Prediction of Residue-Level Base Preference for DNA-Binding Proteins Using a Multimodal Learning Framework

Abstract

Specific recognition between proteins and DNA governs essential cellular functions, yet the binding specificity of most DNA-binding proteins (DBPs) remains experimentally uncharacterized and high-throughput experimental methods still lag far behind the ever-growing protein data set. Here, we present 3M-BSIP (3-modal base-specific interaction predictor), a multimodal deep learning framework capable of predicting the DNA-binding interface and residue-level base preference for DBPs. 3M-BSIP constructs a holistic protein representation by integrating three distinct modalities: a protein language model capturing evolutionary conservation; a graph neural network encoding 3D structural features; and a hierarchical encoder characterizing physicochemical properties of protein surface pointclouds. Comprehensive benchmarking demonstrates that 3M-BSIP achieves SOTA performance on both interface-residue identification and base-preference predictions. Crucially, the model maintains robust performance on computationally constructed protein structures and is also tolerant toward those undergoing profound structural rearrangements upon DNA binding. Two real-world applications of 3M-BSIP demonstrate that our model not only can accurately predict the specific binding motif of structurally unknown transcription factors but also, more importantly, can guide rapid reprogramming of the base preference of DBPs based on a purely computational pipeline. Therefore, 3M-BSIP holds great potential in annotating the binding motifs of all DBPs in a high-throughput manner and designing new DBPs with desired base specificities.

1. Introduction

Recognition of specific DNA sequences by proteins is a fundamental process that orchestrates nearly all aspects of cellular life, from regulation of gene expression by transcription factors (TFs) to the maintenance of genomic integrity and epigenetic markings by DNA modification enzymes. − Dysregulation of these molecular recognitions could potentially lead to numerous human diseases, such as cancer and inherited genetic disorders. Thus, a deeper understanding of protein–DNA binding specificity is essential for both diagnostics and therapeutic development. − Unfortunately, current knowledge of the biological functions or even the binding motifs of DNA-binding proteins (DBPs) lags far behind their vast existing scales found in nature. The JASPAR database contains the binding motifs for only 53% of curated human TFs and 32% for Arabidopsis thaliana. This knowledge gap is compounded by the vast diversity of TFs across taxa45% of plant TF families are absent in mammalsunderscoring the lack of our understanding of DBP specificity for even model organisms. − Moreover, the universe of sequence-specific DBPs is much broader than TFs alone. A previous study found that over 22% of noncanonical DBPs exhibit sequence-specific DNA binding, including proteins from unexpected functional classes such as protein kinases and mitochondrial proteins. This disparity highlights an urgent need for more efficient and high-throughput methods to dissect the DBPs’ function and specificities.

Experimental methods have been pivotal in revealing the protein–DNA recognition mechanism, resulting in extensive binding affinity and target specificity data. Protein-binding microarrays (PBMs), chromatin immunoprecipitation sequencing (ChIP-seq), and systematic evolution of ligands by exponential enrichment with massively parallel sequencing (SELEX-seq) , methods have been applied to systematically characterize the specific DNA-binding motifs for thousands of DBPs, particularly for TFs. Moreover, crystallographic and cryo-EM methods could provide further atomic-level details regarding their interacting interfaces. Despite that, experimental approaches still face inherent challenges: in vivo methods like ChIP-seq are often limited by lower resolution and confounded by chromatin accessibility, while in vitro assays cannot exhaustively explore the full combinatorial sequence space of DBPs or recapitulate the cellular context. Critically, the scalability of these experimental methods remains a bottleneck relative to the sheer number of DBPs across all species. These issues underscore the necessity of computational approaches capable of learning and generalizing the complex principles of biomolecular recognition by leveraging the existing experimental data. The recent success of deep learning methods in modeling critical biomolecular interactions, including protein–protein and protein–ligand binding, − offers a promising avenue to investigate the protein–DNA recognition mechanism.

The computational pursuit of DBPs’ specific binding with DNA has evolved through several paradigms. Early approaches sought to infer binding specificity, typically represented as position weight matrices (PWMs), directly from protein sequence by leveraging homology within the same protein families for which extensive experimental data were available, such as C2H2 zinc fingers or homeodomains. − However, these methods are constrained by their reliance on well-characterized families and often fail to identify critical interface residues. More advanced method like rCLAMPS learns the probabilistic “recognition codes” by mapping interface residues to their corresponding positions within the PWM, offering atomic-level revelations of protein–DNA binding. Meanwhile, the increasing number of resolved DBP-DNA complex structures enabled the applications of structure-based deep learning methods on structural modeling (e.g., RosseTTAFoldNA and AlphaFold3) , or interface prediction (e.g., GraphBind and GeoBind). − A recent geometric deep learning-based model, DeepPBS, was designed to predict the PWM patterns for given protein–DNA structures. Likewise, a graph-based model, NA-MPNN designed by David Baker’s group, was developed to achieve a similar task. Despite these efforts, to date, predicting DNA-binding specificity from a sole DBP structure remains challenging. The development of such end-to-end protein-centric approach would avoid the usage of any DNA structure as prior knowledge that could potentially introduce large biases due to the highly flexible nature of DNA conformation.

Here, we develop a novel multimodal deep learning framework, 3M-BSIP (3-modal base-specific interaction predictor), to predict the residue-level base-preference heatmap for a provided DBP structure from either experimental determination or computational modeling. The prediction results quantify the intrinsic recognition propensity of each DBP residue toward four canonical nucleobases (A, T, C, and G). By integrating three distinct feature modalities via a cross-attention mechanism, 3M-BSIP has successfully learnt a general DNA-recognition pattern based only on DBP’s evolutionary and structural features, independent of any pre-specified DNA context (as required for DeepPBS and NA-MPNN). 3M-BSIP achieves accurate DNA-binding interface and residue-level base preference predictions. The necessity of this multimodal design was validated by comprehensive ablation studies. Importantly, 3M-BSIP is robust to the inputs that are either computationally constructed or structurally highly flexible upon DNA binding. Finally, we performed two real-world applications of 3M-BSIP: 1) the model accurately predicts the promoter binding motif of a newly identified bacterial TF; 2) more intriguingly, we computationally evolved a known Type IIB restriction endonuclease, BsaXI, to preferentially target to alternative DNA motifs (base-preference alteration). These successful attempts potentiate the broader applications of 3M-BSIP in binding-motif annotations of functionally unknown DBPs or DBP reprogramming with desired specificities.

2. Materials and Methods

2.1. Problem Formulation

This study formulates the prediction of protein–DNA interactions as a multitask learning problem. For a given protein with a length of L amino acids, we set out to learn a mapping function that takes the multimodal feature representation of each residue R _i (where i = 1, ..., L) as input and predicts three critical properties simultaneously: (1) Interface-residue prediction: To predict whether each residue R _i is located in the protein–DNA interacting interface, represented by a probability p(R _i ) ∈ [0,1], where a higher value indicates a greater likelihood of being an interface residue. (2) Contact-strength ranking: To rank interface residues by their binding intensity to DNA, represented by a score s(R _i ) for each residue, where a higher score indicates stronger contact strength. (3) Base-preference prediction: To predict the nucleotide preference for each interface residue R _i , represented by a four-dimensional probability distribution q(R _i ) ∈ R ⁴ over the four DNA nucleotide types.

2.2. Data Sets and Benchmarks

The model was trained and evaluated on three distinct data sets designed to assess its performance on experimentally determined structures, computationally predicted complexes, and external benchmarks (Figure S1).

2.2.1. Experimental Structure Data Set

The primary data set for training and testing was derived from the Protein Data Bank (PDB). To construct a comprehensive experimental structure data set, we first extracted all protein–DNA complex structures with a resolution <4.0 Å from the PDB before February 28, 2025. To ensure data quality and focus on typical DNA-binding domains, we only kept the complex structures with the protein length ranging between 40 and 600 aa and DNA-duplex longer than 8 bp. Notably, single-stranded DNA complexes are excluded in our work. The final data set contains 3,729 structures.

2.2.2. Computationally Modeled Data Set

To further evaluate the model’s performance on noncrystallographic data set, we also constructed a computationally predicted protein–DNA data set created by AlphaFold3³¹. Specifically, from the JASPAR CORE database (2024 release), we randomly selected 10 transcription factors (TFs) from each of the six major taxonomic groups (vertebrates, insects, fungi, nematodes, plants, and urochordates). For each TF, its canonical DNA-binding sequence was generated by selecting the nucleotide with the highest score at each position of its position specificity matrix (PSM), these selected nucleotides were then concatenated to form the consensus binding motif. We then employed AlphaFold3 to predict the apo forms of these 60 TFs (referred as AF3_Test_apo) as well as their corresponding DNA-bound complex structures (AF3_Test_holo).

2.2.3. External Benchmark Test Set

To examine the model’s generalization capability on an independent external benchmark, we adopted the DNA_178_test set from a previous study. This test set was originally constructed based on the BioLiP database, with a protein sequence identity threshold of 0.3 determined by the CD-HIT method. To ensure the complete independence of this test set from our study, we removed 47 structures from this test set that either overlap with our training set or contain protein chains exceeding 600 amino acids. The final test set contains 131 protein–DNA complexes (referred as Geo_Test_131).

2.3. Structure Preprocessing and Data Set Partitioning

All protein–DNA complex structures from the curated data sets were first subject to a uniform preprocessing procedure using PDBFixer v1.11, including the replacement of modified aa or nt with their chemically closest natural counterparts, fixing of any missing heavy atoms, and the removal of all water molecules, ions, and other small-molecule ligands.

A specialized deredundancy process was then performed for the experimental structure data set. Notably, different from former studies, we only removed those structures in which both protein and DNA sequences were identical and the RMSD of DNA backbone atoms, calculated after superimposing the corresponding protein structures, was less than 1.5 Å. This treatment allows us to save a large number of structurally similar proteins with either varied DNA sequence preferences or DNA-binding modes that might dictate their distinct base preference. As a result, we finally retained 2,714 protein–DNA structures.

The above structure data set was then partitioned into training, validation, and test sets. Notably, the DNA-binding specificity within the same structural family is often not conservedstructurally similar proteins may recognize remarkedly different DNA sequences through distinct binding modes, such as TFs. We therefore aimed to evaluate whether 3M-BSIP can generalize across binding modes within a shared structural scaffold. To ensure the model’s ability to generalize to unseen but structurally related proteins, we first performed hierarchical clustering on the above data set based on the proteins’ TM scores. Within each cluster, 80% of the cluster members were allocated to the training set, 10% to the validation set, and the remaining 10% to the test set, which ensures a proportional distribution of structurally similar proteins across the data subsets. The final training, validation, and test sets contain 2,091 (PDB_Train_2091), 255 (PDB_Valid_255), and 368 (PDB_Test_368) structures, respectively.

2.4. Multimodal Protein Feature Extraction

For each protein residue, we extracted three complementary feature sets to describe its sequence evolution information, three-dimensional geometry, and surface physicochemical properties. We reason that the effective fusion of these three distinct levels of information is key to accurately capturing the specificity of protein–DNA recognition.

2.4.1. Sequence and Structural Sequence Embeddings

To enable the language model to perceive local three-dimensional conformations while processing the sequence, we adopted a strategy that combines the aa sequence with its corresponding structural sequence (3Di). Specifically, we first employed the Foldseek method to extract the protein’s aa sequence and its 3D structural sequence (3Di) from the coordinate file. The 3Di sequence encodes the protein’s 3D structural information into a one-dimensional string, describing the local geometric environment of each residue. Subsequently, we interleaved aa sequences and 3Di sequence to form a more information-rich input sequence. This combined sequence was then fed into a structure-aware pretrained large protein language model, SaProt (650M). We selected SaProt due to its pretraining regimen that integrates extensive structural knowledge. Finally, we extracted the last hidden-state layer as the contextual sequence embedding for each residue.

2.4.2. Geometric Graph Representation

To capture the internal geometric relationships within the protein, particularly the intrinsic geometric features that are invariant or equivariant to global rotations and translations, we represented each protein’s C_α-backbone as an undirected graph, G _s = (V,E). In this graph, nodes v _i ∈ V represent residues, and an edge e _ij ∈ E connects a pair of residues if their C_α atoms are within a spatial distance of 6.0 Å. The following features were used for nodes and edges that encode equivariant information:

• Node Features: The features for each node (residue) consist of two components: (1) the backbone dihedral angles (ϕ, ψ), which describe the local conformation of the peptide chain, and (2) local coordinate frame, which is a 3 × 3 orthogonal coordinate matrix constructed from the N, C_α, and C backbone atoms of each residue. This local coordinate frame establishes a residue-centric, rigid reference system. By representing neighboring geometries within this invariant frame, the model can learn features that are equivariant to global rotations or translations.

• Edge Features: The features connecting two nodes (residues i and j) are also composed of two parts: (1) the Euclidean distance between their C_α atoms and (2) the relative direction. The relative direction is obtained by projecting the vector connecting residue to residue onto the local coordinate frame of residue i. This representation nullifies the effect of global protein rotation on the edge features, enabling the model to learn the fundamental relative spatial arrangement between residues.

2.4.3. Optimization-Based Molecular Surface Generation and Featurization

To provide high-quality input for the downstream geometric deep learning model, we employed an optimization-based strategy to create protein’s molecular surface. Compared to traditional discretization methods like Marching Cubes, our strategy yields much smoother pointclouds with more accurate normal vectors, thereby finer capturing the geometric and physicochemical properties of molecular surface. The entire process comprises two stages: pointcloud generation and featurization.

• Pointcloud Generation: We employed dMaSIF to generate protein surface pointclouds. We chose this method for two primary reasons: first, it generates a smooth surface via a differentiable optimization process, providing an ideal basis for the subsequent extraction of precise geometric features; second, its high computational efficiency is well suited for the task of protein–DNA interface prediction, which requires scanning the entire protein surface.

• Surface Point Featurization: Based on the generated pointclouds, we designed and implemented a comprehensive featurization scheme to capture the rich geometric and physicochemical information crucial for protein interactions, including:

  • Geometric Features: It is composed of the 3D coordinates of the points, its unit normal vector, and the local curvature estimated from the variation of normal vectors in the local neighborhood.
  • Chemical Features: The local charges were included, calculated by distance-weighted interpolation of charges from neighboring atoms, and the atom types of the nearest atoms (one-hot encoded for C, N, O, and S).
  • Local Environment Features: Comprising the distances to its K-nearest neighbors (invariant feature); the angles between its normal vector and those of its neighbors (invariant feature); a local coordinate system constructed for each point (equivariant feature); and the relative positions of its K-nearest neighbors within this local frame (equivariant feature).

2.5. Generation of Supervisory Signals

To generate quantitative, multichannel supervisory signals for model training, we developed a pipeline to construct a four-channel residue-level base-preference heatmap for each protein. This map has a L × 4 (where L is the protein length) dimension and quantifies the interaction propensity of each residue with four DNA nucleotides (A, T, C, and G), detailed below.

• Identification of the Core DNA Segments: As the crystal structure of the protein–DNA complex typically contains flanking DNA sequences, we first extracted the core DNA segments that are directly involved in contacting with DBPs. For each nucleotide position j along a given DNA strand, we computed the number of protein–DNA heavy-atom pairs, c _j , within a Euclidean distance threshold of τ = 4.5 Å. The core segment [s*,e*] was then defined as the contiguous window of a predefined length L _n (equals to 8 bp in our model, which covers most of the DBP binding motifs) that maximized the sum of the protein–DNA contact counts:

  $$(s^{*},e^{*})=arg\underset{s}{\max }\sum _{j=s}^{s+L_n-1}{c}_j$$
• Chemical-Aware Distance Metric: We defined a chemical-aware distance, d _chem(a,b), between a protein atom a and a DNA atom b by multiplying the geometric distance by a discount factor α that is determined by the interaction type:

  $$d_{chem}(a,b)=\parallel {\mathrm{x}}_a-{\mathrm{x}}_b{\parallel }_2·\alpha (t_{\mathrm{a}},t_{\mathrm{b}})$$

  Here, x_a and x_b are the atomic coordinates, t _a and t _b represent the set of chemical types assigned to each atom (e.g., hydrophobic, hydrogen bond donor/acceptor, and charged), and α­(t _a ,t _b ) is a predefined discount factor, for which the hydrophobic, HB, and salt-bridge interactions are attributed to a value of 0.85, 0.82, and 0.8, respectively. The minimum distance between a protein residue R and a DNA nucleotide N, denoted as d _min (R,N), was then determined by minimizing d _chem (a,b) over all heavy-atom pairs:

  $$d_{min}(R,N)=\underset{a\in R,b\in N}{\min }{d}_{chem}(a,b)$$

2.5.1. Derivation of Residue Score and Base Allocation Coefficient

For each protein residue R, we first identified its closest nucleotide partner, N*­(R), within the core DNA segment according to d _min (R,N). If d _min (R,N*­(R)) > 4.5 Å, the residue R was labeled as a noninterface residue, and its corresponding base-preference heatmap vector was set to zero. Otherwise, d _min (R,N*­(R)) was transformed into a continuous, residue geometry score s _geo (R), representing each residue’s contribution to DNA binding, via a Gaussian kernel with σ = 3.0 Å:

$$s_{geo}(R)=\exp \left(\frac{-d_{min}{(R,N^{*}(R))}^2}{2{\sigma }^2}\right)$$

• To further quantify the residue’s preference for the four nucleotides, we computed the minimum geometric distance, b(R,N), between the side-chain heavy atoms of residue R and the base heavy atoms of nucleotide N with a cutoff distance of τ = 4.5 Å, yielding b̃(R,N) = min(b(R,N),τ). Based on this proximity measurement and the true base type y(B*) ∈ A,T,C,G of the partner nucleotide, we formulated a 4-dimensional logit vector l(R):

  $$l_y=\frac{12}{1+\exp (\mathrm{b̃}/2)},l_{i\ne y}=\frac{2}{1+\exp ((4-\mathrm{b̃})/2)}$$

  This formulation assigns a significantly higher logit to the correct base, creating a strong supervisory signal. The logits were then converted into a probability distribution over the four nucleotides using the softmax function, yielding the residue-nucleotide-level base allocation coefficient p(R) = sofma(l(R)).

2.5.2. Generation of Residue-Level Base Preference Heatmap

The final four-channel supervisory signal (i.e., contact strength) for residue R, denoted h(R), was obtained by element-wise multiplication of the residue score and the base allocation coefficient:

$$h(R)=s_{geo}(R)\odot p(R)$$

This operation uses base allocation coefficient to allocate the residue geometry score s _geo (R) into residue-nucleotides level. Stacking these vectors for all residues resulted in the final L × 4 matrix being used as the ground truth for training.

2.6. 3M-BSIP Model Architecture

The proposed 3M-BSIP model is an end-to-end, multitask learning framework (Figure ). Its information flow follows four main stages: (1) Feature input: The model receives the three preprocessed feature modalities as described in the previous section. (2) Parallel encoding: These three feature sets are fed into three specialized encoders to extract high-dimensional representations for each modality. (3) Model fusion: The outputs from each encoder, depicted by residue-level feature vectors, are passed to a sequence-centric cross-attention module to generate a unified multimodal feature representation. (4) Prediction heads: The final fused features are passed to three parallel prediction heads to perform multitask predictions.

2.

Architecture of 3M-BSIP. (A) Multimodal feature representation of the input protein. Three distinct feature sets are generated: a geometric graph representing the C_α backbone structure; sequence features comprising both the primary aa and 3Di sequences; and a featurized surface pointclouds. (B) Parallel encoder modules for feature extraction. Each feature modality is processed by a specialized encoder: GVP-GNN for the structure graph; SaProt PLM for sequence features; and GAT coupled with a transformer layer for the surface pointclouds. (C) Modal fusion via cross-attention mechanism. The encoded representations from the three parallel streams are projected into a common latent space and then integrated using a cross-attention mechanism to generate a unified feature representation for each residue. (D) Multitask prediction output. The final fused representation is fed into three parallel prediction heads for desired prediction tasks: namely, interface-residue classification, contact-strength ranking, and base-preference classification. These information can finally be mapped onto the protein structure to guide further structural modeling.

2.6.1. Geometric Graph Encoder

To process the SE(3)-equivariant graph features as described above, we employed a graph neural network based on the Geometric Vector Perceptron (GVP) architecture. This encoder iteratively updates the representation of each residue through 3 layers of message-passing layers. In each layer, the model aggregates geometric information from the spatially neighboring residues and their connecting edges. We construct a deep and stable network by applying residual connections and GVP-specific layer normalization between layers, thereby learning complex geometric dependencies within the protein.

2.6.2. Molecular Surface Encoder

This encoder adopts a hierarchical design to first capture the fine-grained local topology of the surface pointclouds and then effectively aggregate it into a global, residue-level representation. Its workflow consists of two stages: We first construct a graph from the surface pointclouds, where each point is a node. The raw 46-dimensional feature vector for each point is projected into a hidden dimension via a linear layer. Then, a stack of 3 graph attention network layers is employed to propagate information across neighboring points. Each GAT layer utilizes 8 attention heads and incorporates residual connections. Layer normalization and an ELU activation function are applied after each GAT layer. Then, for each residue, we aggregate its features by mean pooling over the surface points belonging to this residue. If a residue contains no surface points, then a learnable default feature vector was assigned. These aggregated residue-level representations are then fed into a 2-layer transformer encoder, which captures the long-range dependencies among the surface properties of different residues via a self-attention mechanism, producing the final output of the surface encoder.

2.6.3. Multimodal Fusion Module

To effectively integrate information from three different modalities, we designed a sequence-centric fusion module. Through two independent cross-attention operations, the sequence features “query” the geometric graph features and the surface features, respectively. Each operation first refines the sequence features via a self-attention block and then uses these refined features to query the other modality via a cross-attention block. This two-step process selectively absorbs and integrates key information from the structural and surface modalities. Finally, the original sequence features and the geometry- and surface-enhanced sequence features are concatenated and passed through a linear layer to output a highly condensed and fused representation.

2.6.4. Prediction Heads

The final fused features are passed simultaneously to three structurally identical but parametrically independent multilayer perceptron (MLP) prediction heads for the multitask learning of interface-residue classification, contact-strength ranking, and base-preference prediction.

2.7. Model Training and Evaluation

2.7.1. Training Strategy and Implementation Details

The model was implemented using Torch 2.5.1+CUDA 12.1 and DGL 2.4.0. We used the Adam optimizer by applying different learning rate for the SaProt backbone compared to that of the downstream modules (5e-5 vs 5e-4), along with a weight decay of 1e-5. The training process incorporated several optimization strategies. A linear learning rate warmup was applied for the first 5 epochs; subsequently, the learning rate was managed dynamically, halving whenever the validation loss did not improve for 5 epochs with a minimum value of 1e-6. All experiments were conducted on computer nodes equipped with NVIDIA V100 (32GB) GPUs.

2.7.2. Loss Function

Model training was guided by a composite loss function L _total , defined as the weighted sum of three distinct components, each targeting a specific prediction task:

$$L_{total}=w_{cls}{L}_{cls}+w_{pref}{L}_{pref}+w_{rank}{L}_{rank}$$

• Interface-residue classification loss ( L _cls ): To address the inherent class imbalance between interface and noninterface residues, we employed a weighted binary cross-entropy loss. A positive class weight (pos _weight = 1.5) was applied to increase the penalty for false negatives, encouraging the model to effectively identify true interface residues.

• Base-preference loss (L _pref ): For residues correctly identified as part of the interface, a standard cross-entropy loss was used to supervise the base preference prediction. This loss measures the dissimilarity between the predicted 4-dimensional probability distribution and the true base preference heatmap, guiding the model to learn specific recognition patterns.

• Contact-strength ranking loss (L _rank ): Recognizing that the relative importance of interface residues is more critical than their absolute contact strength values, we formulated this task as a listwise ranking problem instead of a simple regression. We implemented a ranking loss based on the LambdaRank algorithm, specifically designed to directly optimize NDCG. This λ loss focuses on ensuring that the residues ranked highest in the prediction correspond to those with the strongest true contact strength, particularly within the top-k predicted residues. This approach is more robust and aligns better with the biological goal of identifying key functional hotspots at the protein–DNA interface.

The weights for each component (w _cls = 2, w _pref = 1, w _rank = 1) were used to balance their respective contributions to the total loss during training.

2.7.3. Evaluation Metrics

We evaluated the model’s performance according to the prediction accuracies for interface residues and base preferences. The former was assessed by checking the model’s ability to discriminate the interface and noninterface residues. A residue is considered as an interface residue if it contacts any of the four nucleotides (A, T, C, and G). The latter was evaluated by predicting each potential residue–nucleotide interaction pair. Additionally, we assessed the model’s ability to rank residues by their contact strength using Normalized Discounted Cumulative Gain (NDCG).

2.8. A General Computational Screening Pipeline for Potential Mutations Leading to Specificity Alteration

To identify potential mutation sites that could regulate the DNA specificity of DBPs, we establish a computational screening pipeline by integrating 3M-BSIP-based base-preference prediction with physics-based energy evaluations. For a given DBP complex, the protein–DNA contacts are defined using the chemical-aware minimum heavy-atom distances between protein residues and DNA nucleotides within an 8 bp recognition region. Residues within 4.5 Å of any core DNA atoms are classified as contacting residues, and their closest nucleotide identity is assigned as the ground-truth targeting nucleotide. Prior to mutation screening, 3M-BSIP should be first applied to the WT structure to predict the contact strength and the residue-level base-preference heatmap; random-mutation screening proceeds only when at least one key ground-truth interaction is reproduced by the model, ensuring that the model has predictive capacity on the target complex.

Candidate protein mutants can be generated using Rosetta 3.14 by randomly sampling 1–3 residues within an 8 Å neighborhood of each validated DNA-binding residue and substituting them with alternative amino acids across 1,000 independent trials. Each generated mutant is then subjected to local relaxation using FastRelax with a MoveMap restricted to the mutation sites and their 15 Å neighbors. Only those mutants exhibiting a Rosetta full-atom energy (using REF2015 as the scoring function) lower than −1.0 kcal/mol relative to the wild type are retained to ensure structural stability. Each accepted mutant is finally processed by 3M-BSIP for base-preference prediction. The mutants predicted with an altered DNA specificity are designated as primary hits. For these hits, we can further evaluate their binding energy differences between the original and altered DNA sequence using Rosetta and only keep the ones with favorable binding to the altered sequence. For BsaXI, the initial Rosetta mutation structural-stability filter retained 754 mutants. These candidates were evaluated by 3M-BSIP, which identified 77 mutants with predicted specificity alterations. A subsequent Rosetta binding-energy filter retained 23 candidates.

3. Results

3.1. Structural Characterizations of Protein–DNA Interface for DBPs

We aim to develop a multimodal deep learning method to predict, for a given DBP sequence/structure, its DNA-binding interface and residue-level base preference. The model was trained based on a high-quality structural data set consisting of 2,714 protein–DNA complexes derived from protein data bank (PDB) (Figure A). For each protein–DNA complex, an amino acid is determined as an interface residue if any of its heavy atoms are within 4.5 Å of any DNA heavy atoms. To better evaluate the protein–DNA interactions, we introduced a “chemical-aware distance”, d _chem (a,b), to refine the geometric distance between any two atoms a and b by applying a discount factor α based on the interaction type, i.e., hydrophobic, hydrogen bond or electrostatic interactions, with a value of 0.85, 0.82, and 0.80, respectively, which is determined by the relative strengths of these three interaction types. − We then derived a continuous “residue-geometry score” s _geo (R) for each interface residue based on the minimum d _chem (a,b) between each pair of interacting residue and nucleotide. The residues with s _geo (R) ≥ 0.25 are designated as “hotspot interface residues”, which are presumably critical for DNA binding or sequence-specific recognition, while others are considered as “passive interface residues” (Figure B).

1.

Definition of the chemical-aware distance and structural characterization of the DBP data set. (A) Schematic of the data set construction. (B) The quantitative chemical-aware framework. Left: An illustration of the “chemical-aware distance”, which refines the Euclidean distance by applying an interaction discount factor α according to the interaction type between atom types t _a and t _b . Right: The formulas define the chemical-aware distance d _chem (a,b), the continuous residue geometry score s _geo (R) and classification of interface residues. (C) Distribution of top 25 protein families within the curated data set. Blue bars indicate the number of structures for each protein superfamily. Golden and green bars represent the calculated ratios of passive and hotspot interface residues for each family, respectively. (D) Representative structures from four distinct superfamilies. (E) The DNA-binding motifs of the protein families shown in (D), visualized as sequence logos. (F) Background-normalized residue-nucleotide contact enrichment. Top: amino-acid contact counts and log2 enrichment heatmap for all interface contacts. Bottom: amino-acid contact counts and log2 enrichment heatmap for hotspot interface contacts.

Our collected structural data set consists of ∼230 discernible DBP superfamilies, classified by InterPro, including polymerases, restriction endonucleases, and TFs (see Figure C for the top 25 superfamilies). As expected, the sequence-independent members, such as polymerases and reverse transcriptases, exhibit considerably less hotspot interface residues comparing to the well-known sequence-dependent members, such as homeodomain-like, p53, and restriction endonucleases (Figure D–E). The nucleotide contacts of interface residues were characterized across the curated DBP data set. For each residue-nucleotide pair, a log2 enrichment score was calculated after reweighting by the A/T/C/G background composition of the DNA-binding regions (Figure F). This analysis revealed distinct nucleotide contact enrichment patterns for several amino acid types in the interface. These patterns became more selective among hotspot residues, indicating that the hotspot subset is enriched for interactions that contribute to base-preference recognition.

We then analyzed the base-preference patterns of residue-nucleobase contacts separately from those of backbone-only contacts (Figure S2). The enrichment pattern of base contacts overlapped only partially with that observed for the hotspot residues. Thus, the base preference of some hotspot residues may be associated with contacts to the DNA backbone, in which residues recognize sequence-dependent DNA conformation through backbone interactions. These findings suggest that our defined chemical-aware distance can accurately pinpoint the critical DBP residues (namely, hotspot interface residues) responsible for base recognition, via either forming direct nucleobase contacts or establishing backbone interactions.

3.2. The Multimodal Architecture of 3M-BSIP

Building upon the chemical-aware distance framework derived from our training data set, we developed 3M-BSIP (3-Modal Base-Specific Interaction Predictor), a deep learning model aiming to predict both the DNA-binding interface and base preference across DBP families. The model’s architecture is founded on the central hypothesis that the evolutionary conservatism, folding patterns, and physicochemical landscape of DBPs collectively determine the DNA-recognition preference.

To achieve this, the model first generates three distinct, complementary feature sets for the input protein (Figure A): 1) Primary aa sequence combined with a 3D-interaction (3Di) description that discretizes the local structural features for each residue; 2) graph representation of the internal protein structure, in which the C_α atoms form nodes, attributed with dihedrals (Φ,Ψ) and local coordinate frames; while the edges are attributed with inter-residue distances and relative orientations; 3) Pointcloud representation of the protein surface, with each point annotated with its local geometry (curvature etc.), physicochemical properties (charge and atom type, etc.), and the surrounding neighbors (see Materials and Methods 2.4 for more details of the feature selections).

The above feature sets were then fed into three parallel encoders, each specialized for its respective data type (Figure B). The structure-aware sequence feature is processed using a pretrained protein language model SaProt that uses an ESM-2-like architecture to capture long-range dependencies in the sequence. Meanwhile, the internal structure graph feature is processed by a geometric vector perceptron graph neural network (GVP-GNN). This SE(3)-equivariant architecture ensures that the predictions are robust to arbitrary rotations and translations of the input structure. The protein surface pointclouds are processed by a specialized graph neural network. First, a graph attention network (GAT) learns representations of local chemical patches at the point level. Then, point-level features are aggregated into a residue-level representation via mean pooling over surface points within a certain residue. Finally, a two-layer transformer encoder captures the long-range dependencies between residue-level features (Figure S3).

The encoded feature representations were finally integrated via a cross-attention fusion module (Figure C, Figure S4), which allows efficient information communication between multimodal features, e.g., the significance of a specific residue in the sequence can be contextualized by its structural position and surface accessibility. Finally, the integrated representation is passed to three distinct output heads to simultaneously perform multitask predictions, i.e., interface-residue identification, interface-residue contact-strength ranking, and its base preference (Figure D). The resulting predictions can be finally mapped onto the input protein structure, thereby guiding the structural modeling of the protein–DNA complex (see the following section). This end-to-end architecture built on multimodal information empowers the predictive accuracy of 3M-BSIP.

3.3. Ablation Studies Validate the Necessity of Multimodal Integration

To evaluate the model’s capacity to handle the binding-motif diversity within related structural families, we benchmarked its performance on the test-set DBPs against their DNA-binding motif divergence from the most structurally analogous proteins within the training set. We defined the sequence divergence using the average Levenshtein distance between the binding motif of a test protein and those of its five most structurally similar proteins in the training set (higher TM-scores). As shown in Figure A, a clear negative correlation was observed for the interface-residue prediction task, with the Matthews Correlation Coefficient (MCC) dropping from approximately 0.94 to 0.36 as the Levenshtein distance increased from 0 to 7.2. Similar trends were also observed for other two tasks: base-preference prediction and contact-strength ranking, the latter was measured via NDCG@5 (Figure B–C). Notably, our model performs fairly well even when the average Levenshtein distance reaches up to 4 (equivalent to 4-bp difference between two DNA sequences), with the performance metrics above 0.6 for all three tasks. These results demonstrate the model’s resilience to the evolutionary distance of the test samples from the training set.

3.

Model performance and ablation evaluation. (A–C) Analysis of model generalization on the test set DBPs against the sequence divergence from the most structurally analogous proteins within the training set. Results of interface-residue prediction (A), base-preference prediction (B), and contact-strength ranking (C) are plotted. The DNA-sequence difference is measured by the average Levenshtein distance. The red line indicates the linear regression fit, with the shaded area representing the 95% confidence interval. (D–F) Ablation studies were conducted to evaluate the contributions of different feature components. The complete model, serving as a baseline, is compared with single- or dual-modality variants (D). Moreover, different fusion strategies (E) and different central modalities used for the cross-modal fusion (F) were also compared. Model performance is measured based on five metrics: interface-residue MCC, interface-residue PR-AUC, base-preference accuracy (for interface or hotspot residues), and NDCG@5 for contact-strength ranking. (G–H) Assessment of model robustness across protein family (G) and aa type (H).

Next, we conducted rigorous ablation studies to dissect the contribution of each feature component to the model performance by employing the complete 3M-BSIP model as the baseline. First, we validated the necessity of a multimodal approach by comparing the complete model to single- or dual-modality variants. Treating the sequence-only model as a foundational configuration, we found that the inclusion of either structural or surface information could profoundly improve the model performance, particularly for the base-preference prediction (Figure D). More strikingly, including both feature sets would greatly enhance the model performance across all tasks, with the base-preference accuracy increasing from 0.44 to 0.60 (Figure D). This finding demonstrates that all three feature streams provide indispensable and complementary information. Further evaluation of various fusion strategies suggests that cross-attention mechanism outperforms other tested methods, including simple concatenation and other advanced methods (Figure E, see Supporting Information Section 1 for more details). Moreover, employing the sequence feature as the central query to attend to structural and surface information results in the best model performance (Figure F). Notably, the final model works consistently well across all tested metrics, achieving an interface MCC of 0.65, an interface base-preference accuracy of 0.60, and NDCG@5 of 0.67.

Finally, to examine the model’s robustness and influences of training data imbalance, we analyzed the model’s performance across different protein families and aa types. Notably, our model performs reasonably well for the majority of DBP families, while it gives relatively poor predictions for a few chromatin-associated or PTM-dependent protein families such as P-loop NTPases, HMG box domains, and signal response regulators (Figure G), including proteins with limited sequence specificity or atypical DNA-recognition modes, such as conformation-dependent binding, which are not fully represented by A/T/C/G labels. In addition, our model shows a general good performance across all aa types, with the interface base-preference accuracy exceeding 0.50 for all 20 aa types (Figure H). These findings suggest that 3M-BSIP has learned the generalizable biophysical principles of protein–DNA recognition rather than merely overfitting to patterns of specific, over-represented classes during training.

3.4. 3M-BSIP Outperforms the GeoBind Method on Interface-Residue Prediction

We compared the performance of 3M-BSIP with that of the state-of-the-art method GeoBind on the interface-residue prediction task. The evaluation was conducted on the Geo_Test_131 benchmark, a nonredundant subdata set derived from the original GeoBind test set by removing the entries that either overlap with our training data or bind with nondouble-stranded DNA. The result shows that 3M-BSIP surpasses GeoBind on the majority of tested samples, particularly for smaller proteins, which is likely because 3M-BSIP was primarily trained on proteins less than 600 aa length (Figure A). Analysis of subsets with varying sequence divergence shows that 3M-BSIP outperforms GeoBind on all tasks for all subsets, although this advantage diminishes as protein sequence divergence increases (Figure B). Specifically, 3M-BSIP achieves a higher interface MCC on the complete test set (0.545 vs 0.523), and the difference becomes less for proteins with a Levenshtein distance of >5 (0.518 vs 0.512). A similar trend can be observed in the contact-strength ranking of NDCG@5, for which 3M-BSIP demonstrates profound advantages over GeoBind at all levels of sequence divergence (GeoBind ranks each residue based on its contact probability score). Notably, 3M-BSIP could still achieve a superior performance on the base-preference prediction, although the accuracy decreases as the Levenshtein distance becomes larger (Figure B). Finally, we also compared their performances on nonexperimental structures via creating an external AF3_Test_holo data set comprising 60 AlphaFold3-constructed protein–DNA complexes. On this data set, 3M-BSIP again outperforms GeoBind on both interface-residue prediction and contact-strength ranking tasks. For the hotspot base-preference prediction, which is inaccessible in GeoBind, 3M-BSIP achieves an accuracy of 0.58 (Figure C).

4.

Benchmark performance and robustness of 3M-BSIP. (A) Comparison of 3M-BSIP with GeoBind on the Geo_Test_131 benchmark set for interface-residue prediction task. Each point represents a protein complex, colored by the length. (B) Performance comparison on Geo_Test_131 and its subsets divided by the DNA sequence divergence. Comparisons of 3M-BSIP with GeoBind were performed only for the interface-residue and NDCG@5 predictions, while the base-preference accuracy was reported only for 3M-BSIP. (C) Performance on the external AF3_Test_holo data set, which consists of 60 AlphaFold3-predicted protein–DNA structures. (D) Correlation analyses of 3M-BSIP between apo and holo forms of the input protein on three prediction tasks. Structures derived from experiments and computational modeling are colored in yellow and blue, respectively, scored by the corresponding TM-score between the apo and holo structures. The ZnF456 protein, detailed in (E), is marked with asterisks. (E) Case study of the transcription factor ZnF456, for which both its holo (PDB: 6KI6) and apo forms (PDB: 8THO) are available, different zinc finger domains are colored in cyan, blue, and purple, respectively. The interface heatmap derived from 6KI6 serves as a ground-truth (top), while the prediction was conducted based on 8THO (bottom), discrepancies between the two are highlighted with red rectangular boxes.

3.5. Well-Tolerance to 3M-BSIP toward the Structurally Flexible DBPs

DBPs may undergo substantial structural arrangements upon DNA binding. One critical challenge for an end-to-end structure-based prediction model is thus its robustness to input DBP conformations. We therefore further evaluated the performance of 3M-BSIP on both apo (unbound state) and holo (DNA-bound state) forms of DBPs for two different data sets: one consisting of 54 proteins, for which both their apo and holo forms are available from experimental methods; the other includes 60 proteins, for which their apo and holo forms were computationally predicted by AlphaFold3. Our results show that regardless of using the apo or holo form as the input structure, all three prediction tasks could reach comparable accuracy levels, with the correlation value R² > 0.71 (Figure D). This finding indicates that our model succeeded in learning the intrinsic interface properties of DBPs despite the induced structural changes upon DNA binding.

As a showcase, we chose a TF, ZnF456, for more structural analysis. Its apo (PDB: 8THO) and holo (PDB: 6KI6) forms have a relatively low TM-score (0.369), suggesting its striking structural changes upon DNA binding (Figure E). Nevertheless, using the apo form as an input, our model could still achieve an interface-residue MCC of 0.58, a base-preference accuracy of 0.60, and NDCG@5 of 0.86 (marked with asterisks in Figure D).

Comparison of the predicted base-preference heatmap with the holo form (ground truth) reveals that most of the protein–DNA contacts could be successfully captured by our model, only two residues (F16 and S28) are wrongly assigned and three (S47, S48, and T56) fail to be identified as interface residues. Notably, although F16 was wrongly identified as an interface residue, it is still close to the DNA backbone (∼5.1 Å) in the holo structure; S28 was predicted to recognize adenine, whereas the holo structure shows that it interacts with the backbone of one guanine; the remaining three sites (S47, S48, and T56) are predicted as noninterface residues (Figure E). Structural inspections reveal that S47 loses its helical structure in the apo form compared to its holo form (Figure S5). In addition, S48 and T56 are largely buried in the apo form but become exposed upon DNA binding. These findings are supported by additional integrated gradient attribution analyses, indicating that the structural (for S48) or surface (for S47 and T56) features in the apo state strongly disfavor the interface-residue classification (Figure S5). Together, although our model is still sensitive to the side-chain accessibility in the apo state, its multimodal architecture ensures overall robust performance.

3.6. Binding-Motif Prediction for an Uncharacterized Bacterial TF

As a real-world application, we first applied 3M-BSIP to a newly identified and structurally uncharacterized TF from Streptomyces (termed 51910). A homology search reveals that the predicted structure of 51910 shares a low TM-score (0.538) with its most similar structure from our training set (PDB ID: 7JU3), suggesting its unique structural fold (Figure A). Sequence alignment analysis further indicates low sequence conservation between these two proteins, particularly at the DNA-binding interface region (highlighted by triangles in Figure B), suggesting that the DNA recognition mechanism of 51910 could not be directly deduced based on known DBP structures.

5.

Experimental validation of the DNA-binding motif and key interface residues for a novel Streptomyces transcription factor (51910) with unknown structure. (A) Structural superposition of the AlphaFold3-predicted model of 51910 (blue) and its closest structural homologue (gray, PDB ID: 7JU3) in the training set. (B) Sequence alignment of 51910 and 7JU3. Interface residues of 7JU3, identified from its complex structure, are indicated with gray triangles. (C) Top: Residue-nucleotide interaction heatmap predicted by 3M-BSIP for 51910. Bottom: A protein–DNA complex model reconstructed using the ntM (nucleotide-Mers) method based on the predicted heatmap. (D) Schematic of experimental validation. A 138-bp promoter fragment (P₁₃₈) containing a single TTAxT motif was used alongside a minimal 25-bp element (P₂₅) and its nucleotide-converted version (P_25conv). (E) EMSA results of wild type and double mutant (R59A/R63A) of 51910, using P₁₃₈, P₂₅, or P_25conv as the targeting DNA. (F) Design of the P ₁₃₈-gusA reporter system in Streptomyces coelicolor M1146. (G) Representative chromogenic readout of 51910 and its variants.

We thus employed 3M-BSIP to generate a residue-level base-preference heatmap for 51910, which identified several key residues with high DNA contact-strength rankings, including R18, K21, R59, and R63 (Figure C). Notably, these four residues are all nonconserved in protein 7JU3. To translate this heatmap into a consecutive DNA motif, we designed a structural reconstruction method based on a prior-constructed library of nucleotide-Mers (ntM) (Figure S6), which searches a geometry-constrained library of B-form DNA fragments to combine residue-level predictions into a contiguous local model. We were thus able to assemble a physically reasonable protein–DNA complex (see Supplementary methods Section 2–3 and Supplementary results for more details). This reconstruction allows us to resolve the specific DNA-binding motif of 51910 as “TTAxT” where x represents any nucleobase. This structural modeling shows that the R59 and R63 side chains are involved in forming direct HB interactions with T5 and A3, respectively, while R18 and K21 interact with the DNA backbone (Figure C).

To validate the binding motif and key interface residues predicted by our computational predictions, we designed a series of biochemical and cellular experiments (Figure D). Our initial DNaseI footprinting assays on the promoter region of 51910 located a continuous 138-bp protected segment P₁₃₈, within which only one “TTAxT” motif exists. We thus further designed a minimal 25-bp DNA probe (P₂₅) containing the “TTAxT” motif. Electrophoretic mobility shift assays (EMSA) confirmed that 51910 can bind to both P₁₃₈ and P₂₅ in a dose-dependent manner, with the latter showing a relatively weaker binding (Figure E). Notably, the double mutant of 51910, R59A/R63A, could completely abrogate its binding to both P₁₃₈ and P₂₅, verifying their critical roles in mediating protein–DNA recognition. Moreover, to evaluate the functional role of the “TTAxT” motif, we designed a “TTAxT” to “CCGxC” converted motif (P_25conv). The assay shows that the binding of 51910 to P_25conv was greatly attenuated compared to the unconverted one, confirming the “TTAxT” motif is a potential binding site for 51910.

To test the cellular consequence of these interface residues, we placed P₁₃₈ upstream of gusA in Streptomyces coelicolor M1146 and expressed WT 51910, R59A, R63A, or R59A/R63A from the constitutive ermE promoter (Figure F). Notably, WT 51910 could profoundly reduce the chromogenic gusA signal relative to the promoter-reporter-only strain, consistent with the gene-repression role of 51910. In contrast, each single mutant and the double mutant largely lost this regulatory activity (Figure G). Together, both the biochemical and cellular experiments validated that residues R59 and R63 are functionally important for DNA recognition and transcriptional regulation. Therefore, 3M-BSIP is a reliable prediction tool that can be applied to annotate the binding motifs for the vast amount of functionally unknown DBPs or identify key interface residues for DNA recognition.

3.7. Reprogramming the DNA Binding Specificity of a Type IIB Restriction Endonuclease

Considering the accurate prediction power of 3M-BSIP, we raised the question of whether one can computationally evolve certain DBPs with altered binding specificity. To achieve this, we chose the Type IIB restriction endonuclease, BsaXI, as an example case, and integrated physics-based energy screening with 3M-BSIP to design new variants with alternative targeting motifs. For the wild-type (WT) system, 3M-BSIP predicts that the BsaXI residues, i.e., T354, K411, N414, T416, and M417, are critical interface residues, among which T354 and T416 show strong base preference toward the 8-site cytosine (Figure A-B), which is supported by the recent cryo-EM structure of BsaXI-DNA complex (PDB 8W2Q, not included in our training set). To design new BsaXI variants, we randomly introduced mutations into the BsaXI’s DNA-contact interface, then employed 3M-BSIP to predict the corresponding base-preference heatmap (see Figure S7, and Materials and Methods 2.8 for more details of the screening pipeline).

6.

Computational reprogramming of the DNA-binding specificity of BsaXI. (A) Experimentally determined structure of the wild-type (WT) BsaXI-DNA complex (PDB ID: 8W2Q). The zoomed-in inset highlights the critical DNA-contact interface residues around the 8-site cytosine. (B) The base-preference heatmap of WT BsaXI predicted by 3M-BSIP. (C) Summary of the three computationally designed BsaXI variants (Mut1, Mut2, and Mut3), detailing the respective mutation sites, target DNA sequence alterations, structural stability (Rosetta ΔΔG), and experimental validation outcomes. (D) SPR measurements of the binding affinities (K_D) or WT and mutant BsaXI against the original (control check, CK) and altered (Seq1, Seq2) target DNA sequences. (E–G) Structural modeling of the rewired protein–DNA interfaces for Mut1 (E), Mut2 (F), and Mut3 (G), with the red boxes indicating the mutated residues.

Our screening yielded three BsaXI mutant candidates predicted to exhibit both altered sequence specificity and higher structural stability compared to the WT (Figure C). Remarkably, further surface plasmon resonance (SPR) assays verified that two of the designed BsaXI variants, T354I/N414L (Mut1) and T354I/N414R/T416N (Mut2), indeed reverse the base preference at the 8-site from G:C to A:T and T:A, respectively (Figure D and S8). The corresponding specificity ratio, defined as the dissociation constant for the altered DNA divided by that of the original DNA, decreased from 1.586 to 0.663 for the double mutant, and 3.542 to 0.836 for the triple mutant. However, for the single mutant Q415V, we did not observe a complete base specificity reversal. Nevertheless, the SPR results demonstrate a substantial increase of the binding affinity toward the altered DNA sequence by ∼3-fold compared to WT, with a decrease of the specificity ratio from 3.542 to 2.883 (Figure D). This high success rate (2/3) of our method therefore potentiates the future attempts to computationally evolve the DBPs of interest with desired base specificity.

Structural modeling further provided mechanistic insights into the altered target preferences of the obtained variants. For Mut1, the introduced T354I mutation could establish nonpolar contacts with the methyl group of the target thymine at the 8-site; while the N414L substitution likely fine-tunes the orientation of the T416 side chain that further anchors the thymine base (Figure E). For Mut2, the altered base preference is mainly driven by the T416N substitution, leading to forming direct HB with the targeting adenine base, accompanied by an N414R-induced M417 side chain rotation, which directly stacks against the complementary thymine (Figure F). Notably, the binding affinity increase in Mut3 likely originates from a conformational shift induced by the Q415V mutation, thereby positioning T416 closer to the adenine to form bidentate HBs (Figure G). Therefore, we believe that 3M-BSIP achieves high accuracy via deciphering the fundamental roles governing the protein–DNA recognitions.

4. Discussion

Here, we report a multimodal deep learning framework, 3M-BSIP, that could accurately predict both the DNA-binding interface and residue-level base preference for DBPs. In comparative benchmarks, 3M-BSIP consistently outperforms the existing methods in interface-residue predictions. More importantly, its capability of predicting the residue-level base preference, which is still inaccessible in other computational approaches, provides critical molecular insights into the biological functions of DBPs, particularly for those with unknown structure or binding motifs. The robustness of 3M-BSIP is underscored by its high tolerance on allowing the computationally constructed structure as input and also those undergoing profound conformational changes upon DNA binding. Experimental validations for two different DBPs further prove that 3M-BSIP serves as a powerful computational tool for deciphering protein–DNA recognition, potentiating its applications in large-scale functional annotation and protein engineering.

The predictive power of 3M-BSIP originates from the synergistic integration of multimodal features that construct a holistic representation of the DBP’s functional landscape. The sequence embedding capture fundamental evolutionary constraints at the DNA-binding interfacea property where residue conservation exceeds that observed in protein–protein interaction (PPI) interfaces. However, these sequence-derived features alone are insufficient to resolve the unique binding patterns between protein and DNA. Among more than 70 DBP superfamilies, over 100 DNA-binding structural folds might exist; these folds, however, achieve the DNA shape recognition by sharing only a limited number of folding patterns (e.g., Helix–Turn–Helix and AT-hook). ,, In this regard, 3M-BSIP also includes the geometric graph representation to encode the essential three-dimensional structural context, defining the global fold and spatial arrangement of the DNA-binding domain. Finally, we also incorporate the molecular surface features of DBPs to describe its physicochemical landscape, including electrostatic potential, hydrophobicity, and local geometry. Similar methods, for example, MaSIF, have been successfully applied to predict PPIs, protein–ligand recognition, and protein binder design. ,, The ablation study suggests that inclusion of this surface description could significantly improve the prediction accuracy. Taking together, fusion of these multimodel representations empowers 3M-BSIP to decipher the complex DNA-specificity codes for DBPs.

Beyond predictive performance, modal attribution analysis provides a clear physicochemical interpretation of residue-level base preference across protein–DNA interfaces (Figure S9). Base contacts tended to receive greater surface attribution, consistent with the directional hydrogen bonding or nonpolar packing, as expected for the bidentate HB potential of amide-bearing side-chains toward A/G. Surface features therefore provide information on the local donor–acceptor arrangement and electrostatic environment presented to the nucleobase. By contrast, backbone-only contacts tended to receive greater structural attribution. This pattern is consistent with shape readout, in which three-dimensional geometry constrains groove accessibility and dictates spatial positioning, the nucleotide label is used to reflect the local DNA conformation determined by the sequence, rather than the base chemistry. These model attributions suggest that 3M-BSIP achieves high predictive accuracy by autonomously assigns complementary roles among its feature modalities.

The potential applications of 3M-BSIP are substantial. The ability to make accurate predictions for computationally generated proteins would confer 3M-BSIP as a powerful tool for large-scale functional annotation, thereby bridging the gap between the vastness of sequence space and the scarcity of functional characterization. , Critically, its resilience to conformational changes of DBPs allows us to make reliable predictions directly from the apo DBP structures, widely broadening the applications of 3M-BSIP due to their high prevalence in DNA recognition. , Moreover, the predicted residue-level base-preference heatmap could guide us to computationally evolve certain DBPs, as evidenced in the current work. This potential application would provide a valuable computational tool to design new DBPs for diverse functional applications, including programmable transcriptional regulation, targeted genome editing with enhanced specificity for therapeutic purposes, and synthetic biology platforms requiring precise DNA recognition modules. Notably, the presented multimodal framework could be inherently applied to other challenging problems, such as predicting the interactions with RNA or noncanonical DNA structures like G-quadruplexes, for which predictive models remain underdeveloped. ,

Despite that, we also acknowledge the current limitations of 3M-BSIP. First, the predicted residue-level base-preference heatmap could not be directly translated into sequence-based motifs like PWMs that can be used for genome-wide scanning. Although the ntM method has been developed to model the protein–DNA complex from the predicting results, their dependence on high-fidelity side-chain conformations limits their accuracy when applied to computationally modeled structures. , In addition, the current single-protein input cannot encode the cooperative recognition among multiple DNA-binding units, which limits the model’s application in multi-TF complexes. Extending the framework to higher-order protein–DNA assemblies is an important direction for future work. Finally, the model’s protein-centric design may weaken its sensitivity to the recognition events dominated by DNA shape readout. , One promising future attempt is to replace the current canonical base-preference labels (A/T/C/G) with a DNA structural alphabet, which would allow the model to recognize the base preferences in a context of certain DNA backbone conformation rather than in a local environment. , However, the overarching constraint on developing more sophisticated models remains the limited availability of high-resolution protein–DNA complex structures, a fundamental challenge for the field.

Source code and a link to the training weights have been made available at https://github.com/LiSY-SJTU/3M-BSIP. All data used for training and evaluation are publicly available through the PDB (https://www.rcsb.org/).

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/jacsau.6c00925.

• Additional methods and Figures S1–S11 are available (PDF)

⊞.

S.L. and W.H. contributed equally.

This work was supported by the National Natural Science Foundation of China [22677111]; the Natural Science Foundation of Shanghai [20ZR1425400, 21JC1403100, 23J21900800]; and the National Key Research and Development Program of China [2024YFE01997000].

The authors declare no competing financial interest.