AI-Designed Binders: From Predictions to Automated SPR Validation

If an AI model gives your team hundreds of promising protein designs, how quickly can you work out which ones deserve more attention?

As I continue learning about protein design and biophysical analysis, I keep coming back to this question. Generating candidates is exciting. Building a practical way to evaluate them is what makes those predictions useful.

The prediction is where testing begins

In its October 20, 2026 report, Anthropic described a campaign that generated 318,484 designs, selected 211 for experimental testing, and identified 17 that met its binder criteria.

Adaptyv Bio and Twist Bioscience used surface plasmon resonance, or SPR, to investigate binding in different protein-production formats. The criteria included reproducibility, an affinity threshold and no detected binding to an unrelated control. One target was assessed by only one laboratory, so the evidence was not identical across all reported hits.

The work explored possible antivenom research, but binding results did not establish neutralization or efficacy in living animals. It remains a company technical report and data release.

The broader lesson for me is simple: a prediction becomes more useful when we can connect it to experimental evidence.

What SPR adds to the decision

SPR follows binding near a sensor surface in real time. Its curves, called sensorgrams, help show how quickly an interaction forms and how quickly it comes apart.

That adds context to affinity, often reported as K_D: a lower value indicates tighter binding under the measured conditions.

Before trusting a ranking, I would ask:

• Does the binding result repeat?
• Do the controls support the claimed specificity?
• Do the curves support the fitted model and reported values?

Those questions matter because the measurement can also be influenced by the assay itself. For example, Schuck and Minton showed how molecular transport near a sensor can distort apparent binding rates.

As candidate numbers grow, the practical challenge becomes collecting this evidence without creating an overwhelming preparation workload.

Why Alto’s cartridge design matters

This is where I find the design of Alto digital SPR particularly relevant.

Nicoya specifies 16 channels, sample inputs as low as 2 µL, and automated serial dilutions on the cartridge. Actual sample requirements depend on the assay. For teams working with limited protein, these capabilities can make it easier to plan comparisons, repeats and controls while keeping sample demands manageable.

The automation advantage also starts with the consumable.

Alto’s sensors, sample wells and assay fluid handling are integrated into a disposable cartridge. Its well spacing supports single- and eight-channel pipetting, and the cartridge is designed for robotic handling. That gives an automation team a defined unit to prepare, move and load. These features are described in the Alto Automation Suite documentation.

For me, the useful benefit is the connection between sample preparation and measurement: the instrument can carry out programmed sample movements and dilutions after the cartridge is loaded.

Connecting SPR to a larger automated workflow

The Alto Automation Suite adds an API—a software interface—for checking instrument status, starting runs and retrieving results. It follows SiLA 2 laboratory automation standards, and Nicoya provides preconfigured labware definitions and workflows for the Opentrons Flex.

With the required liquid handling, robotic transfer and scheduling in place, the documented workcell supports unattended SPR operation.

It also creates an interesting option for scaling: one shared liquid handler could prepare cartridges for multiple Alto instruments while those instruments run their assays. The number it could support depends on preparation times, assay schedules, robotic access and sample-storage requirements, so that configuration needs to be validated during integration.

The aim is to make experimental feedback easier to collect and connect to the original design.

Adding a stability check with SUPR-DSF

Binding is one part of understanding a candidate. Its response to stress is another.

SUPR-DSF complements that workflow by measuring changes in a protein’s natural fluorescence during heating. Nicoya describes a 384-well format, 10 µL sample volumes, and standard qPCR plates and seals.

That provides a practical way to compare thermal behaviour across suitable candidates or buffer conditions, using familiar consumables.

Its plate format also supports automated sample preparation and robotic loading. A manufacturer application note documents robotic loading and measurement of 16 plates, with sample preparation performed as separate modules.

The methods answer different questions: SPR examines binding behaviour; DSF examines thermal unfolding behaviour. Intrinsic DSF needs a usable fluorescence signal, and a melting temperature alone does not establish shelf life or biological activity.

These product capabilities describe potential workflows; the Anthropic campaign was not a demonstration of Alto or SUPR-DSF.

What I’m taking from this

For teams designing proteins with AI, I think the most useful next step is to plan the experimental feedback alongside the computational pipeline.

Decide what evidence a candidate needs, then choose a workflow that makes collecting that evidence practical. Low sample requirements, automation-compatible consumables and connected instruments can help teams spend less time preparing samples and more time interpreting results.

---

These posts share my own perspective as I continue learning and are not official company statements.

AI assisted with research and drafting.

If you have questions about SPR, protein stability or how these systems could fit into your workflow, I’m happy to chat.

Desmond Gilmour

Email: [e-mail]

Phone: [phone]