Wet-lab feedback lifts agent-designed aptamer hit rates to 42% after zero-shot structure models fail
A Google Research, DeepMind and Stanford preprint shows three rounds of wet-lab feedback raised alpha-thrombin aptamer hit rates from 0% to 42.1% while 12 zero-shot biomolecular models scored near random chance.
Can current 3D biomolecular foundation models design functional single-stranded DNA aptamers zero-shot?
In an October 17 arXiv preprint (arXiv:2610.24806) evaluating 4,405 synthesized and sequenced constructs against human alpha-thrombin and myoglobin, researchers from Google Research, Google DeepMind and Stanford benchmarked 12 zero-shot prediction metrics across nine tools (including Boltz-2, ESMFold2-NA, DNABERT-2 and ViennaRNA). Ten of the 12 metrics scored near or below random chance (ROC-AUC 0.364 to 0.523); only target-agnostic 2D thermodynamic stability and ensemble purity provided useful zero-shot filtering. [1][2]
When the team placed a frontier reasoning agent inside a three-round iterative wet-lab loop using high-throughput sequencing feedback, experimental alpha-thrombin binding hit rates rose from 0.0% (0/167, Round 1) to 26.0% (54/208, Round 2) and 42.1% (131/311, Round 3; 37.1% strictly de novo). On myoglobin, when Round 2 enrichments collapsed target specificity from 75.0% to 19.1% due to G-quadruplex (G4) cross-reactivity, the agent identified the G4 artifact from assay data, constrained G-runs in Round 3 and restored specificity to 72.4%. [1]
A critical biophysical caveat separates screening hits from therapeutic affinity. In single-clone bead-based fluorescence titrations, only chimeric designs seeded with known literature aptamer motifs achieved nanomolar dissociation constants (KD of 24.6 to 132.7 nM), whereas strictly de novo sequences remained in the micromolar range (>1 μM). Closed-loop assay feedback is therefore essential for filtering false positives, not a substitute for affinity maturation. [1]
Take it further
Read the aptamer wet-lab benchmark and zero-shot ablation preprint For computational biologists and wet-lab automation teams. Details the 12-metric zero-shot failure modes and Round 2 G-quadruplex specificity recovery.
Sources & details 2 sources
Discuss in your voice app
Take the reporting, visual explanations and original sources into a conversation. Copy the brief, paste it into your preferred assistant, then start its voice mode.
Paste the brief into your voice app to discuss this reporting.Reporting sources
- arXiv · Aptamers as a sandbox for lab-in-the-loop agentic discovery (2610.24806)
Original October 17, 2026 preprint (arXiv:2610.24806v1) by Cutts et al. Reports 4,405 constructs across three wet-lab rounds, 12 zero-shot baseline ROC-AUCs (0.364–0.523) and bead-titration affinity bounds.
- GitHub · jwohlwend/boltz biomolecular interaction model baseline
Open-source biomolecular structure prediction repository for Boltz-2, one of the zero-shot structural baselines evaluated against wet-lab aptamer enrichment.
Images & figures
Iterative wet-lab feedback lifted thrombin hit rates to 42.1% and repaired Round 2’s specificity collapse.
Euangelion · original data figure · Source · CC BY 4.0 · Download figure