Summary (Overview)

• Closed-loop agentic discovery: This paper demonstrates a lab-in-the-loop framework where LLM-based AI agents design, test, and iteratively refine DNA aptamers against protein targets across three design–test–improve cycles, with wet-lab feedback driving each round.
• Dramatic hit-rate improvements: Statistically significant, bead-corrected on-target hit rates rose from 0.0% → 26.0% → 42.1% on α-thrombin and from 10.2% → 3.7% → 33.5% on myoglobin across rounds (~4,400 total sequenced constructs).
• Autonomous failure diagnosis: Agents diagnosed a G-quadruplex specificity collapse (on-target specificity dropping from 75.0% to 19.1% on myoglobin) from raw sequencing data and recovered specificity to 72.4% by reallocating structural motif families.
• Quantitative validation: Single-clone bead-binding titrations confirmed pooled sequencing enrichment correlates with physical binding (Pearson r = 0.90), with agent-engineered literature motifs reaching preliminary apparent K_D = 24.6–133 nM.
• Computational predictors fall short: 12 zero-shot predictor metrics (9 tools) performed near or below random chance (ROC-AUC = 0.36–0.52) on the Round 3 myoglobin library, underscoring the value of empirical lab feedback over in silico proxies.

---

Introduction and Theoretical Foundation

Background and Motivation

Despite breakthroughs in AI-driven molecular generation, existing discovery pipelines operate largely in silico, with wet-lab feedback reserved for final validation rather than iterative improvement. Whether LLM-based reasoning agents can meaningfully accelerate de novo discovery by iterating with real-world lab data remained unproven.

The Bottleneck

• AI co-scientists (multi-agent systems) have begun accelerating biological research, but initial studies validate designs in only a single experimental pass.
• Self-driving laboratories in chemistry/materials science have closed the loop via robotic automation, but biomolecular design faces a harder challenge: synthesis, expression, and functional validation remain slow and expensive.
• In domains like games and mathematics, iterative search works because verification is computational (rulesets, compilers); in biomedicine, physical measurement is the bottleneck.

Why Aptamers as a Sandbox

The authors argue that single-stranded DNA (ssDNA) aptamers are ideal for closed-loop discovery because:

• The nucleotide sequence is the functional molecule — no cloning, expression, or purification steps required.
• Fast validation cycles: 1–2 weeks per round, $1–$3 per candidate.
• Quantitative multiplexed readouts: NGS returns positive and negative labels for every designed sequence.
• Computational predictors are poorly calibrated for nucleic acids due to sparse structural data, making empirical feedback essential.

---

Methodology

Closed-Loop Pipeline Architecture

The pipeline comprised five stages executed in series:

1. Agentic candidate design — AI agents propose ~1,000 candidate sequences per round.
2. Pooled oligonucleotide array synthesis (IDT oPools).
3. Multiplexed bead-based affinity pulldown against target proteins.
4. Next-generation sequencing (Illumina MiSeq, 1×52 bp single-end reads).
5. Return of count matrices and metadata to agents for iterative redesign.

Agentic Design Platforms

• Round 1: Google's AI Co-Scientist (general-purpose multi-agent system) generated initial candidates from zero experimental data.
• Rounds 2–3: A domain-specialized four-stage pipeline with: (1) analyst agent (independent DESeq2 differential enrichment, bare-bead regression, failure diagnostics), (2) strategist agent (explore-vs-exploit allocation), (3) parallel designer agents (sequence generation with folding/filtering), and (4) a scoring stage (3D structure prediction with Boltz-2, ESMFold2-NA).

Computational Tools Provided to Agents

Statistical Hit-Calling Pipeline

A two-stage statistical pipeline was applied uniformly across rounds:

1. Stage 1 (DESeq2 differential enrichment): Raw replicate counts normalized via median-of-ratios size factors, negative binomial dispersion estimation, Wald statistics with Benjamini–Hochberg FDR adjustment to obtain target-versus-input enrichment $y_{i,t} = \log_2\text{FC}_{i,\text{target vs. input}}$ and bead-versus-input enrichment $x_i = \log_2\text{FC}_{i,\text{beads vs. input}}$.

2. Stage 2 (OLS bare-bead background regression): An ordinary least-squares linear model was fit:

$$\hat{y}_{i,t} = \beta_{0,t} + \beta_{1,t} x_i$$

with bead-corrected residual $r_{i,t} = y_{i,t} - \hat{y}_{i,t}$. A construct was called a hit if padj < 0.05, $\log_2\text{FC} \geq 1.0$, and $r_{i,t} > 1.00$.

Targets

Three proteins spanning a difficulty gradient:

---

Empirical Validation / Results

R.2 Hit Rate Progression

Across three rounds, statistically significant, bead-corrected on-target hit rates among newly designed candidates:

All random-sequence negative controls remained at 0.0% across every round.

R.2 Zero-Shot Predictor Benchmark

On the Round 3 myoglobin library (N = 1,441), 12 open-license predictor metrics performed near or below random chance:

• Boltz-2 iPTM: ROC-AUC = 0.499; ΔiPTM: 0.391
• ESMFold2-NA, RhoDesign, RiboDiffusion: 0.496–0.510
• DNABERT-2, HyenaDNA, RNA-FM: 0.364–0.523
• G4Hunter: 0.366; ViennaRNA ΔΔG G4: 0.476
• ViennaRNA −ΔG_MFE: 0.729; −ED (folding purity): 0.763

Only target-agnostic 2D thermodynamic stability showed meaningful signal, but it has no knowledge of the protein target.

R.3 G-Quadruplex Specificity Collapse and Recovery

• Round 2 myoglobin collapse: 80.9% (72/89) of binders in the counter-screened benchmark were cross-reactive; on-target specificity crashed from 75.0% → 19.1%.
• Autonomous diagnosis: The analyst agent found that 97% of cross-reactive binders were G-quadruplexes acting as non-specific polyanionic stickers.
• Round 3 recovery: Agents cut unclamped G4s from 20.4% → 1.3% and pivoted 72.3% of the pool into non-G4 scaffolds:
  • H-type pseudoknots: 80.4% hit rate (37/46)
  • Extended palindromic hairpins: 61.8% (47/76)
  • Bulge hairpins: 45.9% (67/146)
  • Specificity recovered to 72.4%

R.5 Single-Clone Validation

Eight 5'-FAM-labeled full-length α-thrombin candidates (T01–T08) were individually synthesized and titrated (0–1,000 nM):

• Endpoint fluorescence at 1 μM scaled quantitatively with pooled enrichment (Pearson r = 0.90, p = 0.0021; Spearman ρ = 0.88, p = 0.0038).
• Top-enriched candidates (T01–T04): preliminary apparent K_D = 24.6–133 nM.
• Moderately enriched de novo scaffolds (T05–T08): K_D > 1 μM (micromolar onset).

The tightest binder, T01 (NU172-derived G4–duplex chimera with C10A bulge and 9-bp GC-rich stem clamp), achieved preliminary apparent K_D = 24.6 ± 8.2 nM.

---

Theoretical and Practical Implications

Value of Empirical Feedback

The paper demonstrates that empirical lab data is much more valuable than existing computational proxies. Key evidence:

• 3D co-folding models (Boltz-2) anticorrelated with specific binding — the model rewarded compact G-quadruplex folds that adhered non-specifically to solid supports.
• A supervised classifier trained on early rounds (before bead-background subtraction) learned G-tract frequency as a positive feature and inverted its accuracy on subsequent pools — highlighting the hazard of active learning on confounded data.

Iterative Layering of Selective Pressures

The closed-loop paradigm allows each cycle to layer on a new selective pressure:

1. First optimize for target affinity
2. Then for counter-screened specificity
3. Future rounds: activity in complex biological matrices (serum, cell lysate, cellular uptake)

Generalizability

The principles demonstrated are modality-agnostic. The three-step wet-lab workflow can be mapped onto robotic liquid handlers, enabling self-driving loops that co-optimize sequences, buffers, cofactors, and stringency. This sets the stage for extending multi-round agentic discovery to protein–protein binder design.

Limitations and Caveats

• K_D values are preliminary (fluorescence did not reach saturation plateau within 1 μM).
• Commercial licensing restricted computational evaluations to open-license tools.
• No controlled comparison between agent architectures (AI Co-Scientist in R1 vs. custom harness in R2–3).
• Fine discrimination against close structural homologues (e.g., myoglobin vs. hemoglobin) not tested.

---

Conclusion

This work establishes the feasibility of closed-loop, agent-guided biomolecular discovery across multi-round experimental campaigns. Key takeaways:

1. LLM-based agents can learn iteratively from wet-lab data, diagnosing failures and refining designs across rounds without prior target-specific training data.
2. The aptamer sandbox (1–2 week cycles, $1–$3 per candidate) provides an ideal testbed for lab-in-the-loop paradigms.
3. Empirical feedback outperforms computational proxies — all 12 zero-shot predictors performed near or below random chance.
4. Agents reasoned across the entire experimental stack: structural hypotheses, statistical counter-selection filters, wet-lab protocol suggestions, and synthesis/readout constraints.

Future Directions

• Prospective trials with matched-budget no-feedback and non-reasoning control arms to isolate the contribution of agentic reasoning.
• Head-to-head benchmarking against evolutionary, Bayesian, and supervised ML baselines.
• Extending single-clone validation to truncated minimal cores, additional affinity-maturation rounds, and label-free kinetic assays (BLI, SPR).
• Co-optimization of sequences and assay protocols in future loops.
• Extension to more complex modalities such as protein–protein binder design, leveraging the closed-loop principles demonstrated here.