how sequence changes affect binding and thermal behaviour across successive rounds.
Each measurement still answers its own question. A higher melting temperature does not establish better biological activity, and a thermal shift does not generally translate directly into a binding-affinity ranking.
Consistency across rounds deserves attention, too. A manufacturer application note reports thermal measurements across 16 plates of lysozyme, providing a practical example of assessing repeatability within a larger screening workflow.
For me, the starting point is to agree on what the next round needs to resolve, then choose the measurements and controls that can answer it.
Some rounds may benefit from testing more designs. Others may benefit from taking a closer look at uncertain candidates. Preserving both the successful results and the reasons behind unsuccessful ones gives scientists and computational systems more useful information to work with.
Desmond Gilmour
Email: [email protected]
Phone: +1 519-546-9876
I work at Nicoya Lifesciences. The views expressed here are my own and do not necessarily reflect those of Nicoya. This is a personal publication, not an official company communication.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.