https://www.biorxiv.org/content/10.64898/2026.10.19.759856v1

A rapid, automation-friendly, cell-free and purification-free pipeline for binding and kinetic characterisation of de novo designed binders with high quality standardised per-design data

Key figures

• Figure 1, p. 5: Connects computational design, DNA assembly, cell-free protein synthesis (CFPS), and flow-induced dispersion analysis (FIDA), followed by affinity and kinetic measurements for selected candidates from the same CFPS preparation. The reported three-day experimental timeline begins with receiving DNA, excluding computational design and gene synthesis (Introduction, p. 4).
• Figure 2a–c, pp. 6–7: Shows 117 computationally passing designs from the unseeded hGH campaign after one partial-diffusion round and 16 hours, versus 77 and 18 after five rounds and 36 hours for the two hotspot campaigns on an NVIDIA A100. These are computational pass counts; subsequent experiments tested 32 selected candidates.
• Figure 3a–d, pp. 8–9: Shows amplification and Golden Gate assembly of linear DNA and expression of binder–mChartreuse fusions. DNA gels support assembly; SDS-PAGE examines five selected purified fusion proteins. The reporter is genetically attached to the binder, allowing commercial hGH to remain unlabeled.
• Figure 4b–e, pp. 10–11: Identifies 30/32 designs with an appropriate fusion-size readout and 11 with a target-dependent hydrodynamic-radius increase greater than 0.1 nm at 5 µM hGH: 11/32 total designs, approximately 34%, or 11/30 size-qualified fusions, approximately 37%. HGH_17 and HGH_20 were excluded because their fluorescent species were smaller than expected, consistent with complete or partial fusion cleavage.
• Figure 4f–g, p. 11: Reports equilibrium and nonequilibrium titrations for HGH_1, HGH_6, HGH_11, and HGH_29, selected for the largest screening radius shifts. Their fitted dissociation constants span 0.93–9.98 nM. These four were not demonstrated to have the highest affinities among all 11 hits.
• Figure 5c–f, pp. 12–13: Reports classification AUCs of 0.58, 0.43, and 0.50 for AF2-IG ipAE, Boltz-2 ipSAE_min, and AF3 ipSAE_min within the selected 32-design cohort. Campaign hit counts are 8/20, 2/9, and 1/3; Fisher’s exact P = 0.65 does not establish equivalent campaign success rates.

1) Thesis (one sentence)

To address the gap between rapid computational binder generation and slower quantitative experimental validation, Ruiz Espi and colleagues combine iterative protein design, cell-free expression of fluorescent binder fusions, and FIDA against human growth hormone to obtain expression and size measurements across a 32-design cohort, identify 11 screening hits, and derive affinities and kinetic parameters for four selected candidates from unpurified material, supported by gel analyses, fluorescence calibration, and equilibrium and nonequilibrium binding titrations (Figs. 1–4, pp. 5–11).

2) Evidence card (three bullets only)

• Strongest result: The same CFPS preparations support screening and quantitative follow-up without routine binder protein purification. For the four largest-shift candidates, FIDA-fitted affinities range from 0.93 ± 0.36 to 9.98 ± 2.24 nM, association rates from approximately 3.77 × 10⁵ to 5.83 × 10⁶ M⁻¹ s⁻¹, and dissociation rates from approximately 6.7 × 10⁻⁴ to 1.8 × 10⁻² s⁻¹ (Fig. 4f–g, p. 11; Methods, p. 20).
• Method enabler: The binder–mChartreuse fusion supplies the fluorescent FIDA indicator and a concentration estimate through fluorescence-area calibration; hydrodynamic radius supplies the size-sensitive binding readout. Assay establishment includes purified HGH_3 for calibration and five purified fusions for SDS-PAGE, while DNA purification remains part of the workflow. Purification-free therefore describes routine binder protein handling, and avoiding labeling means avoiding a separate target-labeling step (Figs. 3–4, pp. 9–11; Methods, pp. 18–20).
• Critical limitation: Absolute affinity and kinetic accuracy remain incompletely validated: the fits cover four fluorescent fusions against one target, without an independent binding/kinetic assay. Methods specifies injection programmes and proprietary software but not the fitted binding/transport model, fixed versus estimated parameters, exact follow-up indicator concentrations, or equilibration interval. Fluorescent concentration calibration does not establish each design’s binding-competent fraction; equilibrium, ligand depletion, and transport-model assumptions consequently remain consequential (Fig. 4f–g, p. 11; Methods, “Binding affinity and kinetics,” p. 20).

Optional

Key comparisons (1–3 lines)

• Compared to: The paper contrasts its workflow with heat-treated bacterial-lysate FIDA screening using fluorescent targets; its integration replaces bacterial binder production with CFPS and moves the fluorescent indicator onto the binder (Discussion, p. 15).
• Win: Expression estimates, apparent-size assessment, and target-dependent binding measurements share one preparation (Fig. 4a–e, p. 11).
• Tradeoff: A fluorescent fusion, resolvable binding-associated size change, and FIDA access are required; comparative accuracy and total processing time were not established in a matched platform benchmark (Discussion, pp. 14–15).

Methods I might copy (protocol hooks)

Computational selection. Generate 300 RFdiffusion scaffolds and five ProteinMPNN sequences per scaffold; evaluate with AlphaFold2 Initial Guess. Retain ipAE <8 Å and pLDDT >80. Use the top 15 candidates, at most two per scaffold, to seed partial diffusion, stopping when at least 100 designs pass or five partial-diffusion rounds finish. Rank passing candidates by Boltz-2 ipSAE_min. These are study-specific settings, not established universal thresholds (Results, pp. 5–6; Fig. 2a, p. 7; Methods, p. 17).

DNA assembly. The reported Golden Gate reaction totals 16.06 µL and uses 30 cycles of 37 °C for 60 seconds and 16 °C for 60 seconds, followed by 60 °C for 10 minutes. Fixed insert volume, rather than DNA-mass normalization, makes insert-to-vector molar ratios construct-dependent. Supplementary Tables 1–2 separately describe backbone PCR: 50 µL reactions and 34 cycles with 56 °C annealing for 20 seconds and 72 °C extension for 45 seconds (Methods, p. 18; Tables 1–2, pp. 21–22).

Cell-free expression. Combine 18 µL ALiCE reagent with 2 µL purified assembled DNA at 72–120 ng/µL in a low-binding 384-well plate with a permeable lid. Incubate for 36 hours at 25 °C, 800 rpm, and 85% relative humidity (Methods, p. 19).

Screening and calibration. Dilute CFPS material 1:100 into 10 mM sodium phosphate, pH 8.5, with the reported 0.05% Pluronic acid. Measure fluorescent binder alone and with 5 µM hGH in duplicate on FIDA Neo; the screening criterion is ΔRₕ >0.1 nm. The paper uses 480 nm excitation and emission above 515 nm. Calibration against purified HGH_3 gives R² = 0.997 in the Results; the plotted fit labels R² = 0.9974. The reported minimum indicator concentration is approximately 2.5 nM for SNR >100, above the tightest fitted K_D of 0.93 nM, making explicit depletion treatment important without itself invalidating the fit (Results, pp. 8, 10; Fig. 4e, p. 11; Methods, pp. 19–20).

Equilibrium and kinetic follow-up. Hold indicator concentration fixed and titrate hGH across eight nonzero concentrations from 0.61 nM to 10 µM, plus zero. PreMix uses premixed binder and target. The kinetic CapMix programme injects indicator for 10 seconds at 50 mbar, then analyte for 60 seconds at 1,200 mbar; the equilibrium programme instead uses a 180-second analyte step at 400 mbar. Figure 4g contains nonequilibrium radius-versus-concentration curves from which rates are inferred, rather than directly plotted dissociation time courses. The statistical meaning of the reported ± quantities is not defined (Fig. 4f–g, p. 11; Methods, p. 20).

Automation and capacity. The implemented Tecan Fluent 780 workflow covers gene-fragment amplification, DNA purification, Golden Gate assembly, and a second DNA purification, using four positions per construct in a 384-well plate for up to 96 constructs. The Discussion’s 288 designs every five days is proposed processing capacity; the demonstrated dataset comprises 32 designs and four quantitative follow-ups (Methods, pp. 20–21; Discussion, pp. 14, 16).

Cautions / Limitations

• Size and specificity have limited scope: Fusion radii near 2.9–3.1 nm and hGH near 2.4 nm support gross-size assessment, without independently establishing native oligomeric stoichiometry or the inferred cleavage of HGH_17/HGH_20. Lysate background tests gross nonspecific association, not specificity against unrelated targets (Fig. 4b,h, p. 11; Results, p. 10; Discussion, p. 15).
• Confidence conclusions depend on selection and labels: Figure 5 concerns a small, prefiltered cohort across 11 backbone trajectories. Its 11-positive/21-negative grouping includes the two fusion-size failures previously excluded from binding analysis, mixing integrity/expression failure with intact-protein nonbinding; weak discrimination here does not establish universal metric failure (Results, pp. 10, 12; Fig. 5a–d, p. 13).
• Reporting discrepancies affect reuse: Figure 4c and the Results support hit-associated ΔRₕ of approximately 0.11–0.43 nm, whereas the Discussion states 0.43–1.11 nm. The supplementary introduction calls Tables 1–2 Golden Gate conditions, although their actual captions and contents specify backbone PCR; Golden Gate conditions are in Methods, p. 18 (Fig. 4c, p. 11; Discussion, p. 15; Tables 1–2, pp. 21–22).

Open questions / Theoretical implications (2–5 bullets)

• How does target/fusion size affect sensitivity and false-negative rates? The 0.1 nm criterion was demonstrated for one combination, and small relative radius changes can be difficult to resolve (Fig. 4b–c, p. 11; Discussion, p. 15).
• Does hGH binding produce receptor competition or functional antagonism? Predicted complexes converge on the receptor-binding face, but experimental complex structures, GHR competition, and signaling-blockade measurements are not reported (Fig. 5a–b, p. 13; Discussion, p. 14).
• Can separate integrity, expression, size, screening, affinity, and kinetic labels improve prospective design selection? Model retraining and prospective gains remain proposed outcomes rather than demonstrated results (Discussion, p. 16).