From high-throughput screening to AI learning at the bench

A crash course on aptamers

I was fascinated by Google’s recent publication, which incorporates AI into the aptamer discovery workflow. I explored this area as a PhD student and the field has advanced significantly during my subsequent years of medical training. Google’s new paper is a remarkable example of how far it has come.

An aptamer is a short snippet of DNA or RNA that attaches to a target like a protein.

Aptamers look like this:

☝🏾 That is a 15-nucleotide sequence of single-stranded DNA.¹ I drew it squiggly because aptamers fold in interesting ways, which is what allows them to stick to things. I discovered that the above 15-nucleotide aptamer strongly binds simple sugars, including glucose.

I wasn’t intending to find sugar-binding aptamers, however!

My PhD research was actually looking for aptamers that bind Cryptosporidium parvum, a waterborne pathogen that can spread through contaminated water, including public pools. If your local waterpark gets shut down, Cryptosporidium could likely be the culprit. The thinking was this: maybe we could design a way to detect Cryptosporidium in water by finding an aptamer that binds it strongly and then attaching that aptamer to something that gives off a signal when it latches on, a colour change or a flash of fluorescence, so a sample could be checked on the spot instead of sent away to a lab.

I was also trying to discover aptamers that bind human α-thrombin, a protein involved in blood clotting (Google’s new study also used thrombin as one of its targets). One aptamer that binds thrombin has been known for decades and is thoroughly characterised, so the field uses it to test new discovery methods. That is why thrombin shows up in both my work and Google’s: new methods should be able to find aptamers that are already known.

One sequence confused us for weeks

In our 2011 study, we started with a DNA library, basically a tube holding every possible 15-letter sequence at once, around a billion of them with about 56,000 copies of each. We used deep sequencing to find thrombin aptamers after a single selection step, rather than the usual repeated cycles of selection and amplification.²

At the time, combining an over-represented library with deep sequencing in this way was a novel approach, part of a broader effort to make discovery faster. We found the well-known thrombin aptamer, which proved the viability of our method. We also found a family of previously-undescribed variants, whose abundance tracked how tightly they bound.

Surprisingly, we found one sequence, the GCTATCATCGCAACG sequence drawn above, which was the most abundant sequence in the whole experiment behind only the thrombin aptamer itself.

Was this a new thrombin-binding aptamer?

Answering that took weeks at the bench. First we asked if the sequence was an artifact rather than a binder at all.³ Nope. Then we measured how tightly it actually held thrombin. Not very well at all.

So what was it binding to!?

We went back through everything the DNA had touched. The answer turned out to be sitting in the buffers, the salt solutions a protein is stored and washed in. Ours carried simple sugars, glucose and α-methyl mannoside, left behind by the way the thrombin had been prepared. GCTATCATCGCAACG was binding those sugars!

The signal was real, but did not belong to the thrombin target we were selecting for.

AI comes for aptamer discovery

In the years after I graduated and completed my medical training, researchers began using machine learning to help choose which sequences to test. Google’s 2021 study combined experiments with neural-network predictions. Now its recent paper takes another step: AI agents propose sequences, a human team tests them, and the results help guide the next round of experiments.

The new study, which went up on 17 October, reports that by the third round, roughly four in ten newly-designed thrombin candidates and one in three myoglobin candidates counted as on-target hits: they showed evidence of binding the intended protein under the study’s screening criteria. That is promising because researchers can focus follow-up work on candidates that pass this initial test.⁴

What struck me is how the role of computation changed. When I was grinding away at my PhD in 2011, sequencing made it possible to peer deeper into a selection experiment. Now in 2026, machine learning and automation can use experimental results to iteratively improve subsequent rounds.

Google runs into a familiar frustration

Google’s team hit the same kind of trouble with myoglobin as I did with thrombin. In their protocol, they attached each target protein to tiny magnetic beads, and then tested which DNA sequences bound them. Some promising candidates bound strongly but it turned out that they were “sticky.” That is, they tended to cling to things generally, not the intended target (myoglobin). It turned out that Google’s sticky sequences were binding to the bare magnetic beads or other proteins in the experiment.

That is the same trap my sugar-binding sequence set for us fifteen years earlier. A strong signal tells you that something bound but it does not tell you what it bound to. Both cases show why you run controls and follow-up tests before calling a sequence a success.

The remarkable part of Google’s project is that the detection of nonspecific binding was directly AI-related. The AI agents recognized the myoglobin cross-reactivity from the Round 2 sequencing and counter-screen results, and revised their design rules for Round 3 themselves, reducing overrepresented sticky structures and exploring other DNA shapes.⁵

Human researchers still prepared and ran the experiments. AI’s role was to interpret the readout and guide the next candidate pool. However, even as I write this, numerous teams are racing towards the next step: removing human scientists from the bench and allowing AIs to both plan and execute experiments with autonomous “self-driving labs.”

I’m curious: have you seen AI-driven developments reshape a field you studied earlier in your career?

Sources

• Kupakuwana, G. V. et al. (2011). Acyclic Identification of Aptamers for Human alpha-Thrombin Using Over-Represented Libraries and Deep Sequencing. PLOS ONE.
• Kupakuwana, G. V. (2011). High Throughput Screening of Aptamers. Syracuse University dissertation record and abstract.
• Bashir, A. et al. (2021). Machine learning guided aptamer refinement and discovery. Nature Communications.
• Cutts, Z. et al. (2026). Aptamers as a sandbox for lab-in-the-loop agentic molecular discovery. Google Research paper.

Footnotes

1. In our library every one of these sat on the same short run of constant bases, the same in every member, which held it steady while the fifteen folded. ↩

2. Not quite no amplification at all: the selected material was still amplified once, to prepare it for sequencing. ↩

3. The steps that join and copy DNA can favour some sequences on their own, whatever they bind. ↩

4. That does not mean these candidates are ready to become medicines. Binding is an early step; a candidate still needs more testing for strength, selectivity, safety and usefulness in the body. ↩

5. Google has not released the sequences of its own best binders: the paper reports their affinities and shapes and says the full sequence tables will follow with peer-reviewed publication. ↩