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

Figure 1 | Aptamers as a fast, high-throughput sandbox for closed-loop agentic molecular discovery. (A) Wet-lab steps per round from candidate to binding readout across modalities. Candidates designed by AI agents proceed directly from pooled DNA array synthesis to a multiplexed magnetic-bead pulldown and sequencing readout in 3 bench steps (

\sim

10 days per round), as an aptamer’s nucleotide sequence is itself the functional molecule. This direct readout bypasses the 6-step per-round selection, amplification, and

\lambda

-exonuclease ssDNA regeneration cycle of conventional SELEX (repeated over 8–15 rounds before sequencing readout) and the 7-step cloning, expression, purification, and binding-assay cascade of protein/antibody pipelines. (B) Closed-loop agentic discovery pipeline (

\times N

rounds). In each cycle, the agentic AI pipeline ( Analyze , Generate , Design , Rank , transitioning from AI Co-Scientist zero-shot ideation in Round 1 to a domain-specialized tool-equipped agent harness in Rounds 2–3) designs

\sim

1,000 aptamer candidates for pooled oligo synthesis, which are assayed via multiplexed bead pulldown and next-generation sequencing (NGS) to return quantitative feedback for the next design round. (C) Schematic of the representative data and diagnostic package after each round, comprising demultiplexed FASTQ sequences, per-sequence read count matrices, and sample metadata returned from the wet-lab screen alongside the analyst agent’s independently derived DESeq2 differential enrichment profiles, replicate/cross-target correlation matrices, and target-versus-bead

\log_{2}

fold-change (LFC) regression diagnostics used to refine subsequent designs.

Figure 2 | Closed-loop progression of binder hit rate, population enrichment, and zero-shot predictor benchmark across three rounds (R1

\rightarrow

R3). (A) Surface structures and architecture of the benchmark targets: human myoglobin (left; PDB 3RGK, de novo globular target) and human

\alpha

-thrombin (right; PDB 1PPB, positive-control benchmark with electropositive Exosites I and II). (B) Statistically significant, bead-corrected on-target binder hit rates across Rounds 1–3 for newly designed candidates on

\alpha

-thrombin (

0.0\%\rightarrow 26.0\%\rightarrow 42.1\%

) and myoglobin (

10.2\%\rightarrow 3.7\%\rightarrow 33.5\%

), alongside random-sequence negative controls (

0.0\%

, dashed line; full cohort and strictly de novo breakdowns in Supplementary Table S3 ). (C) Kernel density distributions of raw target enrichment (

\log_{2}\text{FC}

over input prior to bare-bead regression) across Rounds 1–3 for myoglobin (left) and

\alpha

-thrombin (right), stacked directly below B . Vertical dashed lines mark

\log_{2}\text{FC}=1.0

, dotted lines mark population medians, and dark-shaded tails denote the fraction with

\log_{2}\text{FC}\geq 1.0

. (D) Zero-shot receiver operating characteristic area under the curve (

\text{ROC-AUC}\pm 95\%\text{ CI}

) across 12 open-license computational predictor metrics (spanning 9 tools) evaluated on the Round 3 myoglobin library (

N=1{,}441

; 292 hits [20.3%]). The vertical dashed line marks random chance (

\text{ROC-AUC}=0.50

).

Figure 3 | Autonomous diagnosis and recovery from G-quadruplex specificity collapse via target-matched structural motif reallocation (R1

\rightarrow

R3). (A) Target-specific failure modes and autonomous specificity recovery across Rounds 1–3, showing on-target specificity (%; coral curves, left axis; percentage of statistically significant binders in the counter-screened benchmark that bind exclusively to the intended target without off-target cross-reactivity) alongside newly designed on-target binder hit rate (%; teal curves, right axis) for myoglobin (left; specificity

75.0\%\rightarrow 19.1\%\rightarrow 72.4\%

, de novo hit rate

10.2\%\rightarrow 3.7\%\rightarrow 33.5\%

) and

\alpha

-thrombin (right; specificity

0.0\%\rightarrow 76.7\%\rightarrow 98.5\%

, newly designed hit rate

0.0\%\rightarrow 26.0\%\rightarrow 42.1\%

; full denominators in Supplementary Table S3 ). (B) Alluvial flow diagrams showing the reallocation of candidate structural motif families from Round 1 to Round 3 for myoglobin (left) and

\alpha

-thrombin (right), aligned column-for-column with A and paired with horizontal bar charts of Round 3 on-target hit rates (

n/N

) and predicted secondary-structure schematics for the top-performing families.

Reasoning Category & Summary	Representative Agent Quotes & Documented Decision Chain
1. Explored atypical structural hypotheses Conducted de novo explorations outside traditional single-loop hairpins and canonical motifs—reasoning directly about 3D scaffold geometry, inter-arm angles, directional bulges, and bipartite pockets to match target surface topology.	"Wide-hinge 44-nt three-way junction scaffold featuring dual 2-nt flexible single-stranded junction spacers. The flexible hinge geometry allows an expanded inter-arm angle (85–110 ∘ ) to wrap around the globular curvature of myoglobin without steric clash." "Dual-asymmetric internal loop (1-nt 5’ vs. 4-nt 3’ bulge) with an 8-bp GC basal clamp, creating an acute directional bend that positions the CAA apical loop and bulge into the myoglobin heme cleft." "13-bp duplex… project[s] the unstacked purine core ( TGAAAT , positions 14–19) into solvent. The exocyclic faces of A16 and A17 are positioned to establish CH– $\pi$ and edge-to-face $\pi$ – $\pi$ contacts against the exposed aromatic ring of Phe43… while the phosphodiester backbone at the T13:A20 neck forms electrostatic salt bridges with wild-type Lys45 (Arg45 in PDB 3RGK)." "Two interleaved stem-loops (H-type pseudoknot) creating a rigid, compact 3D scaffold with defined bipartite helical binding cavities."
2. Leveraged statistical analysis and physics tools to guide design (Documented Round 2 $\rightarrow$ Round 3 Decision Chain) Built statistical counter-selection filters from raw counts to separate target-specific binding from magnetic-bead matrix noise, diagnosed G-quadruplex promiscuity and 3D co-folding failure, and shifted the synthesis portfolio toward non-G4 scaffolds prior to Round 3 screening.	Observed Failure [Round 2 Screen] $\rightarrow$ Autonomous Diagnosis [Between-Round Analysis] : In Round 2, myoglobin bead-corrected on-target hit rate dropped from 10.2% ( 20/196 ) in Round 1 to 3.7% ( 9/244 ) among newly designed myoglobin candidates ( 9.3% [17/183] in the counter-screened benchmark including 8/16 carried-over R1 positive spikes; 9.0% [88/977] non-spike across the full pool; 80.9% [72/89] of binders cross-reactive, with 13.9% [72/517] off-target binding across all non-myoglobin designs). From raw count matrices, the analyst agent found that "The $R^{2}$ of 0.635 means 63.5% of total enrichment variance is explained by non-specific bead affinity… Bead regression removes 85% of false positives" (in its initial QC-filtered analysis) and diagnosed that "thrombin-designed G-quads bind myoglobin at 18.7% yield — double the 9.3% yield of myoglobin-designed sequences… NONE of the top-10 iPTM sequences overlap with actual binders… Using Boltz2 iptm as a design filter is actively harmful — it selects for non-specificity." Catching when linear regression breaks down [Between-Round Analysis] : "Bead-binding regression is NOT appropriate for this [Thrombin] experiment — not merely because $R^{2}$ is low (0.198), but because the regression slope is negative ( $-$ 0.263), which causes regression to inflate false positives rather than remove them… Flag a binder if LFC_beads / LFC_thrombin > 0.5." Proposed Change [Round 3 Strategy] $\rightarrow$ Implemented in Synthesis Order [Round 3 Pool] : Capped pure/unclamped G-quadruplexes at $\leq 15\%$ of the myoglobin design allocation (reducing unclamped naked G4s to 1.3% [7/549]), shifted 72.3% of the portfolio into non-G4 hairpins, pseudoknots, and three-way junctions (plus 26.4% clamped stem–G4 chimeras), calibrated ViennaRNA 2D folding ( "Replace [3D co-folding] with ViennaRNA local folding in buffer conditions… actual binders are less stable [median MFE $-$ 5.1 vs $-$ 9.3 kcal/mol, p = 0.0007]" ), and embedded prospective Boltz-2 calibration bands ( Tiers A--D ) while humans held the wet-lab buffer constant. Subsequent Result [Round 3 Screen] : Myoglobin on-target hit rate among newly designed candidates surged $9.1\times$ from 3.7% ( 9/244 ) to 33.5% ( 184/549 ; $3.6\times$ vs. the 9.3% [17/183] R2 benchmark and $3.3\times$ vs. R1 [ 10.2% , 20/196], reaching 45.9%--80.4% across non-G4 hairpins and pseudoknots vs. 2.4% [3/124] for clamped 3-tetrad G4 chimeras and 0.0% [0/7] for unclamped G4s), specificity recovered from 19.1% to 72.4% , and prospective Tier A ( Boltz-2 $\text{iPTM}>0.7$ ) yielded 0/25 (0.0%) statistically significant hits.
3. Suggested tweaks to wet-lab protocols Proposed concrete adjustments to liquid-handling automation, buffer ions, wash stringency, refolding, and control pooling (logged as autonomous protocol recommendations; wet-lab buffer held constant across Rounds 1–3) .	"Implement automated magnetic bead washing protocols on robotic liquid handlers (e.g., Hamilton Star / Opentrons OT-2)… Standardize wash volume, agitation cycles, and temperature using automated liquid handlers. Automated wash execution directly prevents outlier replicates." "Our buffer is 1 $\times$ PBS (4.5 mM $\mathrm{K}^{+}$ ); Thrombin TBA was selected in 25–100 mM KCl (our $\mathrm{K}^{+}$ is 5–20 $\times$ too low)… Supplement binding buffer with 20 mM KCl." "Swap standard $\mathrm{K}^{+}$ and $\mathrm{Na}^{+}$ cations for $\mathrm{Li}^{+}$ during positive selection rounds to actively destabilize G-quadruplex folding, forcing enrichment of Watson-Crick-stabilized specific binders." "Increase stringency in magnetic bead washing (salt concentration, surfactant, and wash cycle count) and add 0.1 mg/mL salmon sperm DNA." "Switch aptamer refolding from bench snap-cool to slow ramp-down on a PCR cycler and increase binding incubation from 2 min on ice to 30 min at room temperature." "Evaluate positive controls strictly in isolated single-oligo lanes or spike at $\sim$ 1/100 so the candidate library remains measurable."
4. Reasoned across design, synthesis, and experimental readout Connected in silico sequence generation to physical constraints across solid-phase DNA synthesis, full-construct primer folding, and PCR amplification to prevent library dropouts and misfolding.	Eliminating extreme palindromes to prevent library dropout : "Of the wells that got zero reads, 96% are self-complementary ( $\Delta G$ of $-$ 35 to $-$ 47 kcal/mol, Fisher odds ratio $\approx$ 77, p = 5 $\times$ $10^{-165}$ )… During base-by-base phosphoramidite chemistry at IDT, a strand that folds back on itself blocks the coupling step. In the regenerated R2 pool, high self-complementarity = 0 [dropout fell from 18.8% to 1.8%]." 85-nt full-construct primer invasion : "Aptamer binding depends on the full oligo fold (85 nt), not just the 40-nt variable region. Fixed primer handles form competitive secondary structures ( $\Delta G$ $\approx$ $-$ 6 to $-$ 12 kcal/mol) that destroy the aptamer’s binding fold." Spacer length & palindromic penalty : "Adding 7-nt TTTCGCG / CGCGTTT insulation spacers reduced sequencing read counts 2.7 $\times$ to 5.9 $\times$ by extending total oligo length and introducing palindromic CGCG motifs." PCR stalling vs. synthesis dropout : "Only 1.1% of G4 sequences had zero input reads (synthesis failure); the remaining 74.0% were present at low abundance due to PCR amplification bias. Without the valley QC filter, these sequences are no longer discarded — DESeq2’s negative binomial model accounts for them statistically."
5. Evolved individual candidates and scaffolds across rounds Traced promising candidate lineages across selection rounds—diagnosing structural vulnerabilities in early seeds and re-engineering their flanking stems and recognition loops to improve affinity and specificity.	Round 1 $\rightarrow$ Round 2 (Diagnosing primer invasion & recovering active cores) : Uninsulated R1 $\alpha$ -thrombin cores failed ( 0.0% hits) because fixed PCR primers hybridized with the binding motif; modeling full 85-nt constructs in R2 recovered the canonical 15-mer TBA core ( $\log_{2}\text{FC}=+4.84$ ) and identified top G-quadruplex–hairpin hybrid seed 095e2a34c9a4 (embedding the 25-nt core of canonical NU172 ; $\log_{2}\text{FC}=+12.33$ ). Round 2 $\rightarrow$ Round 3 (Locking canonical TBA cores into Watson–Crick stem-clamps) : Engineered 8–12 bp flanking duplex clamps around the R2 TBA core ( tba_hairpin_scaffold ) to prevent primer invasion and reduce bare-bead stickiness; in the subsequent Round 3 screen, all 22/22 ( 100% ) sequenced tba_hairpin_scaffold constructs were statistically significant, bead-corrected hits ( $\log_{2}\text{FC}=+8.68\text{ to }+13.11$ , comprising 5/5 short-stem TBA-Stem and 17/17 long-stem Extended Palindromic Hairpin constructs; 25/25 [ 100% ] across all R3 TBA / NU172 -core designs), with representative 8-bp and 11-bp stem-clamped and loop-edited leads ( T02 – T04 ) validating in single-clone titrations at preliminary apparent $K_{D}=47.4\text{--}133\text{ nM}$ ( Fig. 4 ). Round 2 $\rightarrow$ Round 3 (Prospective mutational maturation of top hybrid seed) : Generated systematic mutational variants ( R3_095_var1 – var10 ) around Round 2 champion 095e2a34c9a4 ( $\log_{2}\text{FC}=+12.33$ ), introducing a 9-bp GC-rich terminal stem ( 5’-TAACCAGCG...CGCTGGTTA-3’ ) and a C10 $\rightarrow$ A strain-relieving bulge transition ( R3_095_var3 ). In the subsequent Round 3 screen, R3_095_var3 ( T01 ) became the #1 clone in the library ( 44.55% of thrombin reads, $\log_{2}\text{FC}=+16.27$ , $25.2\times$ specificity over bare beads) and validated in single-clone equilibrium titrations at a preliminary apparent $K_{D}=24.6\pm 8.2\text{ nM}$ ( $7.2\times$ fluorescence over background; Fig. 4 ).

Table 4

Figure 4 | Orthogonal single-clone bead-binding validation of

\alpha

-thrombin aptamer candidates ( T01 – T08 ). (A) Normalized equilibrium binding curves (

(\text{RFU}-y_{0})/\Delta B_{\max}

) for the four nanomolar

\alpha

-thrombin binders ( T01 – T04 ) titrated across

1

,

10

,

100

, and

1{,}000\text{ nM}

5’-FAM-labeled full-length aptamer against biotinylated human

\alpha

-thrombin immobilized on streptavidin magnetic beads (

1\text{:}50

bead dilution). Curves show non-linear least-squares 1:1 hyperbolic Langmuir fits (

\text{RFU}(C)=y_{0}+\Delta B_{\max}\cdot C/(K_{D}+C)

); T04 (preliminary

K_{D}=132.5\text{ nM}

, dashed purple) is dashed to distinguish its curve from T03 (preliminary

K_{D}=132.7\text{ nM}

, solid orange). The horizontal dashed line indicates

50\%

of fitted

\Delta B_{\max}

(apparent

K_{D}

), and the shaded vertical band spans the preliminary nanomolar

K_{D}

range (

24.6\text{--}133\text{ nM}

; treated as preliminary estimates as fluorescence did not reach an upper plateau within

1{,}000\text{ nM}

). (B) Blank-subtracted endpoint fluorescence at

1\ \mu\text{M}

(

\Delta\text{RFU}=\text{RFU}_{1000\text{ nM}}-\text{RFU}_{0\text{ nM}}

) across all eight individually synthesized full-length candidates ( T01 – T08 , 75–85 nt), color-coded by binding regime: nanomolar binders ( T01 – T04 , teal), micromolar-onset de novo binders ( T05 – T06 , gold), and candidates below the

1\ \mu\text{M}

titration ceiling ( T07 – T08 , grey). (C) Quantitative concordance between Round 3 pooled sequencing enrichment (

\log_{2}\text{FC}

) and orthogonal single-clone fluorescence at

1\ \mu\text{M}

(

\Delta\text{RFU}

,

\log_{10}

scale) across T01 – T08 (Pearson

r=0.90

,

p=0.0021

). The vertical dashed line marks the enrichment threshold (

\log_{2}\text{FC}>9

) separating candidates with preliminary nanomolar

K_{D}

( T01 – T04 ) from moderate-enrichment candidates in the micromolar regime ( T05 – T08 ). Full sequence details, raw RFU titrations, and curve-fitting parameters under both fixed-blank and free-baseline models are provided in Supplementary Table S4 .

Supplementary Figure S1 | Target-versus-streptavidin-bead enrichment regression and two-stage hit calling across design rounds (R1

\rightarrow

R3). Target enrichment (

\log_{2}\text{FC}_{\text{target vs. input}}

,

y

-axis) plotted against bare streptavidin-bead enrichment (

\log_{2}\text{FC}_{\text{beads vs. input}}

,

x

-axis) for

\alpha

-thrombin ( A–C ) and myoglobin ( D–F ) across Round 1 (

N=642

; A :

0/167

[

0.0\%

] new

\alpha

-thrombin designs; D :

20/196

[

10.2\%

] new myoglobin designs [

37/642=5.8\%

full pool]), Round 2 (

N=981

; B :

54/208

[

26.0\%

] new

\alpha

-thrombin designs [

73/981

total hits]; E :

9/244

[

3.7\%

] new myoglobin designs [

96/977

total hits including

8/16

carried-over positive spikes]), and Round 3 (

N=1{,}405

deduplicated constructs [

1{,}441

total library entries]; C :

131/311

[

42.1\%

] new

\alpha

-thrombin designs [

162/1{,}405

full pool]; F :

184/549

[

33.5\%

] new myoglobin designs [

273/1{,}405

full pool]; complete cohort breakdowns in Supplementary Table S3 ). Dashed lines indicate the ordinary least-squares bare-bead background trend (slope

\beta_{1}

); solid boundaries mark the specific binder zone requiring both

\log_{2}\text{FC}_{\text{target}}\geq 1.0

(FDR < 0.05) and positive residual enrichment

r_{i,t}>1.00

above the bare-bead regression line (plus

x_{i}<0.5

in B and

y_{i,\text{thr}}-x_{i}>1.0

for bare-bead pre-hits in C ). Points denote statistically significant

\alpha

-thrombin binders (blue circles), myoglobin binders (teal circles), rejected bead-stickers (coral circles, including 217 G4 designs in E with

\beta_{1}=0.60

), non-binders (grey dots), negative/scrambled/random controls (open triangles,

\triangle

), published and prior-round positive spike-in controls (gold stars,

\star

), and rejected non-specific sticky controls (red diamonds,

\diamond

in E–F ).

Supplementary Figure S2 | DESeq2 differential enrichment volcano plots and control validation across design rounds (R1

\rightarrow

R3). Statistical significance (

-\log_{10}\text{FDR}

,

y

-axis) versus target enrichment (

\log_{2}\text{FC}_{\text{target vs. input}}

,

x

-axis) for

\alpha

-thrombin ( A–C ) and myoglobin ( D–F ) across Rounds 1–3, with dashed lines marking the pre-hit cutoffs (

\log_{2}\text{FC}=1.0

and FDR = 0.05; upper-right quadrant). Apparent volcano pre-hits are decomposed into statistically significant on-target binders versus rejected bead-stickers eliminated by the bare-bead regression gate in Supplementary Figure S1 . Stars (

\star

) mark published positive controls (uninsulated TBA stem in A ; insulated thrombin controls TBA/ thrombin_3 and thrombin_2 in B–C ) and internal prior-round positive spike-in controls discovered in Round 1 ( D ) and carried forward into Rounds 2–3 ( E–F ). Open triangles (

\triangle

) mark random and scrambled negative controls (0% false positives across all rounds). Dotted horizontal lines in B and C denote the display cap (

-\log_{10}\text{FDR}\geq 45

).

Supplementary Figure S3 | Counter-screen discovery and continuation of HER2-selective binders across Rounds 1–3. (A) Structure of the human HER2 extracellular domain (ECD; subdomains I–IV,

630\text{ aa}

,

\sim 70\text{ kDa}

, PDB 1N8Z), introduced in Round 2 as a specificity counter-screen and continued in the Round 3 multiplexed panel alongside myoglobin,

\alpha

-thrombin, and bare streptavidin beads. (B) Statistically significant HER2-selective binder rate across Round 1 (screened on myoglobin,

\alpha

-thrombin, and bare streptavidin beads; HER2 not yet screened), Round 2 (

N=968

), and Round 3 (

N=1{,}405

). Strictly HER2-selective binders (

\log_{2}\text{FC}\geq 1.0

, bead-corrected enrichment

\geq 1.0

,

\text{FDR}<0.05

, with no significant enrichment on myoglobin or

\alpha

-thrombin) increased

3.8

-fold from 1.3% (13 hits) in Round 2 to 5.1% (71 hits) in Round 3, while negative controls remained at 0.0% across all rounds (dashed line). (C) Kernel density distributions of bead-corrected HER2 enrichment (

\log_{2}\text{FC}

) across Rounds 1–3. The vertical dashed line marks the

\log_{2}\text{FC}=1.0

threshold; dark-shaded tails denote statistically significant selective binders (

\geq 1.0\ \log_{2}\text{FC}

), expanding from 1.3% (max

+1.9\ \log_{2}\text{FC}

) in Round 2 to 5.1% (reaching

+4.6\ \log_{2}\text{FC}

) in Round 3. Individual dots in the Round 3 right tail mark top selective binders (

\geq 3.5\ \log_{2}\text{FC}

), including the carried-over Round 2

\rightarrow

Round 3 binder (

\star

). (D) Round 3 volcano plot of bead-corrected HER2 enrichment (

\log_{2}\text{FC}

) versus statistical significance (

-\log_{10}\text{FDR}

) across all

N=1{,}405

candidates, highlighting the 71 HER2-selective binders (solid circles, up to

+4.59\ \log_{2}\text{FC}

or

24\times

over bare beads), the carried-over Round 2

\rightarrow

Round 3 binder (gold star

\star

,

+4.28\ \log_{2}\text{FC}

,

\text{FDR}=0.031

), and negative/random controls (open triangles,

0/24

false positives).

Target	Round & Lineage	Structural Family	Variable Region / Conserved Motif ( 5’ $\rightarrow$ 3’ )	Folding & Enrichment Metrics	Structure–Activity Hypothesis
$\alpha$ -Thrombin	Round 1 $\rightarrow$ Round 2 Primer-Insulation Fix ( T02 – T04 )	Uninsulated vs. stem-clamped 2-tetrad G4 ( TBA core)	Core: GGTTGGTGTGGTTGG (flanked by fixed primers in R1 vs. 6–11 bp duplex clamps in R2–R3)	R1 full-construct primer–core pairing $\Delta G=-7.8\text{ kcal/mol}$ (0.0% hits) $\rightarrow$ R3 clamped stem 100% hit rate (7/7 canonical 6–10 bp TBA clamps; 100% [25/25] across all R3 clamped TBA / NU172 -core constructs); preliminary single-clone $K_{D}=47.4\pm 14.6\text{ nM}$ ( T02 , 11-bp stem) vs. $132.7\text{ nM}$ ( T03 , 8-bp stem) and $132.5\pm 52.8\text{ nM}$ ( T04 , G8A loop edit; Fig. 4 )	Folding the full 85-nt primer-flanked construct in ViennaRNA indicated that fixed primers could hybridize with uninsulated cores; adding protective Watson–Crick stems restored predicted folding and nanomolar physical binding.
$\alpha$ -Thrombin	Round 2 G4–Hairpin Hybrid Seed ( NU172 core)	G-quadruplex–hairpin hybrid motif	Conserved 25-nt motif: GCCTAGGTTGGGTAGGGTGGTGGCG (25-nt core of canonical 26-nt NU172 )	Shared across 17 of top 18 R2 binders ( $\log_{2}\text{FC}>11$ )	Embeds the 25-nt G-quadruplex/duplex core of NU172 within a flanking Watson–Crick stem clamp that insulates the motif from primer interference.
$\alpha$ -Thrombin	Round 3 Top Enriched Binder ( R3_095_var3 / T01 )	Extended palindromic hairpin / clamped G4 hybrid	TAACCAGCGACTAGGTTGGGTAGGGTGGTGGCGCTGGTTA	$\log_{2}\text{FC}=+16.27$ (strongest binder of campaign, $44.55\%$ of R3 pool); preliminary single-clone $K_{D}=24.6\pm 8.2\text{ nM}$ ( $\Delta\text{RFU}=+1{,}825$ , $7.2\times$ over blank; Fig. 4 ); family hit rate 57.5% (42/73)	Extending the Round 2 25-nt hybrid seed ( 095e2a34c9a4 ) with a 9-bp terminal stem ( TAACCAGCG...CGCTGGTTA ) and a C10 $\rightarrow$ A strain-relieving bulge maximized predicted folding purity, pooled enrichment, and single-clone nanomolar affinity.
Myoglobin	Round 2 Basal-Stem vs. Loop Contrast	Stem-loop hairpin scaffold	Shared 14-bp basal stem: GTCGAAGAGTTCAC (paired with GTGAACTCTTCGAC )	72/74 stem-matched designs failed; 2/74 enriched strongly ( $\log_{2}\text{FC}=+2.49,+2.39$ ; $\Delta G_{\text{MFE}}<-10\text{ kcal/mol}$ )	Comparing 74 sequences with identical basal stems indicated that stem stability alone is insufficient for binding and that enrichment depends on the 10–13 nt apical/internal loop sequence.
Myoglobin	Round 3 Top Pseudoknot Lineage	H-type pseudoknot ( -CGCG- tertiary lock)	GTCGAAGAGTTCACACGCGATGTGAACTCTTCGACATCGT	$\Delta G_{\text{MFE}}=-24.9\text{ kcal/mol}$ ; $\log_{2}\text{FC}=+3.61$ ; family hit rate 80.4% (37/46)	Evolving the Round 2 stem-loop winner via constrained inverse folding into a predicted -CGCG- H-type pseudoknot stabilized loop presentation while eliminating G4 cross-reactivity.
Myoglobin	Round 3 Non-G4 Hairpin Families	Extended palindromic & bulge/pocket hairpins	Variable 8–12 bp stems interrupted by asymmetric internal bulges	Extended palindromic hit rate 61.8% (47/76); bulge/pocket hairpin hit rate 45.9% (67/146)	Replacing polyanionic G-quadruplexes (clamped 3-tetrad G4 chimeras: 2.4% hit rate, 3/124; unclamped G4s: 0.0% hit rate, 0/7) with non-G4 hairpins and 3WJs restored on-target specificity from 19.1% to 72.4%.

Table 9

Candidate ID	Round Origin & Architecture	Variable Region Sequence ( 5’ $\rightarrow$ 3’ , 40–42 nt)	Mean Read Counts ( Input / Beads / HER2 / Myo / Thr )	R3 HER2 Raw $\log_{2}\text{FC}$	R3 Bead-Corrected $\log_{2}\text{FC}$	R3 HER2 FDR	Myoglobin & $\alpha$ -Thrombin Off-Target Enrichment
hybrid_chimera_008	Round 3 ( 8-bp stem + G4 chimera )	GACGCCAATGGGTGGGGTTGTTTAGGGAGGGATTGGCGTC	233 / 294 / 7,214 / 223 / 804	$+4.95$	$+4.59$	$0.031$	No significant enrichment (Myo $+0.25$ , Thr $+0.34$ )
constrained_gquad_005	Round 3 ( 8-bp stem + G4 chimera )	CGCCACCATTAGGGTGGGAATAGGGAGGGAATTGGTGGCG	242 / 177 / 4,773 / 207 / 157	$+4.30$	$+4.47$	$0.034$	No significant enrichment (Myo $+0.43$ , Thr $-0.18$ )
bef5bce6520a ( $\star$ )	Round 2 $\rightarrow$ Round 3 carryover ( 6-bp stem + G4 )	AAGAGCGGGGTGGGTTTTCTTTGGGTTGGGCCGCTCAAAA	258 / 504 / 8,657 / 285 / 311	$+5.06$	$+4.28$	$0.031$	No significant enrichment (Myo $+0.18$ , Thr $-0.69$ )
R1_0130	Round 1 $\rightarrow$ Round 3 carryover ( G4--hairpin hybrid )	GGTGGTGGGTTGTGGGTGGTTGATTTTTCTAATGGTCATGAG	343 / 314 / 4,747 / 208 / 560	$+3.79$	$+3.74$	$0.037$	No significant enrichment (Myo $-0.21$ , Thr $+0.69$ )
constrained_gquad_009	Round 3 ( 8-bp GC stem + G4 chimera )	GCTCGCGCGGGTAAGGGTTTTTTGGGTAAGGGGCGCGAGC	160 / 273 / 2,645 / 187 / 244	$+4.04$	$+3.39$	$0.034$	No significant enrichment (Myo $+0.35$ , Thr $-0.16$ )
constrained_gquad_002_v3	Round 3 ( 8-bp stem + G4 chimera )	GACGACGAGGGTAAGGGTTTTTTGGGTAAGGGTCGTCGTC	219 / 196 / 2,075 / 183 / 186	$+3.24$	$+3.22$	$0.043$	No significant enrichment (Myo $+0.27$ , Thr $-0.07$ )

Table 10

Round	Target Protein	Counter-Targets & Matrix Controls Screened	Full Sequenced Pool ( $N_{\text{pool}}$ )	Cohort Breakdown ( New Target Designs / Off-Target Sub-Pools / Variants & Tiers / Controls & Spikes )	Statistical Hit Inclusion Rule	Raw Enriched ( $\log_{2}\text{FC}\geq 1.0$ )	Volcano Pre-Hits ( $\log_{2}\text{FC}\geq 1.0$ , $\text{FDR}<0.05$ )	Statistically Significant On-Target Hits: New Target-Designed Candidates ( $n/N,\%$ )	On-Target Hits: Strictly De Novo Subset ( $n/N,\%$ )	On-Target Hits: Full Sequenced Pool ( $n/N,\%$ )	Negative / Scrambled Controls ( $n/N,\%$ )
Round 1	$\alpha$ -Thrombin	Myoglobin, insulin, bare streptavidin beads	642	167 Thr new / 475 off-target ( 196 Myo, 279 Ins) / 0 var / 0 ext ctrl	DESeq2 $\log_{2}\text{FC}\geq 1.0$ , $\text{FDR}<0.05$ vs. input + beads	1 / 642 (0.2%)	0 / 642 (0.0%)	0 / 167 (0.0%)	0 / 57 (0.0%) (0 / 110 lit)	0 / 642 (0.0%)	0 / 279 Ins decoy (0.0%)
Round 1	Myoglobin	$\alpha$ -Thrombin, insulin, bare streptavidin beads	642	196 Myo new / 446 off-target ( 167 Thr, 279 Ins) / 0 var / 0 ext ctrl	DESeq2 $\log_{2}\text{FC}\geq 1.0$ , $\text{FDR}<0.05$ vs. input + beads (37 at $\text{FDR}<0.05$ ; 24 also $\geq 1.0\ \log_{2}\text{FC}$ )	42 / 642 (6.5%)	37 / 642 (5.8%) (24 at $\geq 1.0$ )	20 / 196 (10.2%) (18 at $\geq 1.0$ )	19 / 171 (11.1%) (1 / 25 lit; 34 / 499 [6.8%] full pool)	37 / 642 (5.8%) (24 / 642 [3.7%] at $\geq 1.0$ )	0 / 53 random (0.0%)
Round 1b *	Myoglobin (Single-target expansion)	Bare streptavidin beads	1,341	1,288 Myo new / 0 off-target / 0 var / 53 scrambled neg	$\log_{2}\text{FC}\geq 1.0$ , $\text{FDR}<0.05$ , OLS bead residual $>+1.00$	166 / 1,341 (12.4%)	148 / 1,341 (11.0%)	86 / 1,288 (6.7%)	86 / 1,288 (6.7%)	86 / 1,341 (6.4%)	0 / 53 (0.0%)
Round 2	$\alpha$ -Thrombin	Myoglobin, HER2, TNF- $\alpha$ , bare streptavidin beads	981	208 Thr new / 683 off-target ( 243 Myo, 229 HER2, 211 TNF) / 0 var / 90 ctrls ( 14 lit, 16 R1 spikes, 60 scrambles)	$\log_{2}\text{FC}\geq 1.0$ , $\text{FDR}<0.05$ , OLS bead residual $>+1.00$ , $x_{i}<0.5$ (min hit res $+1.15$ )	212 / 981 (21.6%) (112 / 211 [53.1%] vs. beads)	161 / 981 (16.4%) (125 / 211 [59.2%] Thr cohort)	54 / 208 (26.0%) (56 / 211 [26.5%] w/ 2 lit ctrls)	10 / 76 (13.2%) (44 / 132 [33.3%] lit)	73 / 981 (7.4%) (71 / 891 [8.0%] non-ctrl)	0 / 60 (0.0%)
Round 2	Myoglobin	$\alpha$ -Thrombin, HER2, TNF- $\alpha$ , bare streptavidin beads	977	244 Myo new / 643 off-target ( 208 Thr, 224 HER2, 211 TNF) / 0 var / 90 ctrls ( 14 lit, 16 R1 spikes, 60 scrambles)	$\log_{2}\text{FC}\geq 1.0$ , $\text{FDR}<0.05$ , OLS bead residual $>+1.00$	177 / 977 (18.1%) (12.4% in R2 sub-analysis)	150 / 977 (15.4%) (28 / 273 [10.3%] Myo cohort)	9 / 244 (3.7%) (new Myo designs; 17 / 270 [6.3%] incl. 8 / 16 R1 pos spikes; 17 / 183 [9.3%] in $N=700$ benchmark)	9 / 238 (3.8%) (0 / 6 lit)	96 / 977 (9.8%) (88 / 977 [9.0%] excl. 8 R1 pos spikes; 86 / 887 [9.7%] non-ctrl)	1 / 60 (1.7%) (0 / 53 primary random)
Round 2	HER2 ECD (Counter-screen)	$\alpha$ -Thrombin, myoglobin, TNF- $\alpha$ , bare streptavidin beads	968	229 HER2 new / 675 off-target / 0 var / 64 ctrls	$\log_{2}\text{FC}\geq 1.0$ , $\text{FDR}<0.05$ , OLS bead residual $>+1.00$ , non-binding on Myo & Thr	51 / 968 (5.3%)	46 / 968 (4.8%)	59 / 968 (6.1%) (all-pool bead-gated)	13 / 968 (1.3%) (strictly HER2-selective)	13 / 968 (1.3%) (strictly selective)	0 / 53 (0.0%)

Table 11

Candidate ID	Synthesis Well & Tube Ref ( Mfg # , Plate Row)	Architectural Family & Motif Origin	Core Variable Sequence ( 5’ $\rightarrow$ 3’ , flanked by constant primers in full-length 75--85 nt construct)	Full Construct Len (nt)	R3 Pooled SELEX ( $\log_{2}\text{FC}$ )	Raw RFU ( 0, 1, 10, 100, 1000 nM )	Specific Signal at $1\ \mu\text{M}$ ( $\Delta\text{RFU}$ , Fold / Blank)	Preliminary Apparent $K_{D}$ , Fixed-Blank Fit ( $R^{2}$ )	Preliminary Apparent $K_{D}$ , Free-Baseline Fit ( $R^{2}$ )	Binding Classification
T01 ( R3_095_var3 )	#579 , Row A ( Tube T07 )	NU172 -derived bimodal G4–duplex ( C10A bulge edit + 9-bp terminal stem)	TAACCAGCGACTAGGTTGGGTAGGGTGGTGGCGCTGGTTA	85	$+16.27$	296, 478, 888, 1586, 2121	$+1{,}825$ ( $7.2\times$ )	$24.6\pm 8.2\text{ nM}$ ( $0.980$ )	$32.0\pm 13.8\text{ nM}$ ( $0.986$ )	Strong nanomolar binder
T02	#325 , Row H ( Tube T02 )	TBA -derived coaxial G4 ( 15-nt canonical TBA loop GGTTGGTGTGGTTGG + 11-bp stem, $\Delta G=-12.0\text{ kcal/mol}$ )	AACACGCCTATAGGTTGGTGTGGTTGGATAGGCGTGTT	83	$+11.88$	276, 278, 331, 702, 812	$+536$ ( $2.9\times$ )	$47.4\pm 14.6\text{ nM}$ ( $0.986$ )	$43.4\pm 16.4\text{ nM}$ ( $0.989$ )	Strong nanomolar binder
T03	#342 , Row B ( Tube T01 )	TBA -derived coaxial G4 ( 15-nt canonical TBA loop GGTTGGTGTGGTTGG + 8-bp stem, $\Delta G=-8.2\text{ kcal/mol}$ )	AGCTGACAGGTTGGTGTGGTTGGTGTCAGCT	76	$+11.89$	218, 306, 312, 337, 427	$+209$ ( $2.0\times$ )	—*	$132.7\text{ nM}$ ( $0.783$ *)	Moderate nanomolar binder
T04	#328 , Row F ( Tube T05 )	TBA -derived loop-edited G4 ( 11-bp stem + central TBA loop G8A TGT $\rightarrow$ TAT specificity edit)	AAACACCGATAGGTTGGTATGGTTGGTATCGGTGTTT	82	$+9.81$	264, 283, 266, 335, 405	$+141$ ( $1.5\times$ )	$132.5\pm 52.8\text{ nM}$ ( $0.974$ )	$146.0\pm 77.6\text{ nM}$ ( $0.975$ )	Moderate nanomolar binder (reduced off-target enrichment in pooled screen)
T05	#581 , Row C ( Tube T08 )	Strictly de novo hybrid G4–stem (non-canonical G4 core + 10-bp GC clamp)	CGGTGCTGCATGGGTGGGTAGGGTGGGACTTCCAGCACCG	85	$+3.00$	278, 269, 280, 262, 396	$+118$ ( $1.4\times$ )	$>1{,}000\text{ nM}$	$>1{,}000\text{ nM}$	Micromolar onset (at $1\ \mu\text{M}$ )
T06	#912 , Row G ( Tube T06 )	Strictly de novo 3-tetrad G4 (3-plane G4 core + 6-bp stem, $\Delta G=-6.9\text{ kcal/mol}$ )	GCGCATGGGTTGGGATGGGTTGGGATGCGC	75	$+6.70$	291, 260, 268, 284, 349	$+58$ ( $1.2\times$ )	$>1{,}000\text{ nM}$	$>1{,}000\text{ nM}$	Trace micromolar onset (at $1\ \mu\text{M}$ )

Table 12

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

Abstract

Figures & tables

AI Agents for Scientific Discovery

From Benchmark to Bench: Can Agents Survive Real-World Drug Discovery?

ToolMol: Evolutionary Agentic Framework for Multi-objective Drug Discovery

Natural-Language-Guided Generator-Agnostic Shortlisting for Protein Binder Design

Abstract

AI Agents for Scientific Discovery

Explore similar work

From Benchmark to Bench: Can Agents Survive Real-World Drug Discovery?

ToolMol: Evolutionary Agentic Framework for Multi-objective Drug Discovery

Natural-Language-Guided Generator-Agnostic Shortlisting for Protein Binder Design