Aptamers as a sandbox for lab-in-the-loop agentic molecular discovery

Zachary Cutts; John Coller; Rachel Chen; Hua Tang; Mathias Bellaiche; Ryder Easterlin; Dina Zverinski; Petar Sirkovic; Avinatan Hassidim; Yossi Matias; Pooja Rao; Annalisa Pawlosky

AI summary · arXiv: 2610.24806v1 · q-bio.BM

Aptamers as a sandbox for lab-in-the-loop agentic molecular discovery

Zachary Cutts, John Coller, Rachel Chen, Hua Tang, Mathias Bellaiche, Ryder Easterlin, Dina Zverinski, Petar Sirkovic,

and 4 more authors

Avinatan Hassidim, Yossi Matias, Pooja Rao, Annalisa Pawlosky

Original paper Submitted to arXiv on 17 Oct. 2026 Equal leadership (co-senior authors): Pooja Rao and Annalisa Pawlosky. 34 pages, 4 main figures, 1 main table, 3 supplementary figures, 4 supplementary tables License: CC BY 4.0 Also in: fr

How to read this summary? Illustrated guide

The essentials

Three rounds of AI-guided aptamer design with experimental feedback increased measured binder hit rates and helped recover target specificity after cross-reactivity emerged.

Question

Can LLM-based reasoning agents design aptamer binders without prior target-specific training data and improve their designs through repeated wet-lab feedback? The study also asks whether experimental measurements add useful information beyond computational predictors.

Method

The authors ran three design–test–improve rounds against human α-thrombin and myoglobin, using agent-generated pools of roughly 1,000 sequences per round and bead-based pulldown with sequencing readouts; the campaign included about 4,400 sequenced constructs. Agents analyzed sequencing data between rounds, while the authors used differential enrichment and bare-bead background correction to call hits; eight full-length α-thrombin candidates were also tested in single-clone fluorescence titrations across 0–1,000 nM.

Findings

Result Among newly designed α-thrombin candidates, bead-corrected hit rates increased from 0.0% in Round 1 to 42.1% in Round 3.
See the passage in the paper
On $\alpha$-thrombin, where uninsulated Round 1 designs failed due to primer–core hybridization (0.0%, 0/167), hit rates climbed to 26.0% in Round 2 (54/208) and 42.1% in Round 3 (131/311; 37.1% [106/286] across strictly de novo candidates free of known thrombin motifs; Fig. 2B; Supplementary Table S3).
Result On myoglobin, agents shifted away from G-quadruplex designs after counter-screening revealed cross-reactivity, and Round 3 specificity and hit rate recovered.
See the passage in the paper
On myoglobin, after keeping the library G4-heavy in Rounds 1–2 (58.2%–60.7% G4s), the agents cut unclamped G4s from 20.4% to 1.3% and pivoted 72.3% of the Round 3 pool into non-G4 scaffolds (Fig. 3B, left).
Result Most tested computational predictor metrics performed near or below chance on the Round 3 myoglobin library, while two target-agnostic folding metrics showed meaningful signal.
See the passage in the paper
Target-conditioned 3D co-folding (Boltz-2 iPTM and $\Delta\text{iPTM}$: $\text{ROC-AUC}=0.499$ and $0.391$), 3D inverse-folding and structure models (ESMFold2-NA, RhoDesign, RiboDiffusion: $0.496\text{--}0.510$), nucleic-acid foundation models (DNABERT-2, HyenaDNA, RNA-FM: $0.364\text{--}0.523$), and G-quadruplex scorers (G4Hunter $0.366$, ViennaRNA $\Delta\Delta G_{\text{G4}}$ $0.476$) all performed near or below random chance ($0.50$).
Result Single-clone titrations supported a quantitative relationship between pooled sequencing enrichment and measured binding, with the strongest literature-derived candidates showing preliminary nanomolar apparent affinities.
See the passage in the paper
Six of the eight constructs (75%) exhibited concentration-dependent binding above blank, and endpoint fluorescence at $1\ \mu\text{M}$ scaled quantitatively with Round 3 pooled enrichment across nearly four orders of magnitude (Pearson $r=0.90$, $p=0.0021$; Spearman $\rho=0.88$, $p=0.0038$; Fig. 4B, C), confirming that pooled $\log_{2}\text{FC}$ reflects genuine equilibrium binding rather than amplification artifacts.

Limitations

The fitted apparent dissociation constants are preliminary because the titration measurements did not reach a saturation plateau.
See the passage in the paper
These fitted $K_{D}$ values should be considered preliminary estimates, as fluorescence signals did not reach a full saturation plateau ($B_{\max}$) within the $1\ \mu\text{M}$ concentration range tested.
The study did not test fine selectivity against close structural homologues.
See the passage in the paper
Catching and correcting assay-level matrix artifacts is an essential first gate, although fine discrimination against close structural homologues (such as myoglobin versus hemoglobin, or $\alpha$-thrombin versus prothrombin) was not tested here and remains a next step for counter-selection campaigns.
The campaign did not include matched control arms that would isolate the contribution of agentic reasoning from other factors.
See the passage in the paper
While our campaign establishes closed-loop feasibility, prospective trials with matched-budget no-feedback and non-reasoning control arms will be needed to isolate the specific contribution of agentic reasoning, and head-to-head benchmarking against evolutionary, Bayesian, and supervised machine learning baselines [31, 32, 33, 36, 37] is an important direction for future work.

Why it matters

The authors argue that fast, multiplexed aptamer assays can provide empirical feedback for iterative agentic molecular discovery, including diagnosis of assay artifacts and redesign after failure. They present this as evidence of closed-loop feasibility in aptamers and a possible basis for extending the approach to more complex biomolecular design tasks.

Written by AI from the paper. Passages are checked word for word against the text we analysed; the rest may contain errors.

AI-generated keywords: AI reasoning agents DNA aptamers laboratory feedback molecular design binding specificity

Read further

Comprehensive Summary
Layman's Summary
Blog article
Original abstract

The paper investigates whether AI reasoning agents can improve molecular designs by repeatedly learning from laboratory results, rather than generating candidates in silico and testing them only once. The authors use DNA aptamers—short DNA molecules that bind proteins—as a test case because they can be synthesized and screened quickly. Each experimental cycle took about 10 days and cost approximately $3 per candidate, providing quantitative results for both successful and unsuccessful designs. The team tested designs against human alpha-thrombin, which has known nucleic-acid-binding grooves, and myoglobin, a more challenging target without a natural nucleic-acid-binding pocket. Across three design–test–improve rounds, agents proposed pools of roughly 1,000 candidates. Researchers synthesized and screened these candidates using magnetic-bead pulldowns, while sequencing measured enrichment on beads coated with the target protein, other proteins, or no protein. Agents then used the resulting count data to revise their designs. Google’s AI Co-Scientist supported initial ideation in Round 1; Rounds 2 and 3 used a specialized pipeline for statistical analysis, design strategy, candidate generation, and computational scoring. Among newly designed candidates, statistically significant, bead-corrected hit rates against alpha-thrombin rose from 0% in Round 1 (0 of 167) to 26.0% in Round 2 (54 of 208) and 42.1% in Round 3 (131 of 311). Excluding descendants of known thrombin motifs, the Round 3 hit rate was 37.1% (106 of 286). For myoglobin, hit rates were 10.2% (20 of 196), 3.7% (9 of 244), and 33.5% (184 of 549) across the three rounds. The temporary decline in myoglobin performance was associated with the agents’ reliance on G-quadruplex structures, which often bound nonspecifically to beads or multiple proteins. Analysis of Round 2 results led the agents to identify G-quadruplexes as a major source of cross-reactivity: 97% of cross-reactive binders belonged to this structural class. In Round 3, the agents reduced unclamped G-quadruplex designs and shifted toward other structures. Myoglobin specificity increased from 19.1% in Round 2 to 72.4% in Round 3, while hit rates in several non-G-quadruplex families ranged from 45.9% to 80.4%. Thrombin specificity reached 98.5% in Round 3. Computational predictors tested on the Round 3 myoglobin library generally performed near random chance. Two sequence-only folding measures performed better, with ROC-AUC values of 0.729 and 0.763, but they could not predict target-specific binding. Separate tests of eight full-length thrombin candidates supported the pooled-screen measurements: enrichment and fluorescence binding correlated strongly (Pearson r = 0.90). Four candidates derived from known thrombin aptamer motifs had preliminary apparent affinities of 24.6–133 nanomolar. Strictly de novo candidates showed weaker binding, at or above the micromolar range in the tested assay. The authors conclude that rapid, counter-screened laboratory feedback can help agents diagnose failures and redirect designs, and that empirical measurements were more useful than the computational proxies tested. However, the study establishes feasibility rather than proving that agentic reasoning itself caused the gains: it did not include matched campaigns without feedback or campaigns using alternative methods. Other limitations include preliminary affinity estimates that did not reach saturation, single-clone validation only for thrombin, and no testing against closely related protein homologues.

Can AI Improve Molecular Designs by Learning From the Lab?

Many AI tools for molecular discovery follow a one-way path: generate a batch of candidates on a computer, then test them in the laboratory. The paper “Aptamers as a sandbox for lab-in-the-loop agentic molecular discovery,” by Zachary Cutts, John Coller, Rachel Chen, Hua Tang, Mathias Bellaiche, Ryder Easterlin et al., explores a more iterative approach. Could AI agents use experimental results—including failures—to make better designs in the next round?

To investigate, the researchers focused on DNA aptamers: short DNA molecules that can bind to proteins. Aptamers are useful for this kind of study because they are relatively quick and inexpensive to make and screen. Each cycle took about 10 days and cost approximately $3 per candidate, giving the team a way to test whether repeated laboratory feedback could shape later designs.

A Repeated Design–Test–Learn Cycle

The team studied two protein targets. Human alpha-thrombin has known grooves that bind nucleic acids, making it a comparatively accessible target. Myoglobin is more challenging because it lacks a natural nucleic-acid-binding pocket.

Across three rounds, AI agents proposed pools of roughly 1,000 candidate sequences. Researchers synthesized the designs and tested them with magnetic-bead pulldowns. The beads were coated with the target protein, other proteins, or no protein, allowing the team to distinguish target binding from broader stickiness. Sequencing then measured how strongly each candidate was enriched in each condition. The agents used those results to guide their next proposals.

Google’s AI Co-Scientist supported initial ideation in Round 1. For Rounds 2 and 3, the researchers used a specialized pipeline combining statistical analysis, design strategy, candidate generation, and computational scoring. The important feature was not just generating candidates, but returning measured outcomes to the system before the next round.

Performance Improved—But Not in a Straight Line

For newly designed candidates targeting thrombin, the statistically significant, bead-corrected hit rate climbed from 0% in Round 1 to 26.0% in Round 2 and 42.1% in Round 3. The final-round rate remained substantial even when descendants of known thrombin motifs were excluded: 37.1% of those candidates were hits.

Myoglobin results were less consistent. Hit rates were 10.2% in Round 1, fell to 3.7% in Round 2, then rose sharply to 33.5% in Round 3. That temporary setback offered a useful lesson: early success or failure did not simply tell the system to produce more of the same. The researchers could examine why candidates were failing and adjust the design strategy.

Learning to Avoid Sticky Structures

A major issue in Round 2 was the agents’ reliance on G-quadruplexes, DNA structures that can be useful but may bind nonspecifically. Analysis showed that 97% of cross-reactive binders belonged to this structural class. In other words, many candidates seemed promising because they stuck to beads or multiple proteins, rather than because they selectively recognized myoglobin.

After that pattern emerged, the agents reduced designs featuring unclamped G-quadruplexes and shifted toward other structural families. Myoglobin specificity rose from 19.1% in Round 2 to 72.4% in Round 3. Several non-G-quadruplex families achieved hit rates between 45.9% and 80.4%. Thrombin specificity also reached 98.5% in the final round.

This is one of the study’s clearest examples of the value of counter-screening. Testing candidates against other proteins and protein-free beads helped reveal not only which sequences worked, but which apparent successes were misleading.

Laboratory Measurements Beat the Tested Predictors

The researchers also compared their experimental results with computational predictions for the Round 3 myoglobin library. Most predictors performed close to random chance. Two sequence-based folding measures did better, with ROC-AUC values of 0.729 and 0.763, but they could not determine whether a structure would bind the specific target. A plausible fold, the results suggest, is not the same as a reliable prediction of target-specific binding.

Separate tests of eight full-length thrombin candidates supported the pooled-screen measurements: enrichment in the screen correlated strongly with fluorescence binding results, with a Pearson correlation of 0.90. Four candidates derived from known thrombin aptamer motifs showed preliminary apparent affinities of 24.6 to 133 nanomolar. Strictly de novo candidates bound more weakly, at or above the micromolar range in the tested assay.

What the Study Does—and Does Not—Show

The work demonstrates the feasibility of a lab-in-the-loop workflow: agents can receive experimental data, identify a failure pattern, and alter subsequent designs. It also suggests that direct measurements can be more informative than the computational proxies tested here.

But the study does not prove that agentic reasoning itself caused the gains. It did not compare the campaign with a matched effort that received no experimental feedback, or with campaigns using alternative design methods. Some affinity estimates were preliminary and did not reach saturation; only thrombin candidates received single-clone validation, and the study did not test binding against closely related protein homologues.

For now, the strongest takeaway is practical rather than sweeping: when experiments are fast and affordable, their results can help guide the next design round. The researchers’ aptamer sandbox offers a concrete way to study that idea—and a reminder that useful AI-driven discovery may depend as much on learning from failed experiments as on generating clever candidates.

Created on 31 Oct. 2026

Was this synthesis useful?

Score: 0

Ask questions about this paper to our AI assistant

How to chat with this paper? Illustrated guide

Compare with the PDF

You can also chat with multiple papers at once here.

AI assistant instructions?

Tools for this paper

Explain it differently 3 free rewrites left today

Save to a collection Group papers in collections to compare them, chat with them and export them. Create a free account Log in

Export BibTeX · RIS (Zotero, Mendeley, EndNote) · Markdown (Notion, Obsidian)

Advanced tools Premium

For researchers who follow a field closely: citation alerts, journal-club slides, version comparison and Zotero export. Discover Premium

Zotero

Citation alerts Citation alerts guide

Who works on it? The people and institutions who publish on its topic

Weekly letter Presented in "Biology and medicine this week" of 26 October 2026. Biology and medicine this week

Journal club Premium+ Slides to present it

Versions What changed between versions

Talk about it in public Premium+ Press kit: release, questions, hype check Press kit guide

Papers around this one

See them on a map

Ranked by how close their title and abstract are to this paper’s. Select some to read them together.

Closest papers

Very close
Prospecting the protein design landscape.
Jakob R. Riccabona, Katharina T. Stonig, J. Meiler et al. · 2026 · FEBS Letters · Not summarized yet

Abstract
Generative AI has driven remarkable breakthroughs in protein design, enabling the rapid, computationally guided creation of high-affinity binders against diverse targets. While remarkable experimental success has been demonstrated, the confidence metrics used to filter and evaluate designs remain optimized for static protein interfaces and can fail when applied to underrepresented or conformationally complex targets. In this review, we outline the current landscape of deep learning-driven protein design pipelines, discuss tailored applications in peptide, small molecule, binder, vaccine, and antibody design, and argue that the integration of ensemble-based methods represents a promising aven…
Very close
Beyond SELEX: AI-driven aptamers from sequence to structure to function.
Bo-Ming Cui, X. Lian, Wei-Wei Song et al. · 2026 · Biotechnology Advances · Not summarized yet

Abstract
Conventional aptamer discovery by systematic evolution of ligands by exponential enrichment (SELEX) is often constrained by limited library capacity, PCR bias, and multi-round screening. The recent convergence of artificial intelligence with aptamer development is driving a paradigm shift from empirical screening toward rational, data-driven design. This Review presents a timely and comprehensive overview of AI integration across the entire aptamer development pipeline, with a particular emphasis on advances reported in 2024-2026. We begin by discussing bioinformatics databases and computational tools for high-throughput SELEX data analysis. We then review state-of-the-art strategies for sec…
Close
Agentic campaign control for high-throughput de novo binder design
Minkyu Jeon, Jinyeop Song, Jina Kim et al. · 2026 · bioRxiv · Not summarized yet

Abstract
Progress in artificial intelligence has produced a rapidly growing ecosystem of methods for de novo protein design. With access to many specialized and often complementary tools, how does one use them effectively, especially with a finite compute budget? Here, we introduce Target-adaptive Rescue–Explore–eXploit (T-REX), an agentic protein binder design framework that orchestrates multiple state-of-the-art protein generative models and structure evaluators. Over the course of a design campaign, T-REX leverages large language model (LLM) agents to reason over accumulating outcomes and decide whether to rescue promising candidates, explore alternative settings, or exploit productive routes, whi…
Close
UnibindMD: An Integrated AI and Molecular Dynamics Platform for De Novo Peptide Binder Design
Xuan Wu, A. Huang, Cong-Mei Fu et al. · 2026 · bioRxiv · Not summarized yet

Abstract
UnibindMD is an integrated computational platform that turns a protein target into validated de novo peptide-binder leads by chaining five orthogonal stages: deep-learning binder design (BindCraft), independent structure prediction and confidence assessment (AlphaFold3-class / Boltz-2), binding-affinity prediction (PRODIGY), developability profiling (ADMET-AI), and all-atom molecular-dynamics validation (GROMACS/CHARMM36m). The platform targets a persistent weakness of single-method design, namely that confidence metrics such as ipTM and pLDDT are self-consistent but not independent, so a design predicted to be confident may still be physically unstable. We demonstrate UnibindMD on three And…
Close
AnewDDE: An Agentic Drug Discovery Engine for Biomolecular Interaction Modelling and Closed-Loop Design
Kai Liu · 2026 · bioRxiv · Not summarized yet

Abstract
Accurate modelling of biomolecular interactions is fundamental to drug discovery, yet current artificial intelligence (AI) workflows remain fragmented across structure prediction, affinity estimation, molecular design, and experimental decision-making. We introduce AnewDDE, an agentic Drug Discovery Engine that connects these capabilities into a closed-loop system for biomolecular interaction modelling and design. We demonstrate several components of this system: AnewFold delivers superior performance on challenging targets, including antibody–antigen complexes and molecular glues, providing more reliable structural hypotheses together with pocket identification and conformational analysis; …
Close
Agent-guided epitope excision enables economical de novo binder design
Jun W. Kim, Artiom Myaskovsky, O. Donosa et al. · 2026 · Not summarized yet

Abstract
De novo protein design campaigns remain bottlenecked by the inference compute time required to generate and rank candidates, limiting wider functional exploration. Here, we introduce and evaluate an epitope excision approach that truncates targets to local interface windows to accelerate both backbone diffusion and structure prediction. Across 15 targets and 3,666 designs, excised windows closely reproduced full-target binding interfaces (median C\alpha RMSD of 0.37 \AA) and made it practical to design binders against the 719-residue extracellular domain of fibroblast activation protein alpha (FAP) by cutting RFdiffusion and AlphaFold3 GPU time by ~ 88% in one of the epitopes. Pairing this w…
Close
Computationally Guided Design and Experimental Validation of ssDNA Aptamers Targeting the SARS-CoV-2 Spike Protein
José C. E. Márquez Montesinos, Ana Maria Bernardez, Yuliet Mazola et al. · 2026 · Biomolecules · Not summarized yet

Abstract
Aptamer development has traditionally relied on iterative, labor-intensive experimental approaches. Recent advances in bioinformatics and computational power now enable the rational in silico design of high-affinity nucleic acid ligands, drastically reducing time and cost. Here, we present an integrated computational pipeline for designing ssDNA aptamers targeting the receptor-binding domain (RBD) of the SARS-CoV-2 Spike (S) protein. A library of 4000 randomly generated 20-nucleotide sequences with defined base ratios (C:G:A:T = 3:3:2:2) was screened based on secondary structure and free energy. The 100 most stable sequences (ΔG ≤ −4.1 kcal/mol) were modeled in three dimensions, yielding 100…
Close
Validation and analysis of 12,000 AI-driven CAR-T designs in the Bits to Binders competition.
Clayton W. Kosonocky, Alex M. Abel, Aaron L. Feller et al. · 2026 · Molecular Systems Biology · Not summarized yet

Abstract
Artificial intelligence (AI) methods for proteins have advanced rapidly, improving structure prediction and design, particularly for de novo binders. However, most evaluations emphasize binding affinity rather than higher-order biological function. We present Bits to Binders, a global competition evaluating de novo binder design in the context of chimeric antigen receptor (CAR) T cells. Teams from 42 countries submitted 12,000 designs of 80-amino acid de novo minibinders targeting human CD20 as CAR binding domains. Designs were screened by pooled CAR-T proliferation, identifying 707 designs exhibiting significant CD20-specific enrichment, with team hit rates from 0.6% to 38.4%. Top-performin…

Go further

How are these papers found?

We compare the title and abstract of this paper with those of every paper summarized on this site, and with the papers Semantic Scholar recommends, those it cites and those citing it (for a much-cited paper, its most recent citations and, from OpenAlex, the most cited ones). Among the prior work and the follow-ups, the papers that use each other in depth, a method it uses and the much-cited follow-ups come a little higher. The levels (very close, close, related) come from the similarity of the texts, not from a judgment of quality; * marks a paper without an abstract, compared by its title. The sentence under a paper is written by an AI from the two abstracts: check it in the papers. A paper marked not relevant by three readers leaves the list.

Updated on 31 October 2026.

Disclaimer: The AI-based summarization tool and virtual assistant provided on this website may not always provide accurate and complete summaries or responses. We encourage you to carefully review and evaluate the generated content to ensure its quality and relevance to your needs.

Aptamers as a sandbox for lab-in-the-loop agentic molecular discovery

The essentials

Question

Method

Findings

Limitations

Why it matters

Read further

Definitions

Can AI Improve Molecular Designs by Learning From the Lab?

A Repeated Design–Test–Learn Cycle

Performance Improved—But Not in a Straight Line

Learning to Avoid Sticky Structures

Laboratory Measurements Beat the Tested Predictors

What the Study Does—and Does Not—Show

Ask questions about this paper to our AI assistant

Tools for this paper

Advanced tools Premium

Papers around this one

Closest papers

Go further

Cite this paper (BibTeX)