ZeroFold: Protein–RNA Affinity Predictor

• ZeroFold is a transformer-based model that predicts protein–RNA binding affinity using pre-structural embeddings from Boltz-2, preserving ensemble information.
• It integrates separate protein and RNA encoders with a cross-modal attention mechanism to fuse sequence-derived signals without relying on explicit 3D structures.
• Demonstrated on the PRADB dataset with a Spearman correlation of 0.65, ZeroFold offers a fast, sequence-driven approach suitable for virtual screening of flexible biomolecules.

ZeroFold is a transformer-based model for predicting protein–RNA binding affinity directly from sequence-derived, pre-structural embeddings, introduced to address a central difficulty in structural biology: RNA molecules exist as dynamic conformational ensembles, so committing to a single predicted structure can discard binding-relevant information (Hanke et al., 24 Mar 2026). In ZeroFold, pre-structural embeddings are extracted from the Boltz-2 trunk before structure decoding, combined for protein and RNA through a cross-modal attention mechanism, and used to predict affinity as $pK_D=-\log_{10}(K_D)$ (Hanke et al., 24 Mar 2026). The model is trained and evaluated on PRADB, a curated dataset of 2,621 unique protein–RNA pairs with experimentally measured affinities drawn from four complementary databases, and on a held-out test set constructed with 40% sequence identity thresholds it achieves a Spearman correlation of 0.65 (Hanke et al., 24 Mar 2026). This suggests a representation strategy for flexible biomolecules in which ensemble-relevant information is retained without requiring explicit predicted structures.

1. Conceptual basis and problem formulation

ZeroFold addresses protein–RNA binding affinity prediction, where the objective is to infer the strength of interaction between a protein sequence and an RNA sequence, typically reported as

$pK_D = -\log_{10}(K_D).$

Higher $pK_D$ corresponds to tighter binding (Hanke et al., 24 Mar 2026). The problem is motivated by applications in understanding gene regulation, studying RNA-binding proteins and disease mechanisms, and designing RNA-targeting therapeutics such as RNA-binding peptides (Hanke et al., 24 Mar 2026).

The paper identifies four obstacles: RNA is highly flexible and often exists as a conformational ensemble rather than a single structure; many protein–RNA interfaces are shallow, heterogeneous, and dynamic; high-quality labeled affinity data are scarce; and structure-based methods usually require experimentally determined 3D structures, which are unavailable for most protein–RNA pairs (Hanke et al., 24 Mar 2026). The core hypothesis is that intermediate representations from a biomolecular foundation model, captured before structure decoding, may preserve ensemble-level information that would be lost by selecting a single predicted structure (Hanke et al., 24 Mar 2026).

These intermediate representations are termed pre-structural embeddings. In the ZeroFold formulation, they are taken from the final Boltz-2 trunk layer before the structure module generates coordinates, and are argued to implicitly preserve information about multiple possible conformations, evolutionary and contextual signals, and ensemble-level structure properties (Hanke et al., 24 Mar 2026). This is especially pertinent for RNA, which the paper characterizes as a flexible biomolecule for which single-structure models are inherently incomplete (Hanke et al., 24 Mar 2026).

2. Representation strategy and architecture

ZeroFold uses Boltz-2 to obtain pre-structural embeddings for both the protein and the RNA, then applies separate encoders, a cross-modal attention module, and a scalar affinity head (Hanke et al., 24 Mar 2026). The overall pipeline is: input protein and RNA sequences; run them through the Boltz-2 trunk; extract final trunk-layer embeddings before structure decoding; encode protein and RNA separately; fuse them via cross-modal attention; and predict affinity (Hanke et al., 24 Mar 2026).

The model uses two representation types from the final Boltz-2 trunk layer. The single or per-residue representations are

$$s_i^{\text{msa}} \in \mathbb{R}^{384}, \quad i \in \{1,\dots,N_{\text{res}}\},$$

and the pair representations are

$$z_{ij}^{\text{pair}} \in \mathbb{R}^{128}, \quad i,j \in \{1,\dots,N_{\text{res}}\}.$$

These are extracted from the final trunk layer of Boltz-2 with recycling_steps 3, before the structure module decodes coordinates (Hanke et al., 24 Mar 2026).

ZeroFold contains separate encoder blocks for protein and RNA. Each chain has a single-stream encoder and a pair-stream encoder, implemented as two successive transition layers with residual connections and dropout while preserving dimensionality (Hanke et al., 24 Mar 2026). The RNA encoder includes an additional nucleic acid type embedding distinguishing RNA from DNA. This is a learned 32-dimensional embedding concatenated to the single representation before the first transition layer, expanding the input dimension from 384 to 416; because of that dimensional change, the first layer is applied without a residual connection (Hanke et al., 24 Mar 2026).

After separate encoding, the model integrates protein and RNA features through a cross-modal attention module intended to model the interaction interface (Hanke et al., 24 Mar 2026). The paper characterizes the mechanism in standard transformer terms, where one modality attends to the other. A generic form is

$$\mathrm{Attn}(Q,K,V) = \mathrm{softmax}\left(\frac{QK^\top}{\sqrt{d_k}}\right)V,$$

allowing the model to aggregate pairwise relevance between protein residues and RNA nucleotides (Hanke et al., 24 Mar 2026). The final joint representation is passed to an affinity prediction head that outputs a scalar estimate $\hat{y}\approx pK_D$ (Hanke et al., 24 Mar 2026).

A notable design choice is that the Boltz-2 trunk is frozen during training. Only the downstream ZeroFold components are optimized: the protein encoder, RNA encoder, cross-modal attention module, and affinity head (Hanke et al., 24 Mar 2026). This makes the system a downstream predictor built on a fixed pretrained biomolecular foundation model.

3. PRADB dataset and curation procedure

To support training and evaluation, the authors construct PRADB, the Protein-RNA Affinity DataBase (Hanke et al., 24 Mar 2026). PRADB merges four sources: ProNAB, BioLiP2, the UTexas Aptamer Database, and PDBbind+ (Hanke et al., 24 Mar 2026). The raw union contains 4,510 protein–RNA pairs with affinity measurements, and after deduplication yields 2,621 unique protein–RNA pairs (Hanke et al., 24 Mar 2026).

The source composition is heterogeneous. ProNAB initially contains over 20,000 protein-nucleic acid affinity records; for protein-RNA, 5,323 entries were present, 5,173 had valid $K_D$ measurements, and after filtering to retain only natural amino acids and unmodified nucleotide bases, 3,588 protein-RNA pairs remained (Hanke et al., 24 Mar 2026). BioLiP2 contributes 425 protein-RNA complexes with affinity data; the UTexas Aptamer Database contributes 258 protein-RNA complexes with labeled affinity values; and PDBbind+ contributes 239 protein-RNA complexes and is regarded in the paper as the most reliable source due to high curation and structural information (Hanke et al., 24 Mar 2026).

For pairs appearing in multiple databases, PDBbind+ affinity values were preferred when available; otherwise, the median affinity from the remaining sources was used (Hanke et al., 24 Mar 2026). After resolving duplicates, PRADB contains 2,621 unique protein–RNA sequence pairs, 512 distinct proteins, and 1,411 distinct RNA sequences (Hanke et al., 24 Mar 2026).

The dataset exhibits broad length distributions. RNA lengths span a broad range with median 22 nt, and protein lengths span a broad range with median 171 aa (Hanke et al., 24 Mar 2026). The affinity distribution is approximately unimodal, centered around $pK_D \approx 7.1$ with standard deviation 1.35 (Hanke et al., 24 Mar 2026). This distribution supports a regression formulation rather than a coarse classification task.

4. Training protocol and leakage-controlled evaluation

A major emphasis of ZeroFold is fair evaluation under sequence-similarity control (Hanke et al., 24 Mar 2026). PRADB is split into training, validation, and test sets in an approximate 8:1:1 ratio (Hanke et al., 24 Mar 2026). To reduce leakage, protein sequences in different splits are required to have $<40\%$ sequence identity, the same criterion is applied to RNA sequences, and the coverage threshold is 80% (Hanke et al., 24 Mar 2026). The authors use MMseqs2 easy-cluster and then audit the splits with bidirectional MMseqs2 easy-search, because greedy clustering can miss leakage cases (Hanke et al., 24 Mar 2026). After auditing, 23 entries that still exceeded the identity threshold across splits were removed (Hanke et al., 24 Mar 2026). The final split sizes are 2,104 training, 210 validation, and 299 test examples (Hanke et al., 24 Mar 2026).

To mitigate over-representation of large sequence families, each training sample is weighted inversely to cluster size:

$pK_D = -\log_{10}(K_D).$0

where $pK_D = -\log_{10}(K_D).$1 denotes protein cluster $pK_D = -\log_{10}(K_D).$2 and $pK_D = -\log_{10}(K_D).$3 denotes RNA cluster $pK_D = -\log_{10}(K_D).$4; the weights are normalized to sum to 1 over the training set (Hanke et al., 24 Mar 2026). This reduces dominance by highly sampled families such as ribosomal proteins with rRNA (Hanke et al., 24 Mar 2026).

Training is conducted for 100 epochs, with each epoch comprising 900 training samples weighted by cluster representation (Hanke et al., 24 Mar 2026). The Boltz-2 trunk remains frozen, and only the downstream ZeroFold modules are trained on PRADB (Hanke et al., 24 Mar 2026).

5. Quantitative performance and comparative evaluation

On the held-out PRADB test set, ZeroFold achieves the following metrics: MAE $pK_D = -\log_{10}(K_D).$5, RMSE $pK_D = -\log_{10}(K_D).$6, Pearson correlation coefficient $pK_D = -\log_{10}(K_D).$7, and Spearman correlation coefficient $pK_D = -\log_{10}(K_D).$8 (Hanke et al., 24 Mar 2026). The paper emphasizes that this correlation level is near a practical upper bound imposed by experimental measurement noise, estimated at about $pK_D = -\log_{10}(K_D).$9–$pK_D$0 on correlation metrics (Hanke et al., 24 Mar 2026). This interpretation rests on the observation that identical protein–RNA pairs can have substantially different reported affinities across assays (Hanke et al., 24 Mar 2026).

The principal structure-based comparator is CoPRA, described as the leading structure-based predictor (Hanke et al., 24 Mar 2026). CoPRA reported PCC $pK_D$1 and SCC $pK_D$2 on PRA310 with 5-fold cross-validation, but its benchmark used a 70% protein-sequence identity threshold and relies on experimentally resolved or predicted structures, making the setting less strict and structurally advantaged relative to ZeroFold’s 40% threshold on both protein and RNA (Hanke et al., 24 Mar 2026). To compare more fairly, the authors define progressively stricter subsets relative to CoPRA’s PRA310 training data: $pK_D$3, all test examples; $pK_D$4, proteins not present in PRA310; $pK_D$5, maximum protein identity $pK_D$6; and $pK_D$7, maximum protein identity $pK_D$8 (Hanke et al., 24 Mar 2026).

The reported comparison is as follows:

Subset	CoPRA PCC	ZeroFold PCC	CoPRA SCC	ZeroFold SCC
A All	0.50	0.63	0.54	0.65
B Protein not in reference	0.47	0.67	0.52	0.68
C Max protein identity < 70%	0.46	0.71	0.55	0.69
D Max protein identity < 40%	0.22	0.54	0.23	0.52

These results indicate that CoPRA performance drops sharply as evaluation becomes fairer, whereas ZeroFold remains comparatively stable (Hanke et al., 24 Mar 2026). A plausible implication is that ZeroFold’s gains are not primarily due to overlap with comparator training data.

The principal sequence-based comparator is DeePNAP, described as the leading sequence-based method (Hanke et al., 24 Mar 2026). DeePNAP reported $pK_D$9 in its own paper, but when evaluated on PRA201 in the CoPRA study it fell to PCC $$s_i^{\text{msa}} \in \mathbb{R}^{384}, \quad i \in \{1,\dots,N_{\text{res}}\},$$0 and SCC $$s_i^{\text{msa}} \in \mathbb{R}^{384}, \quad i \in \{1,\dots,N_{\text{res}}\},$$1 (Hanke et al., 24 Mar 2026). Because DeePNAP was trained on ProNAB, which overlaps with PRADB, the ZeroFold paper again evaluates on filtered subsets (Hanke et al., 24 Mar 2026). The reported comparison is:

Subset	DeePNAP PCC	ZeroFold PCC	DeePNAP SCC	ZeroFold SCC
A All	0.46	0.63	0.35	0.65
B Protein not in reference	0.48	0.54	0.42	0.54
C Max protein identity < 70%	0.44	0.61	0.45	0.56
D Max protein identity < 40%	0.41	0.53	0.34	0.46

The paper states that ZeroFold consistently outperforms DeePNAP across all filtered subsets and all reported metrics (Hanke et al., 24 Mar 2026).

6. Affinity-band behavior, interpretation, and limitations

The test set is also stratified into three affinity bands: low, $$s_i^{\text{msa}} \in \mathbb{R}^{384}, \quad i \in \{1,\dots,N_{\text{res}}\},$$2; medium, $$s_i^{\text{msa}} \in \mathbb{R}^{384}, \quad i \in \{1,\dots,N_{\text{res}}\},$$3; and high, $$s_i^{\text{msa}} \in \mathbb{R}^{384}, \quad i \in \{1,\dots,N_{\text{res}}\},$$4 (Hanke et al., 24 Mar 2026). These thresholds derive from the dataset mean and standard deviation, with mean $$s_i^{\text{msa}} \in \mathbb{R}^{384}, \quad i \in \{1,\dots,N_{\text{res}}\},$$5 and SD $$s_i^{\text{msa}} \in \mathbb{R}^{384}, \quad i \in \{1,\dots,N_{\text{res}}\},$$6, so that $$s_i^{\text{msa}} \in \mathbb{R}^{384}, \quad i \in \{1,\dots,N_{\text{res}}\},$$7 and $$s_i^{\text{msa}} \in \mathbb{R}^{384}, \quad i \in \{1,\dots,N_{\text{res}}\},$$8 (Hanke et al., 24 Mar 2026).

The reported Spearman correlations are:

Affinity band	CoPRA	DeePNAP	ZeroFold
Low	0.09	0.14	0.10
Medium	0.38	0.02	0.50
High	0.27	0.20	0.28

The authors interpret this pattern as indicating that ZeroFold is especially effective at coarse discrimination across the affinity range, but that correlation drops within narrower bands, particularly among stronger binders (Hanke et al., 24 Mar 2026). They specifically note that SCC is only about 0.28 in the stronger-binding regime, and therefore characterize the model as better suited for virtual screening than for fine-grained lead optimization (Hanke et al., 24 Mar 2026).

This interpretation is tied to a broader claim about experimental noise and heterogeneous assay conditions. The paper notes that affinity measurements for identical protein–RNA pairs can vary due to pH, temperature, and methodology, and that such variability limits the maximum attainable correlation (Hanke et al., 24 Mar 2026). The article therefore frames ZeroFold not as solving affinity prediction in an absolute sense, but as approaching the empirical ceiling allowed by currently available labels (Hanke et al., 24 Mar 2026).

The paper is also explicit about its limitations. Protein–RNA affinity datasets remain small, amounting here to only a few thousand examples (Hanke et al., 24 Mar 2026). Measurement noise is substantial; the compiled labels arise from different assays, laboratories, and conditions (Hanke et al., 24 Mar 2026). Performance may vary across protein classes, RNA classes, and affinity ranges, and these subgroup-specific behaviors require more systematic study (Hanke et al., 24 Mar 2026). Fine rank-ordering among strong binders remains difficult, which constrains immediate use for lead optimization (Hanke et al., 24 Mar 2026).

7. Significance and relation to broader methodology

A central methodological claim of ZeroFold is that sequence-derived, pre-structural embeddings can substitute for explicit structural models when the biomolecules of interest are flexible and structurally heterogeneous (Hanke et al., 24 Mar 2026). In this framework, the value of the Boltz-2 trunk lies not in its final coordinate predictions but in the information encoded before a single structure is selected (Hanke et al., 24 Mar 2026). This shifts the representation problem from structure prediction to extracting informative latent states from a pretrained biomolecular model.

The computational implications are also emphasized. Both ZeroFold and structure-based alternatives require a Boltz-2 trunk forward pass, but structure-based methods then require structure decoding before affinity prediction, whereas ZeroFold bypasses that step and directly uses the pre-structural embeddings (Hanke et al., 24 Mar 2026). This makes the model faster and more suitable for virtual screening, proteome-wide affinity profiling, and high-throughput candidate evaluation (Hanke et al., 24 Mar 2026).

The broader scientific significance lies in the claim that pre-structural embeddings are particularly useful for flexible biomolecules such as RNA, where a single predicted conformation may be fundamentally inadequate (Hanke et al., 24 Mar 2026). This suggests possible extensions to mutation-effect prediction, binding-site identification, and ribonucleoprotein assembly modeling (Hanke et al., 24 Mar 2026). A plausible implication is that ZeroFold exemplifies a more general strategy for biomolecular prediction tasks in which latent representations from foundation models are used directly, rather than forcing all downstream inference through explicit coordinate generation.

Within protein–RNA modeling, ZeroFold is therefore best understood as a leakage-controlled, affinity-regression framework built on frozen Boltz-2 trunk representations, with separate protein and RNA encoders, cross-modal attention, and training on the curated PRADB dataset (Hanke et al., 24 Mar 2026). Its reported performance, particularly the test-set Spearman correlation of 0.65 under 40% sequence identity thresholds, positions it as a strong sequence-derived predictor in a domain where explicit structures are often unavailable and, for RNA, often incomplete even when predicted (Hanke et al., 24 Mar 2026).