More Accurate Binding Affinity Prediction Using Protein Homology and Ligand-Based Transfer Learning

Abstract

Accurate and rapid prediction of protein-ligand binding affinities is critical for drug discovery, particularly when evaluating large chemical libraries or new drug molecules from high-throughput generative models. We present UCBbind, a hybrid framework that combines a similarity-based transfer module with a deep learning-based prediction module to efficiently estimate binding affinities of small molecules to target proteins. For each query protein-ligand pair, UCBbind transfers experimental data from highly similar reference pairs when available, and applies the prediction module when no sufficiently similar reference exists. We benchmarked UCBbind on multiple datasets, including the CASF-2016 set, the HiQBind dataset post 2020, and the COVID Moonshot database. Our results show that UCBbind achieves state-of-the-art predictive performance, particularly for test entries with high similarity to well-characterized reference proteins and ligands, and can support downstream tasks such as binding site prediction and binder/non-binder classification.

Graphical Abstract

Introduction

The determination of protein-ligand binding affinities is crucial in drug discovery, especially with the rise of deep learning generative models ^1,2 that create large data sets of generated drug molecules against critical protein targets. Many physics-based and machine-learned scoring functions (SFs) have been developed to predict protein-ligand binding free energies, and can be broadly categorized into either classical scoring functions (CSFs), ^3-13 or machine learning scoring functions (MLSFs).^14-21 CSFs such as the commonly used AutoDock Vina ³ are designed to model intermolecular interactions or missing free energy components, and can benefit both from better design of the functional form, many-body physics, as well as the availability of training data for parameterization of their semi-empirical or empirical functions.^3-9 They are used typically in a molecular docking ²² context, a popular method of choice for screening candidate drug molecules using 3D structural model generation and scoring. Despite its popularity, docking has notable limitations due to uncertainties in scoring functions, which can lead to both false positives and false negatives ^23,24 in assessments of binding free energies.

Data-driven MLSFs by contrast are less reliant on physical interaction modeling and are far more dependent on experimental information, ^11,25,26 culminating in the current development and use of sophisticated machine learning (ML) models ^14-21 whose much larger parameter spaces are optimized on large, high quality datasets.^27-32 Some representative MLSF designs are also based on 3D structural information, including Random Forest Score (RFScore) ¹⁴ that is based on atom-pair distance counts, InteractionGraphNet (IGN) ¹⁵ that uses graph neural network (GNN) to represent raw 3-dimensional protein and ligand structures, RFScore-V3 ³³ that incorporates terms from AutoDock Vina, and also PIGNet, ¹⁶ RTMScore ¹⁷ and GIGN ²¹ which are more advanced GNN-based models. However, most of these methods require protein-ligand complexes as input, limiting their practical utility in virtual screening, as they rely on external tools for pose sampling. To overcome this limitation, numerous models have been developed to perform sampling and scoring simultaneously. These approaches vary in their input requirements, ranging from a protein structure with a specified binding site, ^34,35 to a protein structure alone, ^20,36 or simply a protein sequence such as the co-folding Boltz-2 model ³⁷ in addition to ligand information.

A more scalable approach for scoring protein-ligand complexes involves sequence-based methods, which typically require only drug SMILES (Simplified Molecular Input Line Entry System) ³⁸ and target protein sequences as input information. For example, DeepDTA ¹⁸ employs a CNN architecture consisting of three sequential 1D convolutional layers to extract semantic features from drug SMILES strings and target protein sequences. AttentionDTA ³⁹ extends this work by incorporating a two-sided multi-head attention module to capture the relationship between those features, with the hope of effectively modeling the intermolecular interactions between the protein and ligand but without explicit 3D information. More recent sequence-based models have also integrated graph structures to further enhance predictive performance. Methods such as MMPD-DTA, ⁴⁰ GraphDTA, ⁴¹ and DeepTGIN ⁴² exemplify this trend, utilizing molecular and drug-target graphs to capture local and global structural information alongside sequential features. Such multimodal approaches effectively combine the complementary strengths of sequence and graph representations to achieve superior results.

In essence sequence methods implicitly rely on the underlying biophysical principle of the sequence-structure-function relationship, ^43,44 i.e. that sequence methods generally yield an accurate prediction of a property when a good sequence homologue is found in a reference database. This foundational concept is what leveraged the prediction of protein structure via fold recognition methods ⁴⁵ or NMR chemical shifts by “transferring” chemical shift information from known sequences in the Biomolecular Magnetic Resonance Bank ⁴⁶ to the query sequence through global sequence alignment as done originally in SHIFTX2.⁴⁷ When no similar sequence is found to exist in the reference dataset, features extracted from the query sequence can instead be used to train or query machine learning models. The UCBShift chemical shift prediction programs ^48,49 modernized this approach by creating a hybrid architecture to make chemical shift predictions using machine learning on a set of extracted features and transferring experimental chemical shifts from a reference database when there is significant sequential and structural alignment. Transfer prediction is beneficial for by-passing the current inability to back-calculate chemical shifts for structural ensembles, which is a similar issue to the limitations to predicting binding affinities from CSFs and MLSFs.

Here we present UCBbind, a binding affinity prediction program that leverages the same conceptual framework of transfer prediction. Like UCBShift it is a hybrid NN architecture of two prediction modes: a transfer prediction module that employs protein sequence alignment and ligand chemical similarity (using molecular fingerprints) for experimental binding affinity replication, and a deep learning module based on AttentionDTA to predict binding affinities when homologous sequences are unavailable. The overall transfer prediction module operates as a multimodal framework, effectively fusing the sequential representation of the protein with the structural and chemical feature representation of the ligand to achieve robust prediction. We evaluate UCBbind on the CASF-2016 benchmark set, ⁵⁰ which serves as a standard benchmark for binding affinity prediction. To further assess the model’s generalizability, we tested it on a HiQBind, ³² a new dataset based on a time split of protein-ligand pairs after 2021 to ensure evaluation on previously unseen protein-ligand pairs. More importantly, the data has been cleaned to satisfy steric and energetic criteria.⁵¹ To examine target-specific transferability, we also evaluated UCBbind on a dataset of noncovalent compounds designed to inhibit the SARS-CoV-2 main protease from the COVID Moonshot open-source project.⁵² Finally, we explore UCBbind’s ability to perform related downstream tasks such as binder/nonbinder classification and binding site prediction.

Materials and Methods

Training and Test Datasets.

We trained our UCBbind model on BindingDB, ²⁸ a dataset containing approximately 2.9 million protein–ligand binding affinity measurements. To ensure data quality, we filtered the dataset to retain only entries with reported IC₅₀, K_i, or K_d values, and removed entries with unphysical assay temperatures (outside 20 – 40°C). Entries without reported temperatures were assumed to have been conducted at room temperature. For protein-ligand pairs with multiple reported affinities, we averaged the available values to mitigate the influence of outliers and experimental noise while still preserving information from multiple independent measurements. This procedure produces a more stable and representative estimate of binding affinity for transfer-based learning. After filtering and aggregation, the dataset consisted of 1, 645, 667 unique interactions.

UCBbind was then evaluated on three external and open-source protein-ligand benchmarks with no overlapping entries to BindingDB from: (1) CASF-2016 benchmark set, ⁵⁰ (2) the HiQBind dataset ³² post-2020 that have been prepared to also remove steric clashes, and (3) a target-specific dataset of noncovalent binders and non-binders from the COVID Moonshot project dataset.⁵² The CASF-2016 benchmark consists of 285 protein-ligand complexes with high-quality crystal structures and reliable binding constants. The post-2020 HiQBind dataset, after being filtering duplicate entries and proteins whose Uniprot IDs could not be mapped to sequences in UniProtKB, had 6, 085 data points. The CASF-2016 benchmark set and HiQBind were chosen as diverse benchmarks spanning a broad range of protein-ligand pairs, allowing both Module X and Module Y to be evaluated on general, data-sparse and data-rich interactions. The COVID Moonshot dataset, by contrast, contains protein-ligand pairs with high homology to entries in BindingDB, providing a data-rich regime in which the transfer-based Module Y can demonstrate its full potential. The dataset contains noncovalent compounds, ranging from initial fragment hits to optimized inhibitors, with reported activity against the SARS-CoV-2 main protease. Compounds with IC50 ≤ 1.0 × 10⁻⁵ M were labeled as binders, whereas those with higher IC50 values were treated as nonbinders. Entries reported only as inequality thresholds for nonbinding activity were removed to avoid ambiguity. After filtering, the dataset comprised 914 binders and 452 nonbinders. However, a separate analysis including entries reported as inequality thresholds (e.g., > 99 μM is provided in the Supplementary Information.

Model architecture.

The overall architecture of UCBbind is shown in Figure 1, and involves a decision tree that aggregates experimental data based on similarity or dissimilarity to known sequence information and associated binding affinities in BindingDB. More specifically, UCBbind uses a transfer prediction module, Y-Module, that operates when similarity is found to related proteins and ligands in BindingDB, and a deep learning module, X-Module, that predicts cases when no similarity exists for the query protein-ligand pair. The transfer modules are designed to estimate binding affinity by aggregating and weighting experimental binding data from similar protein–ligand pairs. To identify similar proteins, we first constructed a reference sequence database from the proteins in BindingDB. The query protein sequence was then searched against this reference database using the MMseqs2 algorithm.⁵³ To identify similar ligands, the SMILES string of the query ligand and the possible related ligands in BindingDB were converted to Morgan fingerprints (radius = 2, 1024 bits).⁵⁴ The query ligand is then searched against the reference database using a K-Nearest Neighbors Search.⁵⁵ On the BindingDB reference dataset (8,203 unique sequences, 1,037,995 unique SMILES), each query takes roughly 1s for the protein similarity search and 0.3s for the ligand similarity search. Because both searches scale roughly linearly with the number of proteins or ligands in the reference set, runtimes will increase proportionally as BindingDB grows.

Figure 1: Overall UCBbind workflow.

(a) Overall design of the UCBbind architecture that ombines both a Y-module transfer prediction module that relies on protein sequence and/or ligand SMILES alignments, and a X-Module that utilizes the AttentionDTA machine learning model. (b) Y-module architecture that takes similar protein ligand pairs as input and uses a error prediction model to adjust the weighted average affinity. (c) X-module architecture, where L_d and L_e are the ligand and protein lengths, F_e is the embedding dimension, and L_c is the number of channels of the final CNN layer.

Y-Module.

When one or more similar protein–ligand pairs are found then the algorithm moves through Y-Module, and computes a weighted average of the reference binding affinities, with weights based on protein and ligand similarity, and then applies a residual predictor to correct for systematic errors using features such as uncertainty, number of effective neighbors, and dominance ratio. More specifically, given a set of similar proteins $𝒫=\left\{p_i\right\}$ and a set of similar ligands $ℒ=\left\{{\ell }_j\right\}$, pairs are considered similar if the protein sequence identity exceeds 0.30 and the ligand Tanimoto similarity exceeds 0.60; these thresholds were chosen to balance predictive accuracy with sufficient coverage of transferable pairs (see Figure S1 for sensitivity analysis).

The raw similarity scores $S_{\text{prot}}$ (sequence identity for proteins) and $S_{\text{lig}}$ (Tanimoto similarity for ligands) are first normalized:

$${\tilde{S}}_{\text{prot}}\left(p_i\right)=\frac{S_{\text{prot}}\left(p_i\right)}{{\sum }_{p_k\in 𝒫}{S}_{\text{prot}}\left(p_k\right)}$$ (1)

$${\tilde{S}}_{\text{lig}}\left({\ell }_j\right)=\frac{S_{\text{lig}}\left({\ell }_j\right)}{{\sum }_{{\ell }_j\in ℒ}{S}_{\text{lig}}\left({\ell }_j\right)}.$$ (2)

For each protein-ligand pair (p_i, ℓ_j) that exists in the reference dataset, a weight is computed as:

$$w_{ij}={\tilde{S}}_{\text{prot}}^{\alpha }\left(p_i\right)*{\tilde{S}}_{\text{lig}}^{\beta }\left({\ell }_j\right),$$ (3)

where α and β control the relative influence of protein and ligand similarity, as well as the weighting among highly similar protein ligand pairs. Their optimal values were determined via grid search on the PDBbind refined set, a subset of PDBbind consisting of 5, 316 higher-quality protein-ligand pairs that are not found in CASF-2016 (see Figure S2 for details on the grid search, including stability analyses and handling of weight underflow). This multiplicative formulation emphasizes that a low similarity in either component suppresses the overall transfer weight, reflecting the interdependent nature of protein–ligand binding. The weighted average is then computed to be

$$\hat{B}=\frac{{\sum }_{i,j}{w}_{ij}\cdot B_{ij}}{{\sum }_{i,j}{w}_{ij}}$$ (4)

where B_ij is the experimentally measured binding affinity of the reference pair (p_i, l_j). To reduce noise from low-similarity references, only pairs whose weight satisfies $w_{ij}\ge \tau \cdot {\text{max}}_{i,j}\left(w_{ij}\right)$, with τ = 0.65, are included in the calculation (see Figure S3 for τ selection process. If no such similar pairs exist in the reference dataset, UCBbind defers to the machine learning-based prediction module UCBbind-X.

While the similarity-based transfer described above provides a useful approximation of the binding affinity ˆ, its accuracy depends on the number and quality of similar pairs in the reference dataset. In cases where the distribution of affinities has high variation, the transfer may be unreliable. Hence UCBbind incorporates a residual prediction model that estimates the expected error in the transferred prediction $\hat{B}$, using meta-features derived from the set of retrieved similar pairs. These features include:

• Uncertainty: Weighted standard deviation of affinities $\left\{y_{ij}\right\}$ used in the weighted average,

  $$\sigma =\sqrt{\frac{{\sum }_{i,j}{w}_{ij}{\left(B_{ij}-\hat{B}\right)}^2}{{\sum }_{i,j}{w}_{ij}}}$$

  where w_ij are the normalized weights for each protein-ligand pair.
• Dominance Ratio: Fraction of total weight contributed by the most similar protein-ligand pair,

  $$D=\frac{{\text{max}}_{i,j}{w}_{ij}}{{\sum }_{i,j}{w}_{ij}}$$
• Effective Neighbors: The effective number of contributing pairs defined as the ratio of square of the summed weights to the sum of the squared weights,

  $$N_{\text{eff}}=\frac{{\left({\sum }_{i,j}{w}_{ij}\right)}^2}{{\sum }_{i,j}{w}_{ij}^2}$$

These features were standardized and passed into a linear Ridge Regression model, which applies L2 regularization (α = 1.0) to prevent overfitting. The model was trained on similarity-transfer predictions and their residuals computed from the PDBbind general dataset, learning to predict the residual error $r=B_{\text{true}}-\hat{B}$. Of the 14,127 protein–ligand pairs in PDBbind general, 8,871 produced valid similarity-transfer predictions and were therefore used to train the residual model. The regularization parameter α was selected via 5-fold cross validation. The final corrected prediction was obtained by adding a bounded correction to the transferred prediction:

$${\hat{B}}_{\text{final}}=\hat{B}+\text{tanh}(r)$$ (5)

The hyperbolic tangent function limits the correction magnitude and prevents overadjustment when large residuals are predicted. Because the residual model is low-capacity and regularized, the 6,181 training samples were sufficient to ensure stable and generalizable performance. More complex models such as XGBoost and Random Forest were explored, but linear regression was found to be comparably effective while remaining interpretable. Given the relatively small size of the PDBbind general training set, linear modeling provides a robust solution that reduces the risk of overfitting.

X-Module is a general-purpose ML module designed to predict binding affinity for protein–ligand pairs where transfer prediction is not possible. In this work, we implement X-Module using AttentionDTA, ³⁹ a deep learning architecture designed to model protein–ligand interactions. Inputs are integer-encoded SMILES strings and amino acid sequences, transformed into embeddings used in 1D convolutional neural networks (CNNs). Notably, we expanded the SMILES and protein vocabularies to 69 and 25 tokens respectively to accommodate the greater chemical and sequence diversity in BindingDB compared to the original AttentionDTA training data (which used vocabularies of size 64 and 20). This is followed by a two-sided multi-head attention mechanism that computes cross-attention between subsequences of the ligand and protein. A global max pooling layer summarizes the features, and the concatenated representation is passed through a multi-layer perceptron (MLP) for the predicted binding affinity. The model was trained using mean squared error loss on BindingDB with standard regularization and optimization techniques. We adopted the same hyperparameters and architectural settings reported for the original AttentionDTA model, ³⁹ including three 1D convolutional layers, a multi-head attention mechanism, and a five-layer MLP trained using the Adam optimizer with a learning rate of 0.0001.

Figure 2 shows the maximum protein and ligand similarities between the non-specific target test sets, i.e. CASF-2016 and HiQBind datasets, and the reference BindingDB to characterize how closely the test sets resemble the training data. Reference entries directly overlapping with the test set were removed to avoid obvious data leakage. The maximum similarities shown in Figure 2(a,b) demonstrate that many test entries exceed the cutoff, since the point of UCBbind is to take advantage of highly similar entries via the Y-module and to only use the X-Module when such information is minimally similar. Figure 2c shows the alignment lengths of the most similar proteins, highlighting that the vast majority exceed 100 residues, which demonstrates the statistical significance of highly similar protein matches. The Moonshot binders and nonbinders were excluded from this analysis because each set contains only a single protein (SARS-CoV-2 main protease) and all ligands have a maximum similarity of 1.0.

Figure 2: Similarity comparison of CASF-2016 and HiQBind with BindingDB.

Comparison between test sets and BindingDB for (a) maximum ligand similarities, (b) maximum protein similarities, (c) protein alignment lengths, and (d) maximum protein-ligand pair similarities defined as max(protein similarity ∗ ligand similarity). The similarity threshold of 0.18 corresponds to the minimum protein similarity (0.30) multiplied by the minimum ligand similarity (0.60). The peak at 1.00 arises from ten identical pairs for CASF-2016 and 421 for HiQBind.

We note that Y-Module, however, is designed to leverage highly similar proteins and ligands to transfer binding affinities; this similarity-driven mechanism is intrinsic to its function and is not considered unintended overlap. Consequently, UCBBind’s evaluations reflect realistic use cases where affinity can be inferred from closely related protein–ligand pairs, while ensuring that exact duplicates do not inflate performance.

We used the default settings and applied a sequence identity threshold of 0.30 to retain high-confidence matches for proteins. Proteins with bit scores above this cutoff were retained as “similar proteins.” We applied a Tanimoto Similarity score threshold ⁵⁶ of 0.60 to retain high-confidence-matches for ligands. Ligands with calculated Morgan Fingerprints that scored above this cutoff were retained as “similar ligands.”

Docking protocols for AutoDock4 and Vina benchmarks.

Ligands from the COVID Moonshot dataset were converted from SMILES to 3D structures using RDKit, hydrogens were added, and geometries were optimized using the UFF force field. This step was unnecessary for CASF-2016 set, which already had ligand SDF structure files. Partial charges were assigned according to the docking software: Gasteiger charges for AutoDock4 and EEM charges for AutoDock Vina. The Moonshot receptor (SARS-CoV-2 Mpro, PDB ID: 7CAM) was prepared by removing water molecules, adding polar hydrogens, and assigning Gasteiger charges before conversion to the PDBQT format for both docking programs. The same procedure was done for all the receptors in the CASF-2016 set.

For AutoDock Vina, docking was performed with a cubic search box centered on the ligand coordinates extracted from the SDF, with box dimensions of 20 × 20 × 20 Å and an exhaustiveness of 32. For AutoDock4, grids were generated with 60 × 60 × 60 points at 0.375 Å spacing, and AutoDock-GPU was run with 50 independent runs per ligand. The lowest-energy pose from each docking calculation was selected, and the corresponding predicted binding affinity was recorded. All preparation steps, grid parameters, and docking settings are documented and available in the accompanying scripts to ensure reproducibility.

Training of deep learning benchmarks.

For a fair comparison within the UCBBind framework, both DeepDTA and AttentionDTA were retrained on the BindingDB dataset. The dataset was split into training and validation sets using a stratified random split according to binding affinities to ensure balanced representation of strong and weak binders. The validation set was kept fixed for all experiments. All training hyperparameters—including learning rate, batch size, number of epochs, and network architectures—followed the original specifications in the DeepDTA and AttentionDTA publications. Predictions for external benchmarks (CASF-2016, HiQBind, and COVID Moonshot) were made using the trained models without any additional fine-tuning. This setup ensures that differences in performance reflect the model architectures and the contribution of UCBBind’s transfer module, rather than dataset-specific biases. Boltz-2 was used as provided by the authors without retraining, due to the complexity of its architecture and the lack of publicly available training protocols.³⁷ As a result, Boltz2 was not evaluated on the CASF-2016 benchmark, since it was included in its training data. Finally, we also performed an additional comparative analysis against the GraphDTA model as a more recent prototype that uses both sequence information and graph-based representations.⁴¹

Results

Binding Affinities for CASF-2016 and HiQBind Test Sets.

Table 1 summarizes the test performance of UCBbind and other MLSF and CSF models on the CASF-2016 and HiQBind data sets when trained on BindingDB. Starting with CASF-2016, the sequence-based DTA methods marginally outperform the best-in-class CSF represented by AutoDock-vina, whereas UCBbind achieves the lowest RMSE of 1.40 kcal/mol and the highest Pearson correlation coefficient r of 0.76.

Table 1:

Performance comparison on the CASF-2016 and HiQBind test sets.

Data Set (n)	Method	MAE	RMSE	95% CI	r
CASF-2016 (285)	DeepDTA	1.34	1.65	[1.52, 1.78]	0.69
	AttentionDTA	1.31	1.64	[1.51, 1.77]	0.68
	GraphDTA	1.28	1.67	[1.52, 1.83]	0.64
	AutoDock Vina	1.94	2.46	[2.24, 2.68]	0.58
	AutoDock4	7.55	8.32	[7.47, 9.10]	0.29
	UCBbind	0.98	1.40	[1.24, 1.54]	0.76
HiQBind (6,805)	DeepDTA	1.16	1.49	[1.46, 1.52]	0.60
	AttentionDTA	1.12	1.46	[1.43, 1.49]	0.64
	GraphDTA	1.02	1.37	[1.34, 1.40]	0.68
	UCBbind	0.82	1.24	[1.21, 1.27]	0.75

Performance results are reported in terms of mean absolute error (MAE) and root mean squared error (RMSE), both in kcal/mol, with 95% confidence intervals (CI) for RMSE and Pearson correlation coefficient (r) for different MLSF and CSF models.

We further evaluated UCBbind’s performance on the CASF-2016 set by assessing the individual contributions of its ML UCBbind-X and the homology-based UCBbind-Y, as also shown in Table 2. Of the 285 protein–ligand pairs in the test set, 128 were routed to UCBbind-X and 157 to UCBbind-Y. UCBbind-Y achieved an RMSE of 1.10 kcal/mol and an r-value of 0.84, while as expected UCBbind-X achieved an RMSE and r-value comparable to the standalone AttentionDTA from which it was derived. We also assess UCBbind on the HiQBind dataset, ³² which provides a complementary set of high-confidence protein–ligand complexes not previously seen in CASF-2016. Table 1 summarizes the performance of UCBbind and the same MLSF models, excluding the docking software. The docking methods were omitted due to both the computational cost of evaluating 6,805 complexes and their poor performance on the CASF-2016 set. UCBbind once again achieves the lowest RMSE of 1.24 kcal/mol and a Pearson correlation coefficient r of 0.75. Of the 6805 protein–ligand pairs in the test set, 2570 were routed to UCBbind-X and 4235 to UCBbind-Y. UCBbind-Y achieved an RMSE of 0.88 kcal/mol and an r-value of 0.85, while as UCBbind-X again achieved an RMSE and r-value comparable to AttentionDTA.

Table 2:

Performance breakdown of components of the UCBbind architecture on the CASF-2016 and HiQBind test sets.

DataSet	Method	MAE	RMSE	95% CI	r	Pairs
CASF-2016	UCBbind	0.98	1.40	[1.24, 1.54]	0.76
	UCBbind-X	1.34	1.69	[1.48, 1.90]	0.57	128
	UCBbind-Y	0.69	1.10	[0.89, 1.30]	0.84	157
	UCBbind-X on Y pairs	1.29	1.60	[1.43, 1.76]	0.83	157
	UCBbind-Y (no residual)	0.72	1.15	[0.93, 1.34]	0.83	157
HiQBind	UCBbind	0.82	1.24	[1.21, 1.27]	0.75
	UCBbind-X	1.29	1.68	[1.62, 1.73]	0.48	2570
	UCBbind-Y	0.53	0.88	[0.84, 0.91]	0.85	4235
	UCBbind-X on Y pairs	1.01	1.32	[1.28, 1.35]	0.66	4235
	UCBbind-Y (no residual)	0.54	0.89	[0.86, 0.93]	0.84	4235

MAE and RMSE (both in kcal/mol) with 95% confidence intervals (CI) for RMSE and Pearson correlation (r) are shown for the full UCBbind predictions, individual modules (X and Y), and ablation studies including Module X restricted to pairs processed by Module Y and Module Y without the residual correction.

To better understand the contributions of each architectural component in Figure 1 of UCBbind, we conducted an ablation study evaluating the effect of Module Y and comparing its performance against Module X (largely AttentionDTA) on the same subsets of protein-ligand pairs as summarized in Table 2. For the CASF-2016 protein–ligand pairs routed through UCBbind-Y, removing the residual correction increases the RMSE from 1.10 to 1.15 and lowers the correlation from 0.84 to 0.83. Notably, if these same complexes were instead routed through Module X, the RMSE rises to 1.35 and the Pearson correlation decreases to 0.60. This highlights the effectiveness of homology-based routing in capturing transferable interactions and justifies the use of Module Y for these cases. The performance breakdown of UCBbind-X and UCBbind-Y for HiQBind mirrors the trends observed for the CASF-2016 set, though we note diminishing returns for the residual correction in cases where the weighted average already gives accurate predictions, as in the case with HiQBind.

Binding Affinities and Classification for Moonshot Binders and Non-Binders.

To demonstrate the full capabilities of UCBbind-Y, we evaluated UCBbind’s performance on 914 noncovalent binders found in the COVID Moonshot dataset, defined as compounds with reported IC₅₀ value against the SARS-CoV-2 main protease less than 1.0 × 10⁻⁵ M. As shown in Table 3, UCBbind yields the least error by a considerable margin, with an RMSE of 0.26 kcal/mol and a Pearson correlation of 0.91. Since BindingDB contains a plethora of binding data for the SARS-CoV-2 main protease, the vast majority (913/914) of predictions were made via UCBbind-Y (Table S1). These results indicate that UCBbind-Y makes a major contribution to UCBbind’s overall robust performance, and its predictive power is expected to improve over time as more and more experimental data is acquired.

Table 3:

Performance comparison on the Moonshot binders and nonbinders.

Data Set (n)	Method	MAE	RMSE	95% CI	r
Moonshot binders (914)	DeepDTA	0.57	0.71	[0.68, 0.75]	0.33
	AttentionDTA	0.67	0.85	[0.81, 0.89]	0.26
	GraphDTA	0.66	0.83	[0.79, 0.87]	0.20
	Boltz-2	1.23	1.52	[1.45, 1.60]	0.53
	AutoDock Vina	4.73	4.89	[4.81, 4.97]	0.19
	AutoDock4	1.51	2.13	[1.98, 2.28]	0.02
	UCBbind	0.15	0.26	[0.23, 0.29]	0.91
Moonshot nonbinders (452)	DeepDTA	1.26	1.42	[1.35, 1.48]	0.01
	AttentionDTA	1.76	1.90	[1.84, 1.97]	−0.04
	GraphDTA	1.05	1.18	[1.12, 1.23]	0.14
	Boltz-2	0.79	1.01	[0.94, 1.08]	0.15
	AutoDock Vina	1.97	2.16	[2.06, 2.26]	0.12
	AutoDock4	3.77	4.84	[4.33, 5.36]	−0.19
	UCBbind	0.26	0.58	[0.50, 0.66]	0.52

Performance results are reported in terms of mean absolute error (MAE) and root mean squared error (RMSE) with 95% confidence intervals (CI) for RMSE, all in kcal/mol, along with Pearson correlation coefficient (r) for different MLSF and CSF models.

To examine whether the strong performance observed for binders generalizes, we also evaluated UCBbind on the 452 Moonshot nonbinders. Despite the BindingDB reference dataset being heavily skewed toward binders, as illustrated by the -pIC₅₀ distribution in Figure 3a, UCBbind achieves the lowest RMSE of 0.58 kcal/mol and highest Pearson correlation value r of 0.52, as shown in Table 3.

Figure 3: Data statistics of Moonshot binders and nonbinders.

(a) Distribution of binding affinities for BindingDB, Moonshot binders, and Moonshot nonbinders (b) Classification accuracies of UCBbind and four benchmark models on Moonshot binders and nonbinders.

Notably, all of the 1,012 Moonshot nonbinders were routed to UCBbind-Y, underscoring the model’s ability to identify and exploit relevant weak-binding data in the reference dataset. UCBbind-Y achieved an RMSE of 0.58 kcal/mol and a Pearson correlation of 0.52, reflecting the model’s effectiveness in transferring from the nonbinder regime. This strong performance supports the conclusion that transferable structure–activity patterns extend into the nonbinder region, enabling accurate prediction even in the low-affinity regime. We have reproduced these results for the case in which the nonbinders reported only as inequality thresholds (e.g., > 99 μM) are included, and performance remains consistent (Table S2), showing UCBbind’s robustness to assumptions made when we are only given inequality data. These entries were assumed to have a ground truth IC₅₀ value of 99 μM. We also note that binders and nonbinders are in reality not known in prior. To ensure the robustness of our model, we have also included the regression statistics for the complete dataset (Table S3).

Using UCBbind’s predictions of binding affinity, we can also classify whether or not a query ligand is a binder or a non – binder. A compound is considered a binder if its predicted negative pIC₅₀ is greater than 5, and a non-binder otherwise. This cutoff corresponds approximately to an IC₅₀ value of 1.0 × 10⁻⁵ M, which is a commonly used threshold to distinguish between binders and nonbinders.^57-59

We assessed how accurately UCBbind classified Moonshot binders and nonbinders based on predicted binding affinities, in comparison to other benchmark models. As shown in Figure 3b, UCBbind classified the protein-ligand pairs in the Moonshot database with high accuracy, correctly identifying 99.45% of binders and 81.86% of nonbinders. This performance is particularly notable given the class imbalance in the training dataset (BindingDB), which contains 1, 448, 466 binders and only 197, 403 nonbinders (Figure 3a). Despite this imbalance, UCBbind demonstrates strong generalization to both classes and does not exhibit strong prediction bias. UCBbind also demonstrates superior ranking performance, as evidenced by AUC-ROC measurements that provide a threshold-independent measure of discriminative power (Table S4). The binding affinity distributions and classification statistics have been reproduced to include the nonbinders excluded from analysis due to their IC₅₀ values being reported as inequalities (Figure S4).

Identification of Ligand Binding Sites.

Although UCBbind was developed for binding affinity prediction, we evaluated whether its similarity-based module (Module Y) could also identify binding sites. To do this, we used the PDBbind general set as the reference set from which to transfer pocket coordinates to the complexes in CASF-2016, when a similar protein–ligand pair was available. The ground truth binding sites came from CASF-2016 annotations, enabling direct comparison between transferred pockets and experimentally determined ones.

For this case study, we limited the evaluation to protein–ligand complexes from the CASF-2016 set that were routed to Y-Module Joint and had at least one similar complex in the PDBbind general set with either a different protein or a different ligand. Among such candidates, we selected the most similar complex that met these criteria, and used its binding pocket for transfer to the query complex. This filtering resulted in a final evaluation set of 157 complexes. Binding site prediction accuracy was measured using the distance between centers of the predicted and true binding pockets (DCA), with a success defined as predictions within 4 Å of the true pocket centroid.

As shown in Table 4, Module Y’s similarity-based transfer method achieved high accuracy in predicting binding sites, with a global mean DCA of 0.61 Å and a 97.45% success rate (DCA < 4 Å). For comparison, P2Rank ⁶⁰ — a widely used binding pocket predictor that relies solely on protein structure — showed a lower success rate of 47.13%. Performance varied by similarity category, with Module Y performing best when the transferred pocket was from a protein–ligand pair sharing the same ligand. It is important to note that Module Y’s approach depends on the existence of a sufficiently similar complex in the reference set, and thus does not replace general pocket prediction methods like P2Rank. Instead, it offers a promising strategy for predicting binding sites of protein–ligand complexes lacking solved structures, provided a large and diverse structural database is available for transfer.

Table 4:

Binding site prediction performance comparison between Module Y and P2Rank, grouped by similarity category.

Category	N	Module Y		P2Rank
		Mean DCA (Å)	Success Rate (%)	Mean DCA (Å)	Success Rate (%)
Global	157	0.61	97.45	8.16	47.13
DLDP	11	0.82	90.90	5.69	45.45
DLSP	120	0.63	97.50	8.40	45.00
SLDP	26	0.46	100.00	8.08	57.69

DCA refers to the distance between centers if the predicted and true binding pockets, with a success defined as prediction within 4 Å of the true pocket centroid. Categories indicate the relationship between the query complex and the PDB from which the binding pocket was transferred: DLDP (different ligand, different protein), DLSP (different ligand, same protein), SLDP (same ligand, different protein).

Discussion

Accurately predicting protein–ligand binding affinities remains a central challenge in computational drug discovery. Traditional docking methods are limited by rigid receptor assumptions and approximate scoring functions, while deep learning models, though more predictive, often lack interpretability and chemical insight. UCBbind addresses both challenges by integrating a similarity-based transfer module with a deep learning module: the former provides interpretable predictions grounded in evolutionary and chemical similarity, while the latter ensures generalizability to novel targets and ligands when no sufficiently similar reference protein–ligand pair is available.

Across benchmark datasets—including the CASF-2016 set, a HiQBind-derived dataset, and the COVID Moonshot database—UCBbind consistently outperformed both traditional docking approaches and well-established deep learning baselines. his was especially true for the COVID Moonshot database, where there was rich representation of homologous proteins to the SARS-CoV-2 main protease in the BindingDB reference set. For this case, practically all protein ligand pairs were routed to Module Y, allowing UCBbind to achieve a low RMSE of 0.26 and 0.58 for the Moonshot binders and nonbinders, respectively. Module Y, the similarity-based transfer module, operates via protein–ligand transfer when highly similar complexes are available. For the transfer case, a residual predictor further refines the transferred binding affinity by estimating and correcting systematic deviations, improving predictive accuracy. Carefully chosen similarity thresholds ensure robust predictions while avoiding overreliance on a single reference complex. Module X, currently implemented using AttentionDTA, encodes protein sequences and SMILES with 1D convolutional networks and models intermolecular interactions with attention. It is fully modular and can be replaced with alternative architectures as desired.

Using sequence-based representations allows UCBbind to operate in a data-rich regime, leveraging large reference datasets to capture joint protein–ligand compatibility patterns efficiently. This contrasts with 3D structure-based methods, such as molecular docking, which are sensitive to inaccuracies in protein and ligand conformations, side-chain orientations, and loop flexibility. While experimentally resolved 3D structures were available for datasets like CASF-2016, enabling accurate docking, datasets such as COVID Moonshot required 3D coordinates to be generated from SMILES strings using RDKit, introducing additional structural uncertainty. Moreover, 3D docking is computationally expensive, limiting its applicability to large datasets like HiQBind. By circumventing these limitations, UCBbind captures systematic binding patterns and applies residual corrections, achieving superior performance compared to methods that rely on detailed 3D structures. Additionally, UCBbind is able to generalize to novel proteins or targets without experimentally resolved structures, enabling rapid hit identification even when 3D coordinates are unavailable. This capability is particularly valuable in early-stage drug discovery, where experimental structures may be lacking or expensive to determine, allowing data-driven exploration of large chemical and protein spaces with minimal structural assumptions.

As a data-driven model, however, UCBbind’s transfer component Module Y relies critically on accurate and consistent similarity mappings. Thus, we reason that transfer tends to underperform in two main scenarios. First, transfer is limited when the relationship between sequence similarity and binding affinity is unreliable, as in intrinsically disordered proteins (IDPs) or cases dominated by local effects such as single-point mutations. Addressing these cases will likely require models that can explicitly account for local conformational variability. Second, when the reference affinities exhibit high variation—due to heterogeneous experimental conditions, diverse data sources in BindingDB, or potential measurement errors—transfer predictions can become noisy. In UCBbind, the residual correction model partially mitigates this by learning to compensate for systematic biases in the transferred affinities. Taken together, these observations clarify the circumstances under which the transfer mechanism is most effective and highlight directions for future improvements in handling more heterogeneous or locally sensitive protein-ligand systems.

However, Module Y still demonstrated strong performance on low-affinity or nonbinding pairs, highlighting its ability to transfer meaningful weak-binding patterns despite the reference dataset being skewed toward strong binders. UCBbind’s predictions generalize well to classification tasks, accurately distinguishing binders from nonbinders with minimal bias, outperforming competing deep learning and docking methods. Looking ahead, integrating co-folding models with UCBbind could enhance predictive accuracy by capturing protein-ligand conformational dynamics, allowing the transfer module to incorporate induced-fit effects and more nuanced interaction features beyond those present in static reference complexes.

For all tested datasets, Module X performed worse than Module Y, dropping UCBbind’s overall performance. This observation underscores the need to strategically leverage the explicit chemical and evolutionary knowledge captured by the transfer module. Currently, the UCBbind architecture operates primarily as a conditional selection model, leveraging the highly interpretable transfer prediction from Module Y only when sufficient homology is detected, and defaulting to the generalizable deep learning prediction from Module X otherwise. While this design ensures robustness across varying data sparsity, it represents a limited form of synergy. Future work could aim to move beyond this selection mechanism and establish explicit cross-communication between the two components. This shift would involve integrating Module Y features with the learned latent space from Module X, even when homology is below the strict selection threshold. It may be possible to leverage the strengths of both pathways by developing a gated fusion model that dynamically weighs the contributions of the homology-informed features with the deep-learned features.

Furthermore, more work can be done to improve scalability. While the current implementation of UCBbind employs MMseqs2 and KNN search for identifying similar protein–ligand pairs, the computational cost of exhaustive similarity searches may increase with the size of reference databases. To address this, future versions could incorporate approxi- mate nearest neighbor methods such as FAISS⁶¹ or HNSW graphs,⁶² which enable sublinear-time retrieval with minimal loss in accuracy. Furthermore, embedding-based retrieval approaches us- ing pretrained protein and ligand encoders could replace raw sequence and fingerprint similarities with continuous vector representations, facilitating scalable and semantically meaningful similarity transfer.

While the primary focus of UCBbind is affinity prediction, Module Y can also provide structural insights by transferring binding po cket information from similar complexes. Although this capability is secondary to the regression performance, it illustrates the potential of leveraging experimental data to inform mechanistic binding hypotheses and guide future structure-based studies. Moving forward, as more high-quality binding data become available, UCBbind’s predictive power and transfer-based insights are expected to further improve, enhancing both affinity prediction and ligand prioritization.

Conclusion

In this work, we introduced UCBbind, a novel architecture that successfully addresses key limitations in existing protein-ligand affinity prediction methods by conditionally integrating a similarity-based transfer module with a sequence-based deep learning module. UCB-bind demonstrated superior predictive performance against both traditional docking and established deep learning baselines across diverse benchmarks. The architecture’s reliance on sequence-based representations ensures computational efficiency and robustness against structural uncertainty, while the conditional transfer component provides valuable, interpretable insights grounded in evolutionary and chemical precedent. By demonstrating the power of integrating data-driven learning and knowledge transfer, UCBbind offers a powerful, scalable, and interpretable tool to accelerate hit identification and ligand prioritization in modern drug discovery campaigns.

Supplementary Material

Details of the data formats and usage notes are provided in the Supplementary Information.

DATA AND SOFTWARE AVAILABILITY

The UCBBind code, HiQBind dataset post 2020, and COVID Moonshot data used in this paper are available at https://github.com/THGLab/UCBbind.git. The complete HiQBind dataset, including the protein-ligand structures, metadata information (binding affinity annotations, release year, resolution, ligand name, protein name, protein UniProt ID and various ligand properties) is publicly available in Figshare: https://doi.org/10.6084/m9.figshare.27430305.