Designing Protein Structures from Natural Language

1 INTRODUCTION

• Existing structure-generation models typically need a separate engineered input representation for each property, limiting the design objectives a single model can support.Natural language is presented as a unified conditioning interface for expressing diverse properties.Source: §1 · p. 2
• Prompteina jointly generates an all-atom structure and sequence from text, steering global properties and residue-specific features such as disulfide bonds and binder–target contacts.The model is intended to let language-model agents pass specifications directly to the generator rather than translate them into fixed, model-specific inputs.Source: §1 · p. 2
• Prompteina-Data pairs 1.8M structures with 52.4M captions, and Prompteina is adapted through text-conditioned pretraining followed by fine-tuning on caption–residue correspondences.Figure 2 depicts annotation-to-caption construction, residue–text alignment, and the two-stage training sequence.Source: §1 · p. 2, Figure 2 · p. 3
• Prompteina leads evaluated baselines on 15 of 18 structural-property tasks, supports more properties than any single baseline, and more closely matches intended-function descriptions.In agent-guided binder design, it outperforms every baseline on at least 13 of 20 targets and raises interface confidence by 14% over the strongest baseline.Source: §1 · p. 2
• Nine of 89 assayed GM2A designs were classified as binders, with the strongest reaching an apparent KD of 107 nM across three replicates.GM2A had no reported de novo binders; the campaign was conducted as a proof of concept.Source: §1 · p. 2

2.1 PROMPTEINA-DATA: TEXT–STRUCTURE DATA

• Prompteina-Data contains about 1.8M chain entries drawn from experimental, predicted, and synthetic structures, with up to 59 annotations per entry.Its 52.4M captions average 29 per entry; an LLM-judged sample matched ground-truth annotations with 97.8% recall, 89.8% precision, and 95.7% accuracy.Source: §2.1 · p. 3

2.2 PROMPTEINA: TEXT-CONDITIONED STRUCTURE GENERATION

• Prompteina adds pooled text conditioning for global properties and token-level cross-attention for residue-specific properties to the pretrained Complexa generator.The combination of pooled conditioning and cross-attention performs best across tasks in the reported ablations.Source: §2.2 · pp. 3–4
• Residue spans associated with annotations are matched to caption text spans, and an auxiliary head is trained to predict which residue–token pairs align.When either span cannot be located, the caption–structure pair is retained but that alignment annotation is omitted; sparse positive pairs are reweighted in the alignment loss.Source: §2.2 · p. 4
• After alignment fine-tuning, Prompteina gains finer control over disulfide-bond and salt-bridge counts while retaining global fold control without the alignment head at inference.The reported comparison is in Table 10.Source: §2.2 · p. 4

2.3 TRAINING AND INFERENCE

• Training first adapts new text-conditioning parameters with the denoiser frozen, then jointly trains the denoiser on caption–structure pairs; alignment fine-tuning freezes the denoiser.At inference, a prompt and protein length are inputs; binder design also takes a target structure and optional contact residues, and the model decodes an all-atom structure and sequence.Source: §2.3 · p. 4

3.1.1 CONTROL OF STRUCTURAL PROPERTIES

• Prompteina controls all five structural-property groups and leads baselines on 15 of 18 tasks, with examples spanning global and residue-specific properties.It also supports multiple domains and combined constraints, while preserving designability and novelty.Source: §3.1.1 · pp. 5–6, Table 1 · p. 5, Figure 3 · p. 5

3.1.2 GENERATION FROM FUNCTIONAL SPECIFICATIONS

• At each model’s best prompt format, Prompteina achieves the highest mean on both ProTrek similarity metrics for functional specifications.With pLDDT weighting, it significantly outperforms BioMatrix on both metrics and Pinal on sequence similarity, while retaining the highest mean structure similarity.Source: §3.1.2 · p. 6, Table 2 · p. 6

3.2.1 BINDER–TARGET INTERFACE STEERING

• Across 20 targets, requested and realized interface residue counts track closely, with Spearman’s ρ = 0.994–1.000 and linear-fit R2 = 0.916–0.996.Prompteina can tune interface counts below or above unconditional generation, and this control is robust to prompt paraphrasing.Source: §3.2.1 · pp. 6–7, Figure 4 · p. 7

3.2.2 AGENTIC PROTEIN BINDER DESIGN

• With agent-written specifications, Prompteina leads expected top-k ipSAEmin across all three sequence conditions and accepts richer interface controls than the baselines.Its advantage is robust under AlphaFold3 evaluation; under redesign, its top designs are more diverse than either BoltzGen variant and match agent-guided RFdiffusion3 in diversity.Source: §3.2.2 · pp. 7–8, Table 3 · p. 8
• Adding interface properties yields the largest gain in interface confidence; on CD45, Prompteina’s two best designs outscore every baseline while matching distinct folds and sites specified in prompts.Adding shape yields only a modest gain, while fold plus shape exceeds fold only by 0.002 expected top-k.Source: §3.2.2 · p. 8, Table 4 · p. 8

3.2.3 ITERATIVE BINDER PROMPT OPTIMIZATION

• Iterative feedback benefits Prompteina more than either baseline, with larger top-1 and top-8 gains on the hardest targets.On Claudin-1, the agent tests new topologies and then requests a larger interface, improving the best design after each round.Source: §3.2.3 · p. 9, Figure 5 · p. 9

3.3 EXPERIMENTAL VALIDATION OF GM2A BINDER DESIGNS

• Nine binders were identified among 89 GM2A designs assayed, including six sub-micromolar binders; the strongest had an apparent KD of 107 nM across three replicates.The strongest design realized the requested two-layer alpha/beta architecture around W94, with 21 interface residues versus 22 requested.Source: §3.3 · p. 9, Figure 6 · p. 9

Protein structure generation

• Structure-generation models use custom inputs for folds, motifs, hotspots, and interactions; RFdiffusion3 and BoltzGen combine controls but retain distinct inputs and annotations for each property.Chroma is a natural-language precedent, but uses a separate structure-to-caption model for inference-time guidance rather than text-conditioned denoising.Source: §4 · p. 10

Protein–text multimodal models

• Protein–text alignment and captioning models support retrieval, classification, or description, while text-guided models such as ProteinDT and Pinal generate sequences that must be folded into structures.HelixProtX emits backbone torsion angles from text, but its pretrained parameters are unavailable for evaluation.Source: §4 · p. 10
• Prompteina directly conditions sequence and all-atom structure generation on text, using supervised text–residue alignment to connect phrases with the residues they describe.This differs from Chroma’s inference-time caption-guidance approach and sequence-only text-guided models.Source: §4 · p. 10

5 DISCUSSION

• Prompteina and Prompteina-Data provide a single natural-language interface for controlling global and residue-specific properties, while supporting agent-written prompts and experimental feedback in design.The discussion reports GM2A binders where no de novo binders had previously been reported.Source: §5 · p. 10

A EXTENDED RELATED WORK

• RFdiffusion uses separate task-specific model variants and inputs for fold control, binder design, motif scaffolding, and symmetry.RFdiffusion2 and RFdiffusion3 use dedicated input formats or procedures for properties despite using a single model.Source: A EXTENDED RELATED WORK · pp. 18–19
• Other approaches use text for gradient guidance or generate sequences that are subsequently folded, rather than jointly generating all-atom structures and sequences from text.Chroma's text guidance uses a structure-to-caption predictor, while Pinal generates sequences from functional prompts and folds them with ESMFold.Source: A EXTENDED RELATED WORK · pp. 19–20
• Several multimodal models support multiple modalities, but use dedicated tracks or mechanisms for conditioning properties.ESM3 assigns property tracks their own tokenization and training choices; BoltzGen uses separate representations and distributions for its conditioning mechanisms.Source: A EXTENDED RELATED WORK · p. 20

B.1 DATA SOURCES AND SCALE

• The dataset combines PDB chains, AFDB monomers, and Teddymer dimers into 1.8M entries with 52.4M captions.PDB and AFDB supply monomers; Teddymer dimers are assembled from TED domain pairs, and chains are captioned independently.Source: §B.1 · p. 21, Table 5 · p. 21

B.2 ANNOTATION EXTRACTION AND CATEGORIES

• Annotations cover ten groups, with 59 property types for PDB, 26 for AFDB, and 32 for Teddymer, subject to source-specific exclusions.The groups include fold and topology, secondary structure, motifs, shape, interfaces, ligand binding, domain architecture, and function.Source: §B.2 · pp. 21–22

B.3 RESIDUE-SPECIFIC CAPTIONS

• For 15 properties, captions may include 1-based residue positions; position inclusion is more frequent for properties such as disulfides than for interface interactions.Emergent properties such as flexibility omit positions because individual residues are not independently designable; positional prompts mix with position-free descriptions.Source: §B.3 · p. 22

B.4 CAPTION SAMPLING

• Caption sampling varies how many annotations a structure contributes, with β = 5.0 favoring sparse captions while retaining occasional dense descriptions.A coverage criterion yields averages of approximately 32 captions per PDB entry, 26 per AFDB entry, and 28 per Teddymer chain.Source: §B.4 · p. 23

B.5 LLM CAPTION GENERATION

• Qwen3-235B-A22B generates captions from structured annotations under instructions to state only provided facts, but an audit found unsupported numerical specificity caused most precision loss.The LLM-judge validation covered 80 captions; despite these errors, captions retained high recall, precision, and accuracy on the audited sample.Source: §B.5 · pp. 23–24

B.6 TEXT–RESIDUE ALIGNMENT ANNOTATIONS

• Text spans are matched to source annotations and mapped to property-specific residue sets, while unreliable matches are omitted from alignment masks.Manual validation over 267 hold-out spans reported 99% coverage, 89% character-level precision, and 73% recall.Source: §B.6 · p. 24

C MODEL AND TRAINING DETAILS

• Prompteina adds pooled text embeddings to Complexa's conditioning array and uses frozen Qwen encoders to provide pooled and token-level text representations for conditioning.The token representations are aggregated across language-model layers and projected for cross-attention to protein states.Source: C MODEL AND TRAINING DETAILS · p. 24, §C.1 · p. 25, §C.2 · p. 25
• A text–residue alignment head uses gated cross-attention updates to predict residue–token alignment logits, with positive entries upweighted during training.Cross-attention is gated per denoiser layer; its conditioning pathway is initialized at zero so its contribution can grow during training.Source: §C.4 · p. 26, §C.5 · p. 26

C.6 TRAINING

• Monomer training warms up text pathways with the denoiser frozen, jointly pretrains all parameters, then fine-tunes alignment components with the denoiser frozen.Restricting alignment fine-tuning to late value and output projections preserves fold accuracy; unconditional dropout supports classifier-free guidance.Source: §C.6 · pp. 26–27
• Binder training noises only binder tokens while keeping target tokens fixed, and uses optional hotspot masks to indicate target residues for contact.Training text describes the binder, while the target structure is supplied separately.Source: §C.6 · p. 27
• CATH class accuracy is 92.0% with cosine cycling, 87.3% with a linear ramp, and 51.3% without rebalancing.Rebalancing interpolates cluster weights between empirical frequency and uniform sampling, emphasizing rare feature combinations.Source: §C.6 · p. 27
• The backbone timestep schedule is changed to a uniform distribution to give more weight to intermediate timesteps, while the side-chain schedule is retained.The stated rationale is that structural decisions and informative text conditioning occur at intermediate timesteps.Source: §C.6 · p. 27

C.7 SAMPLING

• At inference, an SDE generates backbone coordinates and side-chain latents, which a frozen VAE decoder converts into all-atom coordinates and sequence.Sampling uses 400 Euler–Maruyama steps, and classifier-free guidance combines conditional and unconditional velocities with a weight controlling text-conditioning strength.Source: §C.7 · p. 28

D ABLATIONS

• 92.0% CATH class accuracy comes from the complete model; removing rare-class rebalancing drops it to 51.3%, and inherited Beta noise yields 68.7%.The complete model also reaches 27.0% CATH architecture accuracy; cross-attention alone gives 86.7% class accuracy, while pooled conditioning alone gives 68.7%.Source: D ABLATIONS · p. 28, Table 8 · p. 29
• Adding token cross-attention raises CATH class and architecture accuracy over pooled text alone, and 3.2× longer CATH-label training does not close the gap.Pooled text and CATH labels perform similarly through AdaLN; the comparison uses a matched 47,275-step budget before extending label training.Source: D ABLATIONS · p. 28, Table 9 · p. 29
• After alignment fine-tuning, disulfide-count correlation rises from 0.054 to 0.294 and salt-bridge-count correlation from 0.566 to 0.663.CATH class accuracy increases from 94.0% to 96.0%, while architecture accuracy changes from 55.5% to 51.1%.Source: D ABLATIONS · p. 28, Table 10 · p. 29

E MONOMER GENERATION EVALUATION

• Text-conditioned monomer generation is evaluated on 18 tasks: 14 single-property and four composite tasks, using 50 structures per prompt.The evaluation covers fold and topology, secondary structure, tertiary structure, global shape and surface, and combinations of CATH class with structural counts.Source: E MONOMER GENERATION EVALUATION · p. 30, Table 11 · p. 31
• With identical lengths across requested counts, text-conditioned helix and strand correlations remain strong while empty-prompt control correlations are near zero.The matched-length design makes length uninformative about the requested count by construction.Source: E MONOMER GENERATION EVALUATION · p. 30, Table 12 · p. 31

E.1 PROMPT PARAPHRASE ROBUSTNESS

• Across four prompt formats, CATH accuracy is at least 95.0%, helix-count ρ at least 0.912, and composite accuracy and ρ at least 86.7% and 0.652.Reversing the order of the two composite properties leaves both metrics similar.Source: Table 13 · p. 32, §E.1 · p. 32

E.2 DESIGNABILITY

• 79.7% of Prompteina designs are designable, versus 2.4–16.5% for the three text-guided full-suite baselines.Designability means scRMSD below 2 Å; each table row compares Prompteina with a baseline on the same evaluated test cases.Source: Table 14 · p. 34, §E.2 · p. 34
• Fine-tuned Prompteina reaches 79.7% designability with prompts and 85.2% without them, versus 29.1% for pretrained Complexa; the prompt advantage replicates with ESMFold2.The matched per-prompt median difference between prompted and empty-prompt designability is zero, and the ESMFold2 comparison is treated as replication, not better ground truth.Source: §E.2 · p. 34

E.3 STRUCTURAL NOVELTY

• Conditional Prompteina and Complexa have mean novelty of 0.230 and 0.217, while matched conditional and empty-prompt Prompteina samples have 0.230 and 0.165.The matched Prompteina comparison uses the same checkpoint, prompts, lengths, and reference set; cross-model values are descriptive because references differ and omit Teddymer.Source: Table 15 · p. 35, §E.3 · p. 35

E.4 FULL PROMPTS FOR FIGURE 1A

• The two Figure 1a prompt examples pair functional aims with detailed fold, secondary-structure, residue-interaction, or shape constraints.One specifies a mechanobiology sensor; the other describes an organophosphate-hydrolysing enzyme.Source: §E.4 · p. 35

F FUNCTIONAL SPECIFICATION EVALUATION

• ProTrek evaluation is not independent of Pinal’s training: ProTrek helped select its training pairs, structure tokens, and candidates.Prompteina uses neither Pinal’s ProTrek-derived data nor its selection procedures.Source: F FUNCTIONAL SPECIFICATION EVALUATION · p. 36
• Prompteina has the highest mean sequence and structure similarity at each method’s best prompt format; its advantages over BioMatrix are resolved for both metrics.The advantages over BioMatrix are +2.73 for sequence and +2.71 for structure similarity; the benchmark averages across 122 sequence-coherent protein families.Source: F FUNCTIONAL SPECIFICATION EVALUATION · pp. 35–37, Table 16 · p. 37
• Against Pinal, unweighted advantages of +0.91 for sequence and +0.26 for structure similarity remain unresolved under the paired family bootstrap.With confidence weighting, the sequence advantage resolves at +1.14, but the structure advantage of +0.72 remains unresolved.Source: F FUNCTIONAL SPECIFICATION EVALUATION · p. 37

G.1 INTERFACE-RESIDUE STEERING

• Across tested phrasings, requested interface-residue ordering is preserved with ρ ≥ 0.976; paraphrased and reordered prompts retain substantial response slopes.Paraphrases retain or increase the key–value slope on every target, while reordered prompts retain 76–109% of that slope.Source: Table 17 · p. 38, §G.1 · p. 38

G.2.1 TARGET PANEL

• The target panel has 20 targets; curated hotspots are available for 18, and the agentic setting lets each method select an epitope.Influenza H1 and the SARS-CoV-2 receptor-binding domain are excluded because prompt expansion produced no usable specifications; IL-17A and TNF-α remain despite Complexa runtime limits.Source: §G.2.1 · p. 38

G.2.2 GENERATION PROTOCOL

• Each method generates 200 designs per target in both agentic and non-agentic settings, with binder lengths sampled across 20 bands.Designs are sampled independently without inference-time search, reward models, or candidate selection, under a common generation budget.Source: §G.2.2 · pp. 38–39

G.2.3 SEQUENCE CONDITIONS

• Three sequence conditions compare model-generated sequences, ProteinMPNN redesign, and redesign with generated interface identities fixed.The gap between the two redesign conditions isolates how much of a method’s score is carried by ProteinMPNN rewriting the interface.Source: §G.2.3 · pp. 39–40

G.2.4 AGENTIC SPECIFICATION GENERATION

• A two-stage language-model pipeline researches targets and produces a structured dossier, then expands it into design specifications for each generation model.Research tools include web and literature search, PDB lookup, and UniProt lookup; the dossier covers target structure, binding biology, known binders, and biophysical properties.Source: §G.2.4 · p. 40
• Methods receive the same dossier, task instructions, and structural query tools; only the method reference describing its channels and accepted values differs.The task instruction specifies the number and diversity of designs and the length band.Source: §G.2.4 · p. 40, Table 18 · p. 40
• The agent encodes its chosen epitope through each method’s available channel and rejects specifications naming absent target residues.Proposed binder lengths are clamped to the target’s range; other agent output is passed through unchanged.Source: §G.2.4 · p. 40

G.2.5 CD45 SPECIFICATION EXAMPLES

• CD45 examples show Prompteina realizing diverse natural-language specifications, while all 200 BoltzGen specifications contain residue-level secondary-structure ranges.The examples illustrate available conditioning channels, not matched design strategies; the input interface determines which parts of the agent’s reasoning reach generation.Source: §G.2.5 · pp. 41–42, Figure 9 · p. 42

G.3 BASELINE AGENT SPECIFICATIONS

• Baseline specifications expose different conditioning channels: RFdiffusion3 accepts atom-level hotspots and hydrogen-bond partners, while BoltzGen accepts secondary-structure ranges and permitted amino acids.Agent-guided runs select epitopes for every design, whereas non-agentic runs use expert-curated hotspots on 18 targets; the table summarizes channel use across 4,000 specifications per method.Source: §G.3 · pp. 42–43, Table 19 · p. 43

G.4 BINDER SCORE CALIBRATION

• ipSAEmin has higher macro average precision than iPAE under every prediction model, while adding binder-aligned RMSD gives no consistent benefit.Adding RMSD lowers macro-F1 for three predictors, including ESMFold2-Fast (0.519 to 0.469), and raises it slightly only for AlphaFold2 (0.441 to 0.448).Source: §G.4 · p. 44

G.5 PREDICTION-MODEL REPLICATION

• Prompteina remains first under AlphaFold3, with a 0.084 margin over Complexa versus 0.055 under ESMFold2-Fast.The agent-guided and non-agentic BoltzGen ordering reverses between folding models, so the authors do not interpret that comparison.Source: §G.5 · p. 45

G.6 INTERFACE CONFIDENCE ACROSS DESIGN BUDGETS

• Prompteina leads at every design budget under self and MPNN-FI, but under MPNN it leads only in the middle of the range.RFdiffusion3 leads for k ≤2 and BoltzGen for k ≥64 under MPNN; no method leads at every budget in that condition.Source: §G.6 · p. 45, Table 22 · p. 45

G.7 GENERATED BINDER–TARGET INTERFACE PROPERTIES

• Prompteina’s median buried-surface-area lead over the best baseline narrows from 149 Å2 on generated backbones to 99 Å2 after MPNN-FI and 55 Å2 after MPNN.Among baselines, full ProteinMPNN redesign increases median interface hydrogen-bond counts by one to three, while other refolds change the median by at most one.Source: §G.7 · p. 46, Table 23 · p. 46
• Prompteina produces larger interfaces than structured-conditioning baselines (median buried surface area 1,020 Å2 vs. 710–871 Å2), but buried surface area explains only part of its score difference.ipSAEmin rises across buried-surface-area quartiles, yet no combination of measured per-design properties predicts ipSAEmin beyond r ≈0.4 on unseen targets.Source: §G.7 · p. 46, Table 24 · p. 46

G.8 BINDER SPECIFICATION ABLATION

• The full specification exceeds fold + shape by 0.091 expected top-k (95% CI [0.041, 0.145]) and scores highest at every selection budget.Fold + shape exceeds fold-only by just 0.002 (95% CI [−0.036, 0.037]); the authors attribute the improvement to specifying interface properties rather than simply adding more properties.Source: §G.8 · p. 47

G.9 STRUCTURAL DIVERSITY

• At k = 32, Prompteina is more diverse than several baselines across redesign conditions, but splits targets with agent-guided RFdiffusion3.Diversity is measured by pairwise TM-score, with lower similarity indicating greater diversity; the comparison retains 19 targets under MPNN and 18 under MPNN-FI.Source: §G.9 · pp. 47–48, Table 25 · p. 47
• Agent-guided BoltzGen varies least in helix fraction across targets (responsiveness 0.035) and is the only method whose pairwise similarity increases with k.The reported similarity rises from 0.463 at k = 8 to 0.494 at k = 64; secondary-structure variation and whole-fold diversity are distinct measures.Source: §G.9 · p. 48, Table 26 · p. 48

H.1 ITERATIVE DESIGN PROTOCOL AND ANALYSIS

• Each iterative loop uses the one-shot 200-design budget in four rounds of 50, with generation, redesign, and scoring otherwise unchanged.After each round, the agent receives all designs so far ranked by the objective, with scores across redesigned sequences and design-level attributes.Source: §H.1 · p. 48
• On Claudin-1, the selected round-1 design switches from phospholipase A2 to growth-hormone topology; round 2 keeps that topology and enlarges the requested interface.The requested interface rises from 20 to 24 residues and from 1500 to 1700 Å2 buried surface area.Source: §H.1 · p. 49, Table 27 · p. 49
• Both iterative baseline runs improve over matched-budget i.i.d. generation on Claudin-1, though by less than Prompteina.The agents extract coherent feedback patterns and use them to focus later batches; the authors suggest Prompteina offers more ways to revise designs.Source: §H.1 · pp. 49–50

H.2 FULL-PANEL EVALUATION

• An independent disjoint-sample check reproduced the same difficulty thirds; changing the ranking metric left the six hardest targets unchanged and shifted at most six targets near the middle–easiest boundary.The independent check verifies target grouping, not gain magnitudes.Source: §H.2 · pp. 50–51
• Prompteina gained most on the hardest third at k=1 and k=8, exceeding other methods there; BoltzGen peaked on the middle third at k=8, while RFdiffusion3 lost ground at k=1.Table 29 reports mean gain in top-k ipSAEmin over each method’s 200 i.i.d. designs, averaged within difficulty thirds.Source: Table 29 · p. 50, §H.2 · p. 50

I.1 CANDIDATE SELECTION

• After screening 1,000 backbones across five threads and 33 sequence variants each, filters retained 6,031 sequences; final ranking combined interface, fold-confidence, and refolding-agreement metrics.The final score aggregated predictions across ESMFold2-Fast and AlphaFold3.Source: §I.1 · pp. 51, 53
• Metric development sought both binder enrichment and rejection of non-binders from all-failure targets; 16 new candidates were tested individually after stating a physical hypothesis and negative control.Source: §I.1 · p. 51
• iPAEtarget-norm was the best-performing criterion for experimental selection, chiefly because it rejected non-binders from all-failure groups better than ipSAEmin.Source: §I.1 · p. 52
• iPAEtarget-norm compares interface PAE with within-target PAE to residues 8–20 Å away, then uses the 25th percentile to prevent a few confident contacts dominating.Source: §I.1 · p. 52

• Nine GM2A binders were identified among 89 loaded designs, including six medium and three weak binders.Source: §I.2 · p. 53
• Eight binders had fitted overall apparent KD values; the remaining weak binder had no fit, and the six medium-binder ESMFold2-Fast structures are computational predictions.Source: §I.2 · p. 53
• 107.4 nM was the strongest binder’s overall apparent KD (95% CI: 80.9–142.5 nM).Its replicate-level fits were 68.8, 129.1, and 137.6 nM.Source: Table 31 · p. 53, §I.2 · p. 53