https://rfd4-proteina.org/assets/paper/RFD4-Proteina.pdf

Biomolecular design through latent reasoning with RFD4-Proteina

Key figures

• Figure 1a-g, pp. 5-6: Defines a shared representation of anchor coordinates and 32-dimensional chemical/conformational latents, recurrent refinement within denoising steps, and externally scored trajectory search. The peptide-DNA example shows that interface area, hydrogen bonding and prediction confidence become informative at different times. Panel g supports a computational benefit from additional intermediate-step refinement; cleavage activity is not experimentally demonstrated for this example.
• Figure 2a-d, pp. 7-8: Benchmarks 24 protein targets, 77 small molecules, 37 enzyme-scaffolding tasks and three DNA targets. High-thinking clustered pass rates are 10.6%, 14.3%, 12.1% and 3.5%, respectively. These quantify distinct structures satisfying prediction-based criteria. Target coverage reaches 22/24, 77/77, 35/37 and 3/3, leaving failures even in the strongest overall model.
• Figure 3a-f, pp. 10-12: Extends the model to packing, inverse folding, predicted specificity, stereochemistry, activity ranking and refinement. Panels a,b improve quality/diversity tradeoffs. Panel c estimates specificity through AF3, but its recipes also differ in redesign radius, recycling and scoring, preventing isolation of backbone-flexibility effects (§S10.3, p. 87). Panel f rescues 100/102 previously nonbinding parents computationally. Panel e’s advantage over PLACER remains uncertain under paired analysis (Supplementary Fig. 22, p. 90).
• Figure 4a-f, pp. 13-14: Demonstrates calcium-dependent binding in 12/18 designs, pMHC discrimination, IDR recognition and species-selective or cross-reactive TrkA binding. Selected YAP1 and Tau153 binders give SPR-fitted affinities of 112 and 527 pM. The 158/188 submicromolar hTrkA result counts binding across two campaigns; it does not mean 158 designs satisfy every specificity objective (Table S34, p. 106).
• Figure 5a-f, p. 16: Panels a-c support control of glycan processing: eight of nine characterized profiles contain >90% oligomannose among occupied sites. These are characterized samples, not a campaign hit-rate denominator (Supplementary Fig. 26, p. 107). Panels d-f show H3pS10/pT11 recognition, with yeast-display response midpoints of 179/133 nM and phosphosite-dependent NGS recovery.
• Figure 6a-n, pp. 18-19: Connects eight-residue, bimetallic active-site design to measured paraoxon hydrolysis, then demonstrates cysteine-dependent amide formation in a lenacapavir synthesis step. The text reports four clear ligase hits from 67 ordered designs. Design 52 converts 79.0% of the S-amine and 8.8% of the R-amine after 16 hours, versus respective no-enzyme controls of 4.3% and 6.2%. Structural overlays in d,k compare designs with AF3 predictions; they do not experimentally establish catalytic geometry.

1) Thesis (one sentence)

To address the limited joint treatment of molecular identity, conformation and chemical constraints in biomolecular design, Corley and colleagues introduce RFD4-Proteina, whose shared chemical representation, recurrent refinement and selective trajectory search improve evaluated computational design outcomes and support experimentally characterized binders, glycoproteins and catalysts, as assessed through model ablations, structure-prediction benchmarks, binding measurements, LC-MS/MS and enzyme assays (Figs. 1-6, pp. 3-20).

2) Evidence card (three bullets only)

• Strongest result: The eight-residue phosphotriesterase campaign produced 89 active designs among 288 tested, spanning 25/47 scaffold families, versus 0/192 RFD3 designs. Purified p2H11 reached apparent catalytic efficiency of 3,600 ± 500 M⁻¹ s⁻¹ at 500 nM enzyme. Separately, 15/1,000 passed computationally using 47 motif atoms and 18 substrate atoms, requiring motif RMSD <1 Å and minimum protein-substrate PAE <1.5 Å (Fig. 6c-g, pp. 17-19).
• Method enabler: Sequence and conformation are generated together, while recurrent passes refine hidden representations at fixed noisy coordinates and latents, and search evaluates emerging structures. Atom-level coordinate, interaction and sequence-probability conditioning supports diverse tasks. Energy-supervised decoder training reduces geometric defects, but supplies a learned preference rather than a universal physical guarantee (Fig. 1a,b; §§S2.1, S3.1.4, S6.1, pp. 32-37, 46, 59-60).
• Critical limitation: Performance remains task-dependent: TrkB yielded 1/94 submicromolar hit, and EA1 and TDP-A1 yielded 0/94 IDR screening hits each. The 77-ligand benchmark establishes computational agreement, without measured affinities for that cohort. Three covalent-adduct enzyme tasks were excluded, restricting conclusions about general covalent-state handling (§S8.2, p. 72; Tables S33-S34, pp. 105-106).

Optional

Key comparisons (1–3 lines)

• Compared to: RFdiffusion3, Proteina-Complexa, BindCraft and task-specific sequence, packing and ensemble methods under the manuscript’s stated protocols (Figs. 2-3).
• Win: Higher clustered pass rates across four benchmarks, with 5.0-10.4-fold greater passing-cluster yield including common AF3 evaluation (p. 8).
• Tradeoff: Additional refinement lowers generation throughput, and predictor-guided search can cost more per passing cluster (Table S29, p. 78).

Methods I might copy (protocol hooks)

Generation and conditioning. Benchmark runs use 60 denoising steps. Protein-binding low/medium/high settings allocate 120/180/300 total trunk passes, or network evaluations, early; small-molecule and enzyme tasks allocate 120/150/180 around intermediate steps. DNA uses the protein low/medium schedules and five passes per step at high thinking. Counts include the initial pass. Figure 2 conditions protein target backbones, generates ligand/DNA conformations and fixes enzyme-task ligand coordinates. These are task-specific settings, not one universal recycling schedule (§S8.3, p. 73; Supplementary Fig. 9, p. 68).

Chemical identity and stereochemistry. Small molecules retain explicit element, charge, connectivity, bond-order and aromaticity inputs. Idealized local stereochemical coordinates align to current noisy coordinates; tetrahedral training uses a signed-volume loss without requiring a whole-ligand reference conformer (§§S2.1.3, S2.3.4, pp. 35, 41). In a separate 43-ligand held-out test, 98.8% chirality and 92.2% E/Z adherence average eligible center/bond accuracy over outputs, then equally over ligands. These are not whole-molecule correctness rates; corresponding AF3-refold values are 62.9% and 66.1%. The 92.3% PoseBusters result excludes connectivity, InChI-conversion and energy tests, leaving those aspects unvalidated by this score (Fig. 3d; §S10.4, p. 87).

Search with a defined stopping rule. On ten ligand tasks, donor/acceptor assignments use pH 7.4-protonated molecules, supplemented by aromatic-ring stacking requirements. Search stops at step 30 if more than two requirements remain unmet; completed structures may miss at most one before AF3 evaluation. At four GPU-hours per target, passing clusters rise from 541 to 674 overall. Apixaban and methotrexate worsen, and SAM yields none under either strategy. Predictor-guided search also reduces computational yield on every standard benchmark (Tables S26-S29, pp. 65-66, 78).

Validation and accounting. Keep all 200 original attempts per target, including failures, and divide passing Foldseek clusters by that denominator; the TM-score threshold is 0.5. Small-molecule passes require protein-backbone RMSD ≤1.5 Å, protein-aligned ligand RMSD ≤5 Å, ipTM ≥0.80 and minimum interface PAE ≤1.5 Å (§§S8.1-S8.2, p. 72; S12, p. 89). Main-figure BindCraft uses one backbone and one sequence; full pipelines appear separately in Supplementary Fig. 16. Generation throughput excludes common AF3 evaluation, while reported GPU costs exclude setup, loading, compilation, CPU scoring and idle allocations. Full campaign elapsed time and cost remain unquantified (§§S8.3, S8.6, pp. 73, 78).

Binding measurements. Calmodulin BLI compares 2 mM CaCl₂ with 3 mM EDTA, representing calcium-loaded versus divalent-depleted conditions, using 300-second association/dissociation phases at 25°C; the two lowest apparent affinities reach the off-rate fitting floor of 10⁻⁵ s⁻¹ (§S13.1.2, pp. 91-92). For pMHCs, the 340 pM, 710 pM and 1.60 nM values are tetramer-staining midpoints with avidity; the >100-fold specificity claim comes from a separate staining-response comparison. Phosphopeptide Hill midpoints and NGS recovery likewise require their own readout labels (Fig. 4b; §§S13.2, S13.6.6-S13.6.8, pp. 92, 96).

Enzyme activity. PTE kinetic assays use 500 nM enzyme, 405-nm initial-rate measurements at 25°C and pH 8, with 1% methanol, 200 µM ZnSO₄ and 25 mM final bicarbonate after a ≥2-hour, 50 mM bicarbonate pretreatment. Turnover is normalized to total protein-chain concentration; active enzyme concentration was not independently measured. Reported uncertainties propagate measurement assumptions and technical variation from one representative biological replicate (§S13.7.4, pp. 98-100). At 100 nM enzyme, the best design’s reported efficiency is 1,700 ± 300 M⁻¹ s⁻¹, warranting further concentration-dependence checks (Supplementary Fig. 31d,e, p. 112).

Ligase endpoints. Assays use 50 µM enzyme, 1 mM activated ester and 0.5 mM enantiopure amine in 100 mM potassium phosphate, pH 6, with 20% DMSO, at 30°C for 16 hours. UPLC measurements use chemical standards and three measurements per condition without background subtraction. These data support product formation and endpoint stereopreference; initial-rate selectivity remains unresolved. Multistate AF3 filtering is mentioned, but the experimental supplement lacks a complete state-by-state computational recipe (§5.3, p. 19; §S13.8, pp. 100-101).

Cautions / Limitations

• Conflicting computational counts: The manuscript reports 15/1,000 versus 0/10,000 phosphotriesterase computational passers, while the website reports 111/1,000 versus 1/10,000. Use the manuscript counts for this comparison (Fig. 6c; PDF pp. 17-19).
• Phosphosite scope: The manuscript demonstrates H3pS10/pT11 binders and treats pS28 as an off-target; the website claims binders to all three sites. The manuscript supports the narrower scope (Fig. 5d-f; PDF pp. 15-16, 109-110).

Open questions / Theoretical implications (2–5 bullets)

• Can stage-specific search rewards be selected prospectively across tasks, given their gains under strict ligand-burial criteria but poorer efficiency under ordinary criteria (Table S22, p. 63)?
• Will ensemble spread predict enzyme activity prospectively? The highlighted retroaldolase intermediate was selected retrospectively, and paired AUROC intervals do not establish superiority over PLACER (§S10.5; Supplementary Fig. 22, pp. 87, 90).
• Can experimental structures and modification-state measurements confirm the proposed carbamylated-lysine, metal-coordination and ligase geometries beyond functional assays and predicted overlays (Fig. 6b,d,i,k; §§S13.7-S13.8)?
• How much pMHC and phosphopeptide discrimination persists in monovalent solution assays, and how broadly does it extend beyond the tested off-target panels (Figs. 4b, 5d-f; §§S13.2, S13.6)?