https://www.biorxiv.org/content/10.64898/2026.09.27.754867v2

Deep-learning-based design of an orthogonal self-labeling protein from K-Ras(G12C)

Key figures

• Figure 1 (pp. 2–3): Defines LUCI-tag as a third self-labeling protein and shows the design strategy: preserve a drug-recognition pocket while redesigning the surrounding K-Ras sequence and removing the membrane-associated hypervariable region.
• Figure 2c–h (pp. 5–6): Establishes monomericity, thermal stability, rapid labeling, tag–probe selectivity, and reversible probe binding in the C12S variant; separates labeling kinetics from noncovalent affinity.
• Figure 3 (pp. 9–10): Validates the adagrasib-bound and inhibitor-free structures and supports an open-pocket preorganization model; the inhibitor-free structure still contains GDP–Mg²⁺, so it is not completely ligand-free.
• Figure 4 (pp. 10–11): Tests three distinct properties: loss of native Ras-associated interactions in MOLT4 lysate, lack of elevated MEK/ERK phosphorylation in the tested cells, and selective probe labeling on lysate gels.
• Figure 5 (pp. 12–14): Demonstrates wash-free nuclear imaging with AG–JF646, extracellular receptor labeling, and three-color imaging with HaloTag7 and SNAP-tag; the three-color experiment includes washing.
• Figures S24–S25, S33–S34 and Tables S4–S5 (SI pp. S39–S40, S46–S47, S58): Resolve nucleotide dependence, the partially reduced buffer dependence of LUCI-tag–CF, differences among dye payloads, and the slower development of fluorescence after rapid covalent labeling.

1) Thesis (one sentence)

To expand orthogonal protein labeling beyond the commonly used HaloTag7 and SNAP-tag systems, the authors extensively redesigned K-Ras(G12C) around its preserved covalent-inhibitor pocket to produce LUCI-tag, a compact self-labeling protein with rapid probe attachment and three-color live-cell utility supported by biochemical kinetics, crystallography, proteomics, signaling controls, and microscopy.

2) Evidence card (three bullets only)

• Strongest result: LUCI-tag labels with AG–AF488 at an apparent second-order rate constant of 5.54 × 10⁵ M⁻¹ s⁻¹, versus 1.97 × 10⁴ M⁻¹ s⁻¹ for the minimally engineered K-Ras(G12C) comparator, approximately a 28-fold gain; AG–TAMRA labeling is complete by the first 30-second gel timepoint at 10 µM protein and 0.5 µM probe, and three tag–probe systems give spatially distinct three-color labeling in cells (Fig. 2d–g, Table S4, Fig. 5c).
• Method enabler: Starting from experimentally established inhibitor-bound Ras folds, ProteinMPNN, LigandMPNN, and LASErMPNN redesign permitted residues while preserving Cys12 and Lys16; surface-charge diversification, single-sequence AF2 and ligand-aware Boltz-1 filtering reduce roughly 20,000 designs to 48 experimental candidates, all reported soluble and monomeric, with about half showing rapid ARS-1323 labeling (Fig. 1d, Figs. S3–S7, SI Computational Methods, pp. S3–S6).
• Critical limitation: The kinetic advantage is payload-specific and the HaloTag7/SNAP-tag rate comparisons use values from an earlier study rather than a fully matched experiment here: LUCI-tag is substantially slower than HaloTag7 with TAMRA, its AF647 and JF646 rates differ about 8.5-fold, and JF646 fluorescence develops more slowly than covalent attachment; thus neither universal payload independence nor universally superior imaging follows from this preprint (Fig. 2f, Table S4, Figs. S33–S34).

Optional

Quote bank (2–4 short excerpts)

• Quote 1: “preserving its covalent-inhibitor binding pocket” (Abstract, p. 1).
• Quote 2: “fluorescence turn-on appears to have a slower rate” (Fig. S33 caption, SI p. S46; combined PDF p. 66).

Key comparisons (1–3 lines)

• Compared to: Cys-light, HVR-truncated K-Ras(G12C) for biochemical assays, and HaloTag7/SNAP-tag with their cognate probes for labeling and imaging (Fig. 2, Table S4, Figs. S32–S34).
• Win: A 169-aa, 19-kDa tag tolerates extensive sequence change, provides a third labeling channel, and labels the negatively charged AF488 payload rapidly; C12S binds three AG–dye conjugates at approximately 0.12–0.14 µM (Fig. 2h, Table S5).
• Tradeoff: Standard LUCI-tag remains sensitive to cofactor-depleted buffers and has less JF646 fluorogenicity than HaloTag7; probe chemistry, nucleotide state, expression, and the imaging protocol still matter (Figs. S25, S32–S34).

Methods I might copy (protocol hooks)

• Construct design / Models: Use inhibitor-bound templates 8AFB, 6OIM, and 6USZ; omit the HVR, retain Cys12/Lys16, and define pocket/cofactor contacts by a 4.8 Å heavy-atom cutoff. The five design categories vary which inhibitor, cofactor, and PPI residues can change. Generate 512 sequences per eligible model/template/category combination; ProteinMPNN and LigandMPNN use temperature 0.2, while LASErMPNN reports first-shell temperature 0.5. Predict the unliganded sequence with ColabFold and the complex with Boltz-1 using five recycles, five models, and an explicit inhibitor–Cys12 covalent constraint. The constraint tests compatibility with a specified adduct; it does not predict spontaneous bond formation. AF2 core metrics omit much of the switch loops; inspect those regions separately. The final panel contains 22 ProteinMPNN, 11 LigandMPNN, and 15 LASErMPNN designs, with category-specific thresholds rather than one universal cutoff (SI pp. S3–S6; Tables S2–S3).
• Conditions / Instruments: Express the selected proteins from peAIP32 with a His–SUMO fusion in BL21(DE3), using overnight autoinduction at 30 °C for at least 16 hours. The reported protein buffer is 20 mM Tris pH 7.4, 100 mM NaCl, 2 mM MgCl₂, 10 µM GDP, and 1 mM TCEP. Purify by Ni-NTA and on-resin release with 1 µM Cth protease overnight at 4 °C. Verify oligomeric state on a BioRad NGC/ENrich 650 column and secondary structure/thermal stability by Jasco J-1500 CD; LUCI-tag does not visibly melt over the measured range up to 95 °C (SI pp. S6–S7; Fig. 2c).
• Readout / Analysis — distinguish binding from chemistry: Measure C12S reversible binding with 50 nM AG–dye and a protein titration in technical triplicate; fit a quadratic single-site model. For fast labeling, globally fit concentration-dependent anisotropy to P + S ⇌ PS → PC with measured mixing/reader dead time; anisotropy counts both PS and PC. Report k_app = k₁k₂/(k₋₁ + k₂), with 500 residual-bootstrap refits for confidence intervals, and corroborate covalency by gels/intact MS. The main AG–AF488 result uses 10 nM probe; the general time-course protocol describes 50 nM, so preserve the experiment-specific concentration. The slower ARS-1323 screen uses an approximate pseudo-first-order slope for candidate ranking, not the same full fit (Fig. 2e; SI pp. S7–S11).
• Readout / Analysis — fluorescence and imaging: Measure fluorescence intensity independently of anisotropy: Fig. S33 uses 100 nM AG–JF646 and 200 nM protein for a 90-minute time course, with three technical replicates. For nuclear cellular kinetics, GFP–SLP–PITX2 permits expression normalization; background-corrected JF646 is divided by the same nucleus’s initial GFP signal, and fixed, unsaturated tracked cohorts are analyzed. Single-tag AG–JF646 imaging can omit washing, whereas multicolor staining uses 500 nM per probe and a wash/incubation sequence. The SI identifies the imaging cells as HEK293FT, called HEK293T in the main text; reproduce the stated subline and the exact figure’s conditions rather than merging protocols (Figs. S33–S34; SI pp. S17–S19).
• Controls / Orthogonality: Use a complete tag × probe matrix, C12S, tag-negative cells, lysate gels, and separate cellular signaling readouts. IP-MS compares purified FLAG-tagged constructs incubated with MOLT4 lysate; it is a biochemical interaction screen, not an exhaustive interaction map in intact cells. Fig. 4b includes full-length Ras, HVR-truncated Ras, and mock controls, and both LUCI-tag and truncated Ras lack the strong signaling phenotype. The experiment therefore supports silence of the tested LUCI construct without isolating sequence redesign as the sole cause. Fig. S42 additionally tests G12C-positive H358 cells at 500 nM AG–JF646 for one hour, but absence of visible background does not establish absence of endogenous target engagement (Fig. 4; SI pp. S11–S13, S19; Fig. S42).

Open questions / Theoretical implications (2–5 bullets)

• How much of the kinetic gain comes from pocket opening versus chemistry? The inhibitor-free, GDP-bound crystal structure and sequence-likelihood analysis support preorganization, with M67E/Y71E accounting for much of the computed preference, but do not independently assign their effects on association, dissociation, or the chemical step (Fig. 3; Fig. S27).
• How far does payload tolerance extend? Table S4 gives 5.71 × 10⁵, 5.54 × 10⁵, 1.27 × 10⁶, and 1.50 × 10⁵ M⁻¹ s⁻¹ for TAMRA, AF488, JF646, and AF647, respectively; oligonucleotides, degradation handles, and other payload classes remain proposed extensions rather than a demonstrated universal property.
• Can fluorogenicity be improved independently of covalent capture? The rapid chemical labeling but slower, approximately ninefold JF646 brightness increase over 90 minutes suggests an additional optimization objective; matched raw intensity, expression normalization, and time courses are needed alongside endpoint images (Fig. S33; Figs. S32–S34).
• What remains of cofactor dependence and biological activity outside the tested settings? LUCI-tag–CF loses about fivefold labeling rate in the depleted-buffer comparison versus roughly 110-fold for LUCI-tag, rather than being perfectly buffer-independent; that comparison also changes salt and adds EDTA. Broader cell types, endogenous G12C target engagement, and in vivo probe behavior remain to be established separately (Fig. S25; SI p. S20; Fig. S42; Discussion).