KSDiffusion: conditional diffusion for kinase-specific phosphorylation site prediction under data-limited and imbalanced regimes

Abstract

Protein phosphorylation governs cellular signaling, making accurate identification of kinase-specific sites essential for understanding regulatory and disease mechanisms. Although computational approaches have shown promise in inferring kinase specificity, most existing methods primarily rely on local sequence patterns and remain limited in their ability to capture broader contextual information. More critically, experimentally validated kinase–substrate data are inherently scarce and highly imbalanced across kinase groups, substantially restricting the generalization performance of purely discriminative models, especially for underrepresented kinases. To address these challenges, we propose KSDiffusion, a unified framework for kinase-specific phosphorylation site prediction explicitly designed for data-limited and imbalanced regimes. KSDiffusion integrates a protein language model with task-aware conditional diffusion-based generative modeling. Specifically, an ESM-2-based encoder is employed to extract context-aware peptide representations enriched with evolutionary and structural information, while supervised contrastive learning further enhances kinase-specific discriminability in the embedding space. To alleviate data scarcity for rare kinase groups, we introduce a conditional diffusion model, termed KS-DiT, which generates biologically plausible and kinase-consistent synthetic representations that directly support downstream prediction. Comprehensive experiments across kinase groups spanning low-, medium-, and large-data regimes demonstrate that KSDiffusion consistently outperforms representative baseline methods. In particular, substantial improvements are achieved for data-scarce kinase groups, with AUC gains of up to 15%, while maintaining competitive performance when sufficient training data are available. These results underscore the regime-dependent effectiveness of conditional diffusion-based augmentation and highlight the value of integrating protein language models with task-aware generative modeling for robust kinase-specific phosphorylation site prediction under realistic data constraints.

Introduction

Protein phosphorylation, catalyzed by protein kinases, is one of the most fundamental and widespread posttranslational modifications (PTMs) in eukaryotic cells [1–3]. This process transfers a phosphate group from adenosine triphosphate to specific amino acid residues of substrate proteins under precisely regulated temporal and spatial contexts [4–6]. In eukaryotes, phosphorylation predominantly occurs on serine (S), threonine (T), and tyrosine (Y) residues. It is tightly controlled by kinase-mediated signaling networks to maintain cellular homeostasis [7]. Dysregulation of phosphorylation dynamics has been implicated in a wide range of diseases, including cancer, diabetes, and developmental disorders [8, 9]. Owing to their central regulatory roles, protein kinases have become one of the most important classes of therapeutic targets. Consequently, systematic characterization of kinase-mediated phosphorylation events is critical for both mechanistic studies and drug discovery [10]. In this context, the accurate identification of kinase-specific phosphorylation sites is essential for elucidating phosphorylation-dependent signaling mechanisms and advancing biomedical research [11].

The rapid development of high-throughput, mass spectrometry-based phosphoproteomics has enabled the large-scale identification of hundreds of thousands of phosphorylation sites across the human proteome [12]. However, these technologies struggle to determine the upstream kinases responsible for specific phosphorylation events. They remain particularly insufficient for resolving kinase-specific sites [13, 14]. Experimental identification of kinase–substrate relationships still relies largely on low-throughput, labor-intensive, and costly biochemical techniques. Examples include in vitro kinase assays, phosphorylation-specific antibody enrichment, and site-directed mutagenesis [15, 16]. As a result, only a small fraction of experimentally identified phosphosites have been confidently annotated with their cognate kinases, leaving substantial gaps in current annotations [17]. This incompleteness severely constrains system-level analyses of cellular signaling networks and limits our understanding of kinase-regulated biological processes [18, 19].

Consequently, computational approaches have become indispensable tools. They scalably predict kinase-specific sites by learning discriminative patterns from established kinase–substrate relationships. Early computational methods primarily relied on manually engineered sequence features combined with simple statistical or machine learning models [20]. For instance, Scansite employed position-specific scoring matrices to scan protein sequences for motifs matching known kinase substrate preferences [21]. NetPhos introduced one-hot encoding of peptide sequences and applied artificial neural networks for prediction, laying the foundation for subsequent predictors [22, 23]. GPS 1.0 further proposed a group-based strategy by constructing hierarchical models according to the evolutionary relationships of kinases at the group, family, and subfamily levels [16, 24]. Although these approaches provided early insights into kinase specificity, their reliance on handcrafted features and shallow model architectures fundamentally limited predictive performance.

To improve accuracy, subsequent studies incorporated richer feature representations and more advanced machine learning algorithms [25]. The KinasePhos series exemplifies this methodological evolution. KinasePhos 1.0 encoded sequences using position-specific scoring matrices and physicochemical properties, employing Hidden Markov Models for prediction [26]. KinasePhos 2.0 introduced the maximum relevance minimum redundancy feature selection strategy and adopted Support Vector Machines to enhance discriminative capability [27]. More recently, KinasePhos 3.0 represented protein sequences as consecutive tripeptide “words” and integrated a pretrained word2vec model with convolutional neural networks to learn high-level sequence representations [28]. With the emergence of transfer learning, GPS 6.0 further expanded its scope by pretraining a universal phosphorylation predictor on nearly 500 000 sites and transferring the learned knowledge to construct 577 kinase-specific models across 185 species [29]. Phosformer proposed an end-to-end deep learning framework that abandons manual feature engineering and explicit kinase category encoding. By feeding kinase domain sequences and substrate peptides directly into a Transformer-based architecture, it captures evolutionary and structural dependencies to achieve state-of-the-art performance [30, 31].

Despite these advances, kinase-specific phosphorylation site prediction continues to face two fundamental challenges. First, existing models rely predominantly on local substrate sequences. They often neglect broader biological contexts, including tertiary structures, evolutionary constraints, and dynamic interactions. Consequently, their predictions frequently lack essential biophysical and physiological constraints [31]. Second, experimentally validated kinase–substrate relationships are inherently data-limited and highly imbalanced. While a small number of well-studied kinases are associated with abundant phosphorylation annotations, many kinase groups possess only a handful of confirmed substrates [30]. Under such conditions, purely discriminative models tend to overfit dominant kinase groups and exhibit poor generalization to data-scarce kinases, substantially limiting their practical applicability.

To address these challenges, we propose KSDiffusion, a unified computational framework explicitly designed for data-limited and highly imbalanced regimes. KSDiffusion integrates protein language model-based representation learning with task-aware conditional diffusion modeling. Specifically, we construct a context-rich encoding module based on ESM-2, a large-scale protein language model capable of capturing evolutionary and structural information directly from raw sequences. To further enhance representation quality, we introduce a supervised contrastive learning (SCL) strategy that encourages discrimination among kinase-specific patterns while preserving biologically meaningful similarities. To mitigate the scarcity of annotations in underrepresented kinase groups, we developed a conditional diffusion model, termed KS-DiT. This model generates biologically plausible synthetic peptide embeddings conditioned on kinase-specific information. Rather than serving as a standalone augmentation technique, this diffusion model is tightly coupled with the downstream prediction task. This integration enables the principled supplementation of training data where severe scarcity hinders discriminative learning.

Importantly, KSDiffusion targets realistic data regimes where data availability gradually transitions from sparse to sufficient, rather than extreme few-shot scenarios. Through systematic experiments across varying data scales, we demonstrate that KSDiffusion effectively improves prediction performance for data-scarce and imbalanced kinase groups while maintaining competitive accuracy when sufficient training data are available. These results highlight the regime-dependent effectiveness of diffusion-based augmentation and provide insights into when and why generative modeling can enhance phosphorylation site prediction. Overall, KSDiffusion offers a biologically informed and robust solution to the longstanding challenges of limited context integration and severe data imbalance. Ultimately, it facilitates a more reliable exploration of kinase-regulated signaling networks.

The main contributions of this work are summarized as follows:

• Problem-driven formulation under realistic data regimes: We systematically investigate kinase-specific prediction under data-limited and highly imbalanced conditions, clarifying the limitations of purely discriminative models and identifying regimes where generative augmentation becomes beneficial.

• A task-aware conditional diffusion framework: We propose KSDiffusion, a unified framework integrating protein language model-based representations with conditional diffusion modeling to generate kinase-consistent synthetic samples tightly coupled to the downstream prediction task.

• Effective integration with discriminative learning: By jointly leveraging original and diffusion-generated samples, KSDiffusion improves generalization for data-scarce kinase groups while maintaining competitive performance when sufficient training data are available.

• Comprehensive evaluation and regime-aware insights: Extensive experiments across varying data scales demonstrate the effectiveness of KSDiffusion and further elucidate the relationship between data availability, generative quality, and downstream prediction performance.

Materials and methods

Dataset construction and problem setting

To support kinase-specific phosphorylation site prediction under data-limited and highly imbalanced regimes, we construct two complementary datasets: (i) a general phosphorylation site dataset, used for representation adaptation and diffusion pretraining; and (ii) a kinase-specific phosphorylation site dataset, used for conditional generation and downstream evaluation. This dual-dataset design enables the model to capture global phosphoproteomic patterns required for learning general phosphorylation-related representations, while simultaneously preserving kinase-specific regulatory characteristics essential for conditional modeling and prediction.

General phosphorylation site dataset

The general phosphorylation site dataset was derived from the PhosphoSitePlus database [3]. All serine (S), threonine (T), and tyrosine (Y) residues across annotated proteins were treated as candidate sites. Diverging from the conventional 7 residue window commonly adopted in previous studies, we extended each candidate site by 30 amino acids in both upstream and downstream directions, resulting in fixed-length peptide sequences of 61 residues. Sequence boundaries were padded with the placeholder symbol “X” when necessary. This design aims to capture richer contextual information surrounding phosphorylation events, including longer-range sequence dependencies and potential structural signals [32].

Experimentally validated phosphorylation sites were treated as positive samples, while unannotated S/T/Y residues were used as negative candidate sites. To reduce redundancy and mitigate bias introduced by homologous sequences, CD-HIT clustering was applied with a sequence identity threshold of 90% [33]. The final dataset statistics are summarized in Table 1, comprising 117 360 phosphorylated sites and 120 001 non-phosphorylated sites. This dataset is used to adapt the protein language model to phosphorylation-related sequence patterns and to pretrain the diffusion model on large-scale phosphoproteomic distributions.

Table 1.

Statistics of the general phosphorylation site dataset.

Category	Count
Phosphorylation site	117 360
Non-phosphorylated site	120 001

Kinase-specific phosphorylation site dataset

Kinase-specific phosphorylation site annotations were collected from GPS 6.0, PhosphoSitePlus, and UniProtKB [22, 32, 34, 35]. Kinases were categorized according to the KinBase taxonomy and the grouping strategy adopted by KinasePhos 2.0, resulting in 10 major kinase groups [27, 36].

Peptide construction followed the same 30 residue extraction strategy described above to ensure consistency across datasets. Redundant sequences were removed using CD-HIT with a sequence identity threshold of 90%. Based on the number of available kinase–substrate annotations, kinase groups were further divided into three data-scale regimes:

• Small-scale groups: containing 500 samples (e.g. CK1, PKL, STE, and TKL).

• Medium-scale groups: containing 500–2000 samples (e.g. Atypical, CAMK, Other, and TK).

• Large-scale groups: containing 2000 samples (e.g. AGC and CMGC).

Detailed sample statistics for each kinase group are reported in Table 2. This dataset exhibits a pronounced long-tailed distribution, with a small number of kinase groups containing abundant annotations and the majority of groups remaining severely data-limited. Such imbalance reflects the current state of experimental kinase research and poses substantial challenges for kinase-specific prediction models.

Table 2.

Sample distribution of the kinase-specific phosphorylation site dataset across kinase groups and data-scale regimes.

Regime	Kinase group	Samples
Small-scale	CK1	245
	PKL	149
	STE	389
	TKL	221
Medium-scale	Atypical	710
	CAMK	1364
	Other	1366
	TK	1909
Large-scale	AGC	2893
	CMGC	3675

This dual-dataset design enables the model to capture general phosphorylation-related representations while preserving kinase-specific regulatory characteristics required for conditional modeling and prediction.

Problem formulation and framework overview

In kinase-specific phosphorylation site prediction, the task is formulated as a binary classification problem that distinguishes true kinase–substrate relationships from non-substrate candidate sites. This formulation preserves kinase specificity and avoids collapsing heterogeneous phosphorylation patterns across different kinase families.

A key characteristic of this problem is the highly uneven distribution of available data across kinase groups. While some kinases are well characterized with abundant experimentally validated substrates, many others remain severely underrepresented. This naturally results in a heterogeneous and highly imbalanced learning setting, where the number of positive samples varies dramatically across kinase groups.

Such data conditions impose significant challenges for standard discriminative models, particularly in low-data regimes where limited positive samples restrict the ability to learn robust and generalizable decision boundaries. These challenges motivate the need for learning strategies that can adapt to varying levels of data availability and effectively capture the underlying distribution of kinase–substrate relationships.

The objective of this study is to predict kinase-specific phosphorylation sites by determining whether a given residue within a substrate protein can be phosphorylated by a specific kinase.

Let denote the set of kinase groups. For each candidate phosphorylation site, we extract a fixed-length peptide sequence centered at the target residue, denoted as , where corresponds to a amino acid window. For a given kinase group , the task is to learn a kinase-specific prediction function , which estimates the probability that peptide is a true substrate of kinase .

The training data for each kinase group consists of a positive set and a negative set . In realistic biological settings, kinase-specific datasets are often highly imbalanced, particularly for underrepresented kinase groups where . This imbalance severely limits the generalization ability of purely discriminative models and motivates the need for data augmentation strategies that can expand the effective support of the positive class.

To address these challenges, we propose KSDiffusion, a unified framework that integrates representation learning, conditional generative modeling, and discriminative prediction. As illustrated in Fig. 1, the framework consists of three sequential stages:

Figure 1.

Overview of the proposed KSDiffusion framework. The model follows a three-stage pipeline comprising context-aware representation learning, kinase-conditioned diffusion-based data augmentation, and kinase-specific discriminative prediction. Peptide sequences are first encoded into embedding representations using a pretrained and fine-tuned protein language model. A conditional diffusion model then generates additional positive embeddings for underrepresented kinase groups. Finally, both real and generated embeddings are used to train a kinase-specific predictor.

• Context-aware representation learning: Raw peptide sequences are encoded into dense embedding representations using a protein language model, capturing evolutionary and structural dependencies.

• Conditional diffusion-based augmentation: A kinase-conditioned diffusion model is applied in the embedding space to generate synthetic positive samples, effectively enriching the distribution of kinase-specific substrates.

• Kinase-specific prediction: A discriminative classifier is trained on both original and generated samples to predict kinase–substrate relationships.

This integrated design enables KSDiffusion to adapt to varying levels of data availability while maintaining biological consistency in the learned representations.

Context-aware representation learning

To leverage shared structural patterns, the ESM-2 encoder is jointly fine-tuned on an aggregated training set containing samples from all kinase groups using masked language modeling (MLM) and SCL. By optimizing relative similarities rather than absolute class frequencies, the SCL objective ensures that even data-scarce kinases form compact, well-separated clusters in the embedding space, thereby significantly enhancing representation quality and downstream learning under severe class imbalance.

ESM-2-based sequence encoding

We employ ESM-2 (650 M parameters) as the backbone encoder for substrate peptide sequences. ESM-2 is a pretrained protein language model trained on millions of protein sequences using self-supervised objectives, enabling it to capture rich contextual information related to protein evolution and structure [37, 38].

Given an input peptide sequence centered at a candidate phosphorylation site, the sequence is first tokenized using the official ESM tokenizer and augmented with special tokens at the sequence boundaries. The tokenized sequence is then processed by the ESM-2 Transformer encoder to produce contextualized embeddings for each residue, along with a pooled sequence-level representation. In contrast to approaches that treat pretrained language models as frozen feature extractors, we fine-tune all ESM-2 parameters to adapt the representations to phosphorylation-related sequence patterns.

To preserve the generalization capability of the pretrained model while enhancing task relevance, we adopt a controlled fine-tuning strategy based on MLM [39, 40]. During training, a subset of amino acid residues is randomly masked, and the model is trained to reconstruct the masked tokens based on surrounding context. The MLM objective is defined as:

(1)

where indicates whether position in sample is masked, denotes the true amino acid token, and is the corresponding hidden representation produced by the encoder. This auxiliary objective encourages the model to retain sensitivity to local and long-range sequence dependencies during task adaptation.

Supervised contrastive fine-tuning

While MLM enhances contextual understanding at the sequence level, it does not explicitly enforce discrimination between kinase-specific phosphorylation patterns. To address this limitation, we further introduce SCL to fine-tune the sequence-level embeddings [41].

In the supervised contrastive setting, peptide samples associated with the same kinase are treated as positive pairs, whereas samples corresponding to different kinases or non-substrate sites are treated as negatives. Given a batch of sequence embeddings , the supervised contrastive loss is formulated as:

(2)

where denotes the set of indices corresponding to samples sharing the same kinase label as sample , and is a temperature parameter controlling the concentration of the embedding distribution.

This contrastive objective explicitly structures the embedding space according to kinase specificity, promoting intra-kinase compactness while significantly increasing inter-kinase separability.

Crucially, the ESM-2 encoder is not fine-tuned in isolation for each individual kinase group. Instead, the fine-tuning process (incorporating both MLM and SCL) is performed jointly across the entire aggregated training set comprising all available kinase groups. This joint training strategy is particularly vital for handling sequence differences under data-scarce conditions. By sharing encoder parameters across diverse kinase groups, underrepresented kinases can leverage the universal structural and physicochemical “grammar” of phosphorylation learned from data-rich groups, thereby preventing severe overfitting. Simultaneously, the SCL objective explicitly forces the model to capture the subtle sequence differences between various groups by pushing embeddings of different kinases apart. This ensures that the shared encoder does not collapse into a generic representation but strictly preserves inter-kinase separability, enabling the downstream diffusion model to capture distinct kinase-specific manifolds even when empirical samples are extremely limited.

Furthermore, this contrastive geometric regularization directly addresses the challenge of extreme data scarcity [41]. In standard cross-entropy-based training, rare kinase classes are frequently overwhelmed by majority classes, leading to collapsed or heavily skewed decision boundaries. SCL mitigates this fundamental issue by optimizing the representation space independently of absolute class frequencies. By forcing even a highly limited number of positive samples to form a tightly bound, well-separated cluster, SCL ensures that underrepresented kinases carve out a distinct and robust manifold in the high-dimensional space. This well-defined localized structure serves as a critical prerequisite for the subsequent generative stage: it provides the conditional diffusion model (KS-DiT) with a clear, localized target distribution to learn from. Consequently, the diffusion model can accurately sample from this structured latent space to generate biologically coherent synthetic samples, rather than producing out-of-distribution artifacts that would otherwise occur in a poorly clustered embedding space.

The overall fine-tuning objective of the encoder is defined as a weighted combination of the MLM and SCL losses:

(3)

where the loss weights control the balance between contextual reconstruction and kinase-specific discrimination. The resulting context-aware embeddings serve as a unified representation for both the conditional diffusion model and the downstream kinase-specific predictor, enabling consistent information flow across generative and discriminative components of KSDiffusion. Additional implementation details and training configurations for the representation learning module are provided in Supplementary Material section A.

Conditional diffusion for task-aware data augmentation

Learning kinase-specific phosphorylation patterns is fundamentally constrained by the severe scarcity and imbalance of experimentally validated kinase–substrate relationships. For kinase groups with limited annotations, purely discriminative models are prone to overfitting dominant patterns and often fail to generalize, while heuristic data augmentation strategies based on random perturbations or oversampling may introduce biologically implausible samples. To address this challenge in a principled manner, we introduce a conditional diffusion-based generative module that performs task-aware data augmentation in representation space.

Rather than generating peptide sequences directly, the proposed module operates on context-aware embeddings learned by the upstream encoder, enabling stable training and direct alignment with the downstream prediction task. By explicitly conditioning generation on kinase-specific information, the model captures group-dependent substrate preferences and produces synthetic samples that are both diverse and biologically consistent.

Conditional diffusion framework

Diffusion models generate samples by learning to reverse a gradual noising process that transforms data into noise through a sequence of stochastic steps [42–44]. Given a peptide-level embedding , the forward diffusion process incrementally adds Gaussian noise according to a predefined schedule:

(4)

where controls the noise intensity at each timestep .

In KSDiffusion, we adopt a conditional diffusion formulation, in which generation is explicitly conditioned on kinase-specific contextual information , such as the learned embedding of a kinase group. During the reverse process, a neural network is trained to progressively denoise while respecting the conditioning signal, thereby recovering peptide representations consistent with the target kinase distribution [45].

Importantly, the conditional diffusion model is not treated as a standalone generator. Instead, it serves as a task-aware augmentation mechanism, ensuring that generated samples preserve biologically meaningful characteristics relevant to kinase-specific phosphorylation site prediction. A detailed derivation of the diffusion objective and sampling procedure is provided in Supplementary Material section B.

KS-DiT architecture for kinase-conditioned generation

To efficiently implement conditional diffusion, we design a diffusion backbone termed KS-DiT, which follows a DiT-style architecture adapted for kinase-conditioned generation. KS-DiT operates directly on peptide embedding representations produced by the context-aware encoder, enabling generation in representation space rather than raw sequence space.

The reverse denoising process is parameterized by a neural network that predicts the injected noise conditioned on the current noisy state, timestep, and kinase context:

(5)

Kinase-specific conditioning is injected into the denoising network through learned embeddings and modulation mechanisms, allowing intermediate feature representations to adapt dynamically to different kinase groups. Compared with standard DiT architectures, KS-DiT is tailored to capture kinase-dependent variations while maintaining a compact and stable structure suitable for data-limited regimes.

Figure 2 illustrates a simplified schematic of the conditional diffusion framework, emphasizing the forward noising process and the reverse generation process guided by the kinase-specific condition. The complete architectural and implementation details, including the layer-wise adaptive modulation mechanism, are provided in Supplementary Material section C.

Figure 2.

Simplified schematic of the conditional diffusion framework (KS-DiT). The forward process gradually adds noise to the initial peptide embedding , while the reverse denoising process reconstructs the biologically meaningful embedding from pure noise . This reverse generation is explicitly guided by the kinase-specific condition , ensuring the synthetic samples align with the target kinase group. Detailed architectural schematics of the internal conditional noise prediction are provided in Fig. S1 of Supplementary Material section C.

Task-aware data augmentation strategy

The conditional diffusion model is employed to generate synthetic peptide embeddings for kinase groups with limited positive samples. Generated representations are combined with original training data to form an augmented dataset for the downstream kinase-specific predictor. Unlike conventional augmentation methods that apply uniform perturbations, KSDiffusion leverages conditional generation to ensure that augmented samples remain consistent with kinase-specific phosphorylation patterns.

To further align augmentation with realistic data regimes, we introduce a regime-aware weighting strategy that controls the contribution of generated samples based on the availability of empirical data. For a given kinase group , the effective training set is defined as:

(6)

where denotes the original positive samples, represents generated samples, and is a data-dependent coefficient that decreases as the number of real samples increases.

In addition, to avoid the introduction of low-quality or redundant synthetic samples, generated embeddings are filtered using a combination of classifier-based confidence scores and diversity constraints. The evaluation and selection criteria for generated samples are detailed in Supplementary Material section D. Together, these strategies enable KSDiffusion to enhance generalization in data-scarce kinase groups while avoiding performance degradation when sufficient empirical data are available.

Integration with kinase-specific prediction

The primary objective of KSDiffusion is to improve kinase-specific phosphorylation site prediction under data-limited and highly imbalanced conditions. To achieve this goal, the proposed framework tightly integrates context-aware representation learning, conditional diffusion-based generation, and a downstream discriminative predictor into a unified learning pipeline.

For each substrate peptide–kinase pair, biologically meaningful embeddings are first produced by the ESM-2–based context-aware encoder. This shared embedding space serves as the common interface between the generative and discriminative components. The downstream predictor operates directly on these embeddings to estimate the probability that a given residue is phosphorylated by a specific kinase, ensuring semantic consistency between generation and prediction.

For kinase groups with limited training data, the conditional diffusion model is employed to generate additional synthetic embeddings conditioned on kinase-specific information. These generated samples are combined with the original training data to form an augmented dataset. Importantly, diffusion-generated samples are not treated as independent perturbations or noise-injected observations. Instead, they are explicitly constrained to follow the embedding distribution learned from real kinase–substrate relationships, thereby preserving task-relevant biological characteristics.

By jointly leveraging empirical and generated samples, the kinase-specific predictor is able to learn more robust decision boundaries for underrepresented kinase groups while maintaining discriminative performance for well-represented kinases. As data availability increases, the contribution of generated samples naturally diminishes, allowing the predictor to rely predominantly on real observations. This integrated design enables KSDiffusion to adapt seamlessly to varying data regimes without requiring separate models or heuristic switching strategies, effectively balancing generative support and discriminative learning.

Training strategy overview

KSDiffusion is trained using a staged strategy that coordinates representation learning, conditional generation, and downstream prediction.

First, the ESM-2 encoder is adapted using the General Phosphorylation Site Dataset (Table 1) to produce context-aware peptide embeddings. The diffusion backbone is pretrained on the same dataset to capture global phosphoproteomic distributions.

Second, the pretrained diffusion model is fine-tuned in a kinase-conditioned manner (KS-DiT) using the Kinase-specific Phosphorylation Site Dataset (Table 2), enabling it to model kinase-dependent substrate distributions.

Finally, a kinase-specific predictor is trained using both original samples from the Kinase-specific Phosphorylation Site Dataset and diffusion-generated synthetic samples to learn discriminative decision boundaries.

These stages are tightly coupled: the encoder provides structured representations for diffusion modeling, the diffusion module expands the effective support of kinase-specific substrates, and the classifier leverages both to improve generalization under varying data regimes. A detailed description of data usage, training procedures, and implementation settings is provided in the Experimental setup section.

Results

Experimental setup

To comprehensively evaluate the proposed KSDiffusion framework, we designed a unified experimental protocol encompassing overall performance comparison, robustness analysis under data-limited and imbalanced regimes, and ablation studies. All experiments were conducted on kinase-specific phosphorylation site datasets described below, using consistent data partitioning, training strategies, and evaluation metrics to ensure fair and reproducible comparisons.

Experimental configuration

To comprehensively evaluate the proposed KSDiffusion framework, we design a unified experimental protocol encompassing overall performance comparison, robustness analysis under data-limited and imbalanced regimes, and ablation studies.

We compare KSDiffusion with representative baseline methods spanning different methodological paradigms, including motif-based approaches (e.g. GPS series), traditional machine learning models with handcrafted features (e.g. KinasePhos 3.0), and deep learning-based sequence models (e.g. Phosformer). All methods are evaluated under identical data splits and evaluation protocols to ensure fair comparison.

Model performance is evaluated using standard classification metrics, including AUC, PR-AUC, F1-score, accuracy, and false positive rate (FPR). To further assess practical utility, we additionally report reclassification-based metrics, including the categorical Net Reclassification Index (cNRI) and Integrated Discrimination Improvement (IDI).

All experiments were conducted on a workstation equipped with four NVIDIA GeForce RTX 3090 Ti GPUs (24 GB memory each). The implementation was based on Python 3.12.0 and PyTorch 2.5.1.

Data construction and usage

Two datasets are used in this study with distinct roles across different stages: a General Phosphorylation Site Dataset (Table 1) and a Kinase-specific Phosphorylation Site Dataset (Table 2), as described in the Dataset construction and problem setting section.

The Kinase-specific Phosphorylation Site Dataset is partitioned into training, validation, and test sets prior to any model training, using a stratified 8:1:1 split while preserving class distributions. The test set remains completely unseen throughout all stages. The training set is used for conditional diffusion training and downstream predictor learning, while the validation set is reserved exclusively for model selection in the final prediction stage.

The General Phosphorylation Site Dataset is used for encoder adaptation and diffusion pretraining. Importantly, it is constructed by explicitly excluding all protein families associated with the kinase-specific dataset, ensuring strict separation between pretraining data and downstream evaluation.

To address class imbalance and enforce kinase-specific discrimination, we adopt a dual negative sampling strategy that reflects two distinct levels of learning difficulty:

• Simple negatives are sampled from non-phosphorylated S/T/Y residues. These samples primarily guide the model to distinguish phosphorylatable substrates from general non-substrate sequences, capturing fundamental structural and physicochemical constraints.

• Hard negatives are sampled from experimentally validated phosphorylation sites associated with other kinase groups. These samples introduce inter-kinase ambiguity, forcing the model to learn fine-grained, kinase-specific patterns rather than relying on generic phosphorylation signals.

This design is particularly important in kinase-specific prediction tasks, where many phosphorylation sites share common biochemical properties across kinases. Without hard negatives, the model may overfit to general phosphorylation features and fail to capture the subtle distinctions required for accurate kinase-specific classification. The ratio between simple and hard negatives is fixed at 1:1 to balance global discrimination and kinase-specific differentiation.

Training pipeline and implementation

This section provides the detailed implementation corresponding to the training strategy outlined in the Training strategy overview section. The training process involves three stages and four key components.

First, the ESM-2 encoder is adapted using the General Phosphorylation Site Dataset (Table 1) to produce context-aware peptide embeddings. Then, the diffusion backbone is pretrained on the same dataset to capture global phosphoproteomic distributions.

Second, the pretrained diffusion model is further fine-tuned in a kinase-conditioned manner (KS-DiT) using the training partition of the Kinase-specific Phosphorylation Site Dataset (Table 2), enabling it to model kinase-dependent substrate distributions.

Finally, the kinase-specific predictor is trained in a supervised manner using both original samples from the kinase-specific training set and diffusion-generated synthetic samples. The validation set is used for model selection, while the test set remains strictly held out.

The diffusion model is trained using the standard noise prediction objective:

(7)

where denotes the noisy embedding at timestep , and is the predicted noise.

Key hyperparameters include a diffusion length of . This choice reflects a practical trade-off between approximation fidelity and computational efficiency in the ESM-2 embedding space. Compared with pixel-space diffusion, where longer schedules are often required, the continuous and structured nature of protein embeddings allows effective modeling with a shorter diffusion trajectory. In practice, is sufficient to capture the underlying distribution while avoiding over-smoothing of biologically relevant features. Additional implementation details, including the meta-learning-based initialization [46, 47] and adversarial fine-tuning [48, 49], are provided in Supplementary Material section E.

Sample generation and selection

After training, the conditional diffusion model is used to generate synthetic embeddings for data-scarce kinase groups. To ensure both quality and diversity, each generated sample is evaluated using a classifier-based quality score:

(8)

and a diversity score based on cosine similarity:

(9)

The final selection score is defined as:

(10)

and only samples exceeding predefined thresholds are retained. This strategy mitigates mode collapse and ensures that augmented samples remain informative and non-redundant [50, 51]. Further details regarding the quality control thresholds and selection procedures are provided inSupplementary Material section D.

Overall comparison with baseline methods

To evaluate the overall effectiveness of the proposed KSDiffusion framework, we conducted comprehensive comparisons against representative baseline methods across 10 kinase groups spanning low-, medium-, and large-data regimes. The selected baselines include motif-based predictors, classical machine learning approaches, and recent deep learning models, thereby providing a broad and fair reference for performance assessment.

Overall, KSDiffusion demonstrates a clear and consistent performance advantage across kinase groups of varying data scales, with the most pronounced gains observed in low-data regimes and stable performance maintained as data availability increases. This trend highlights the effectiveness of conditional diffusion as a targeted solution for data-scarce and imbalanced kinase groups, rather than as a generic replacement for large-scale empirical data.

To visually interpret the quantitative results summarized in Table 3, we present ROC curves for representative kinase groups selected from each data regime in Fig. 3. As shown in Fig. 3a, the ROC curve for the CK1 group (chosen as a representative case for the low-data regime) illustrates a substantial expansion in the area under the curve compared with baseline methods. This visualization aligns with the aggregate trend observed across all low-data groups (AUC gains of 8%–15%), confirming that KSDiffusion effectively learns robust decision boundaries even with limited samples. Similarly, Fig. 3b and c illustrates the performance characteristics in medium-data (exemplified by the Atypical group) and large-data (exemplified by the AGC group) regimes, respectively. These plots further corroborate the regime-dependent effectiveness of our generative augmentation strategy, showing consistent improvements in medium-data settings and competitive robustness in large-data scenarios where baselines approach saturation.

Table 3.

Comprehensive performance comparison across all kinase groups and data regimes.

Method	AUC	PR	F1	ACC	FPR
Low-data regime
CK1 Group
GPS	0.7398	0.7101	0.7146	0.6586	0.5379
KinasePhos (SVM)	0.5471	0.5852	0.6666	0.5000	0.9221
KinasePhos (XGB)	0.5975	0.5758	0.6956	0.5871	0.8130
Phosformer	0.7222	0.7176	0.7073	0.6689	0.4620
KSDiffusion (Ours)	0.8892	0.9084	0.8113	0.8000	0.2600
PKL Group
GPS	0.8446	0.8761	0.7999	0.8265	0.0408
KinasePhos (SVM)	0.7216	0.6825	0.7407	0.7211	0.3660
KinasePhos (XGB)	0.8032	0.7890	0.7547	0.7451	0.3220
Phosformer	0.8100	0.8248	0.7592	0.7346	0.3673
KSDiffusion (Ours)	0.9020	0.9024	0.8400	0.8400	0.1600
STE Group
GPS	0.6765	0.6487	0.6874	0.5830	0.7508
KinasePhos (SVM)	0.5836	0.6631	0.6756	0.5264	0.9614
KinasePhos (XGB)	0.5040	0.5672	0.6666	0.5114	0.9554
Phosformer	0.5262	0.5600	0.6658	0.5051	0.9757
KSDiffusion (Ours)	0.7747	0.8011	0.7262	0.6406	0.6719
TKL Group
GPS	0.5775	0.5865	0.6685	0.5041	0.9917
KinasePhos (SVM)	0.6670	0.6801	0.6724	0.6217	0.5412
KinasePhos (XGB)	0.6115	0.6895	0.6711	0.5160	0.9854
Phosformer	0.6118	0.6087	0.6934	0.5578	0.8842
KSDiffusion (Ours)	0.7856	0.8090	0.7213	0.6600	0.5600
Medium-data regime
Atypical Group
GPS	0.8672	0.8850	0.7999	0.8028	0.1830
KinasePhos (SVM)	0.7882	0.8108	0.7674	0.7138	0.4929
KinasePhos (XGB)	0.8665	0.8926	0.7922	0.7746	0.3098
Phosformer	0.8879	0.9006	0.8111	0.8098	0.1971
KSDiffusion (Ours)	0.9018	0.9160	0.8261	0.8305	0.1441
CAMK Group
GPS	0.7943	0.7744	0.7465	0.7299	0.3357
KinasePhos (SVM)	0.7914	0.7623	0.7702	0.7509	0.3358
KinasePhos (XGB)	0.7705	0.7648	0.7432	0.7216	0.3649
Phosformer	0.7740	0.7736	0.7359	0.6739	0.5620
KSDiffusion (Ours)	0.8054	0.8103	0.7474	0.7291	0.3436
Other Group
GPS	0.7432	0.7473	0.6926	0.6548	0.4587
KinasePhos (SVM)	0.6488	0.6556	0.6976	0.6204	0.6350
KinasePhos (XGB)	0.6925	0.6884	0.7030	0.6423	0.5620
Phosformer	0.6622	0.6430	0.7046	0.5839	0.8248
KSDiffusion (Ours)	0.7583	0.7851	0.7004	0.6608	0.4714
TK Group
GPS	0.9559	0.9562	0.9259	0.9267	0.0628
KinasePhos (SVM)	0.9502	0.9406	0.9234	0.9214	0.1047
KinasePhos (XGB)	0.9450	0.9347	0.9077	0.9031	0.1465
Phosformer	0.9645	0.9639	0.9305	0.9293	0.0890
KSDiffusion (Ours)	0.9285	0.9392	0.8742	0.8774	0.0975
Large-data regime
AGC Group
GPS	0.8630	0.8481	0.8020	0.7982	0.2206
KinasePhos (SVM)	0.7980	0.7688	0.7545	0.7202	0.4206
KinasePhos (XGB)	0.7940	0.7561	0.7488	0.7115	0.4379
Phosformer	0.8432	0.8346	0.7858	0.7702	0.3034
KSDiffusion (Ours)	0.8708	0.8794	0.7925	0.7946	0.1950
CMGC Group
GPS	0.9173	0.9090	0.8490	0.8410	0.2119
KinasePhos (SVM)	0.8508	0.8444	0.7899	0.7562	0.2693
KinasePhos (XGB)	0.8680	0.8384	0.7715	0.7785	0.2802
Phosformer	0.9135	0.9027	0.8706	0.8734	0.1059
KSDiffusion (Ours)	0.8757	0.8847	0.7950	0.7868	0.2533

Best results for each metric are highlighted in bold (highest for AUC, PR, F1, ACC; lowest for FPR).

Figure 3.

Representative ROC curves illustrating the regime-dependent performance of KSDiffusion compared with baseline methods. (a) In the low-data regime (exemplified by the CK1 group), KSDiffusion shows a significant AUC improvement, visually confirming the statistical gains reported in Table 3. (b) In the medium-data regime (exemplified by the Atypical group), the proposed method maintains a clear performance lead. (c) In the large-data regime (exemplified by the AGC group), KSDiffusion achieves competitive performance comparable to state-of-the-art methods. These plots visualize the consistent trend that generative augmentation is most beneficial under data scarcity.

Beyond overall discrimination ability, improvements in PR-AUC and F1 score further confirm the robustness of KSDiffusion under class imbalance. In several low-data kinase groups, PR-AUC increases by ¿10 percentage points, indicating a significantly enhanced ability to recover true phosphorylation sites while effectively controlling false positives. These consistent gains across complementary metrics demonstrate that KSDiffusion not only improves global ranking performance but also enhances practical prediction reliability.

Taken together, the results in Table 3 reveal a data-scale-dependent performance profile, where the benefits of diffusion-based augmentation are most pronounced when positive samples are scarce and gradually diminish as data availability increases. This systematic behavior aligns with the design motivation of KSDiffusion and provides empirical evidence that conditional diffusion is particularly well suited for kinase groups suffering from limited and imbalanced training data.

Performance under data-limited and imbalanced regimes

Low-data kinase groups

Low-data kinase groups, including CK1, PKL, STE, and TKL, represent the most challenging prediction scenarios, with training sample sizes typically ¿500. In these settings, model performance is highly sensitive to data scarcity and class imbalance. As shown in Table 3 (Low-data regime), KSDiffusion consistently and substantially outperforms all baseline methods across these kinase groups.

Specifically, KSDiffusion improves AUC-ROC by 12%–15% relative to the best-performing baselines. Corresponding gains in F1 score range from 8% to 12%, reflecting a more balanced trade-off between precision and recall. Improvements in PR-AUC are particularly pronounced, with absolute increases exceeding 10% in several cases. These results indicate that conditional diffusion-based augmentation effectively enriches the positive sample distribution, enabling the classifier to learn more stable and biologically meaningful decision boundaries under extreme data scarcity.

Medium-data kinase groups

For medium-data kinase groups, including Atypical, CAMK, Other, and TK, where training sample sizes range from 500 to 2000, KSDiffusion continues to outperform baseline methods, albeit with reduced margins. As reported in Table 3 (Medium-data regime), AUC–ROC improvements are typically within the range of 4%–7%, while gains in PR-AUC and F1 score are more modest, generally 3%–5%.

These results suggest that diffusion-based augmentation remains beneficial in moderately sized datasets, particularly for mitigating residual imbalance effects. However, as the number of real positive samples increases, model performance gradually transitions from being constrained by data scarcity to being dominated by representation quality. In this regime, synthetic samples play a complementary role, refining decision boundaries rather than fundamentally reshaping the learned feature space.

Large-data kinase groups

Large-data kinase groups, such as AGC and CMGC, contain >2000 training samples and represent comparatively well-characterized phosphorylation systems. As shown in Table 3 (Large-data regime), most baseline methods already achieve strong performance in these groups, with AUC–ROC values approaching saturation.

Under these conditions, KSDiffusion consistently matches or slightly exceeds the strongest baselines, with AUC improvements typically within 1%–2%. Although the marginal benefit of data augmentation becomes limited, these results demonstrate that the proposed framework does not compromise performance in data-rich scenarios. Importantly, competitive PR-AUC and F1 scores are preserved, indicating that diffusion-based augmentation is self-regularizing rather than intrusive when sufficient real data is available.

Data-scale-dependent performance summary

Across all kinase groups, KSDiffusion exhibits a clear data-scale-dependent performance pattern: substantial improvements in low-data regimes, moderate gains in medium-data regimes, and stable performance in large-data regimes. This systematic behavior supports the central claim that conditional diffusion is particularly effective as a targeted strategy for kinase groups affected by data scarcity and imbalance, while remaining safe and non-degradative when applied to well-sampled datasets. Together, these findings validate the design rationale of KSDiffusion and underscore its practical applicability across diverse kinase prediction scenarios.

Ablation study

To systematically investigate the contribution of each core component in the proposed KSDiffusion framework, we conducted a comprehensive ablation study. The experiments are designed with two complementary objectives: (i) to evaluate the effectiveness of diffusion-based generative augmentation compared with conventional imbalance-handling strategies, and (ii) to assess the contribution of key architectural components within the proposed framework. The results are summarized in Table 4.

Table 4.

Ablation experiment results across three representative data regimes.

Method	AUC	PR	F1	ACC	FPR
Low-data regime
w/o Diffusion (BCE Baseline)	0.6996	0.7312	0.6901	0.5600	0.8600
replace w/ Focal Loss	0.6108	0.6476	0.6761	0.5400	0.8800
replace w/ SMOTE	0.6376	0.6230	0.6533	0.5900	0.6600
Frozen Encoder	0.7308	0.7120	0.7054	0.6500	0.7500
w/o Meta-learning	0.7398	0.7487	0.7209	0.5200	0.7200
w/o Adversarial	0.7234	0.7132	0.7156	0.5800	0.7500
w/o Sample selection	0.7419	0.7229	0.7008	0.5800	0.7300
Full Model	0.7856	0.8090	0.7213	0.6600	0.5600
Medium-data regime
w/o Diffusion (BCE Baseline)	0.8994	0.9052	0.8078	0.7924	0.2881
replace w/ Focal Loss	0.7698	0.7637	0.7152	0.6356	0.6441
replace w/ SMOTE	0.7444	0.7352	0.5222	0.6356	0.5271
Frozen Encoder	0.8683	0.8865	0.7622	0.7914	0.2514
w/o Meta-learning	0.8798	0.8921	0.7556	0.8012	0.2679
w/o Adversarial	0.8577	0.8812	0.7816	0.8198	0.2423
w/o Sample selection	0.8624	0.8798	0.7709	0.8051	0.2392
Full model	0.9018	0.9160	0.8261	0.8305	0.1441
Large-data regime
w/o Diffusion (BCE Baseline)	0.8527	0.8661	0.7752	0.7810	0.2628
replace w/ Focal Loss	0.8287	0.8204	0.7814	0.7696	0.2843
replace w/ SMOTE	0.8294	0.8297	0.7472	0.7778	0.2713
Frozen Encoder	0.8115	0.8527	0.7525	0.6585	0.2875
w/o Meta-learning	0.8515	0.8417	0.7524	0.7617	0.3002
w/o Adversarial	0.8221	0.8726	0.7191	0.7681	0.2870
w/o Sample selection	0.8356	0.8515	0.7322	0.7502	0.2521
Full model	0.8757	0.8847	0.7950	0.7868	0.2533

Best results for each metric are highlighted in bold.

To address the first objective (Table 4, Rows 1–3 versus Full Model), we construct a strict non-generative baseline denoted as w/o Diffusion (BCE Baseline). This baseline employs the same downstream classifier and weighted Binary Cross-Entropy (BCE) loss as the Full Model, but is trained solely on the original imbalanced data without any diffusion-based augmentation. On top of this baseline, we further introduce two representative non-generative strategies for handling class imbalance: Focal Loss (cost-sensitive reweighting) and SMOTE (linear interpolation-based oversampling in the ESM-2 embedding space). This design enables a direct and controlled comparison between diffusion-based augmentation and conventional alternatives under an identical predictive architecture.

To address the second objective (Table 4, Rows 4–7 versus Full Model), we perform ablations on individual components of the proposed framework. Specifically, Row 4 corresponds to freezing the ESM-2 encoder, thereby removing task-aware representation adaptation. Rows 5–7 are obtained by removing one component at a time from the Full Model, including the meta-learning strategy, adversarial fine-tuning, and dynamic sample selection module, respectively. These controlled ablations allow us to isolate the contribution of each component to the overall performance.

Impact of core components in low-data regimes

In low-data regimes, where both data scarcity and class imbalance are most severe, traditional non-generative strategies show limited effectiveness (Table 4). Compared with the class-weighted BCE baseline (AUC = 0.6996), applying Focal Loss and SMOTE directly to the ESM-2 embedding space leads to severe performance degradation, yielding AUC scores of 0.6108 and 0.6376, respectively.

This suggests that, under extreme data scarcity, cost-sensitive reweighting may lead to overfitting on a small number of positive samples, while linear interpolation in SMOTE may not adequately capture the complex structure of the high-dimensional ESM-2 embedding space. In contrast, the conditional diffusion module provides a more flexible way to enrich the support of the positive class distribution, leading to improved performance.

Crucially, removing adversarial fine-tuning (w/o Adversarial) yields the most substantial drop among the internal modules (AUC decreases to 0.7234). This validates our hypothesis that unconstrained diffusion may generate mathematically valid but biologically implausible sequences; adversarial guidance is indispensable for enforcing physical constraints and biological fidelity. Finally, ablating the dynamic screening mechanism (w/o Sample selection) degrades the AUC to 0.7419, indicating that without explicit quality and diversity filters, the inclusion of noisy or redundant synthetic samples can easily corrupt the delicate decision boundaries in low-data regimes.

Component contributions in medium-data regimes

In the medium-data regime, the dynamics of component contributions undergo a discernible shift (Table 4). Interestingly, the standard class-weighted BCE baseline achieves a highly competitive AUC of 0.8994, indicating that a moderate number of empirical samples provides a reasonable foundation for discrimination. However, forcefully applying Focal Loss or SMOTE to this space leads to drastic performance deterioration, plunging the AUC to 0.7698 and 0.7444.

This observation indicates that, even when data availability improves, simple reweighting or linear interpolation strategies may still be insufficient to model the underlying data distribution. Instead, maintaining the quality and consistency of representations—through adversarial fine-tuning and dynamic sample selection—becomes increasingly important.

Consequently, as empirical data availability improves, the primary role of the internal architectural modules shifts from basic distribution support to strict generative quality assurance. As demonstrated in Table 4, removing adversarial fine-tuning (w/o Adversarial) or dynamic screening (w/o Sample selection) causes the AUC to regress to 0.8577 and 0.8624, respectively—dropping significantly below the non-generative BCE baseline. This emphatically proves that without rigorous adversarial physical constraints and diversity-aware filtering, unconstrained generative augmentation risks injecting out-of-distribution noise that corrupts the already established decision boundaries. Thus, these internal modules are paramount for ensuring that synthetic samples remain strictly beneficial and biologically coherent.

Ablation results in large-data regimes

In the large-data regime, where empirical training samples are relatively abundant, the standard class-weighted BCE baseline naturally establishes a robust discriminative performance (AUC = 0.8527, Table 4). Replacing the diffusion module with Focal Loss or SMOTE results in lower AUC values (0.8287 and 0.8294) compared with the full model (0.8757).

This suggests that even in data-rich scenarios, naive reweighting or interpolation may introduce biases that are not well aligned with the complex structure of protein representations. In contrast, the diffusion-based approach remains stable and continues to provide performance gains without degrading model reliability.

Crucially, the ablation of internal modules further emphasizes that generative augmentation in large-data scenarios must be strictly regularized. The Frozen Encoder variant suffers a severe performance drop (AUC = 0.8115), reiterating that generic protein language representations must be fine-tuned for kinase-specific tasks regardless of the available data volume. Furthermore, removing adversarial fine-tuning (w/o Adversarial, AUC = 0.8221) or dynamic screening (w/o Sample selection, AUC = 0.8356) causes the performance to fall considerably below the non-generative BCE baseline. This highlights a profound conclusion: when original data are abundant, injecting unconstrained or noisy synthetic samples actively dilutes the high-quality empirical data. Only through rigorous biological constraints and diversity screening can generative augmentation safely surpass a strong large-data baseline. Interestingly, the removal of meta-learning (w/o Meta-learning, AUC = 0.8515) results in a noticeably milder performance drop compared with its impact in the low-data regime. This observation perfectly aligns with the theoretical foundation of meta-learning, which acts as a few-shot initialization technique and naturally exhibits diminishing returns as empirical data becomes sufficiently abundant.

Summary of ablation findings

Across all evaluated data regimes, replacing the conditional diffusion module with traditional non-generative imbalance-handling strategies (Focal Loss or SMOTE) consistently results in inferior performance (Table 4). These extensive comparisons conclusively demonstrate that in the highly non-linear, high-dimensional ESM-2 embedding space, naive cost-sensitive reweighting or linear interpolation fails to capture the true topological distribution of protein substrates. In contrast, diffusion-based augmentation provides a mathematically principled and flexible mechanism for modeling complex structural manifolds, successfully enriching the minority class support without distorting the established decision boundaries.

Crucially, the ablation study further reveals that generative augmentation alone is not a universal panacea; its success is strictly predicated on the rigorous internal architectural safeguards embedded within KSDiffusion. The consistent performance degradation observed upon removing task-aware encoder fine-tuning, meta-learning initialization, adversarial fine-tuning, or dynamic sample selection proves that these components are highly indispensable. Together, they form a robust quality assurance pipeline: meta-learning enables rapid adaptation under extreme scarcity, while adversarial guidance and explicit diversity screening prevent the injection of biologically implausible or topologically destructive noise. Ultimately, it is this precise synergy between manifold-aware generative augmentation and strict biological regularization that empowers KSDiffusion to achieve robust and state-of-the-art performance across diverse data environments.

Biological plausibility and motif enrichment of novel candidate substrates

Standard evaluation metrics such as AUC and PR-AUC primarily assess global ranking performance, but they do not fully characterize the biological validity of individual predictions. In phosphorylation site prediction, this limitation is particularly relevant because available phosphoproteomic annotations are inherently incomplete, and many true kinase–substrate relationships remain unobserved. As a result, a subset of high-confidence predictions labeled as false positives may correspond to biologically plausible yet currently unannotated substrates.

To investigate this aspect, we analyze the sequence-level properties of high-confidence predicted sites that are not annotated as positives in the test set. Specifically, we select unannotated sites with predicted probabilities and examine their position-specific amino acid distributions. To ensure coverage across different data regimes, we consider three representative kinase groups: TKL (low-data), Atypical (medium-data), and CMGC (large-data).

The resulting sequence logos (Fig. 4) reveal clear and biologically consistent motif patterns. For the CMGC group (Fig. 4c), which is characterized by proline-directed kinases, the predicted sites exhibit a strong enrichment of Proline (P) at the +1 position, consistent with well-established substrate specificity. For the TKL group (Fig. 4a), the predicted sites show a preference for hydrophobic residues, including Leucine (L) at the −2 position and Valine (V) at the −1 position, aligning with known structural binding tendencies. In the Atypical group (Fig. 4b), distinct enrichment of Glycine (G) at the −1 position and Arginine (R) at the −3 position is observed, reflecting family-specific sequence preferences.

Figure 4.

Motif enrichment analysis of high-confidence false positive (FP) predictions. Sequence logos were generated for unannotated test samples with a predicted probability . (a) The TKL group FPs exhibit a preference for distinct local sequence contexts, notably Leucine (L) at the −2 position and Valine (V) at the −1 position. (b) The Atypical group FPs display a motif enriched with Glycine (G) at the −1 position and Arginine (R) at the −3 position. (c) The CMGC group FPs show a canonical proline-directed motif with a dominant Proline (P) at the +1 position. This robust consistency across diverse kinase families demonstrates that KSDiffusion identifies biologically plausible candidate substrates.

These observations suggest that the model captures meaningful sequence determinants associated with kinase specificity, even for sites not annotated as positives. Rather than being arbitrarily distributed, high-confidence predictions tend to follow known motif patterns characteristic of their respective kinase families. This indicates that the model’s predictions are grounded in biologically relevant features and may provide useful candidates for further experimental validation.

Overall, this analysis complements standard performance metrics by offering a sequence-level perspective on prediction quality, highlighting the model’s ability to recover biologically meaningful patterns beyond annotated datasets.

Evaluation of practical utility via reclassification metrics

While metrics such as AUC and F1 score effectively evaluate global ranking and discrimination ability, they may not fully capture the practical utility of a model in real-world screening or decision-making scenarios, where the correctness of individual predictions and their associated confidence are critical. To further assess whether KSDiffusion improves prediction reliability at the sample level, we additionally evaluate our framework using two complementary reclassification metrics: the cNRI and the IDI.

The cNRI evaluates whether the new model improves discrete classification outcomes relative to a baseline under a fixed decision threshold. To avoid potential bias introduced by data-dependent threshold selection, we adopt a standard binary decision boundary of 0.5. The metric quantifies the net proportion of correctly reclassified samples by measuring upward movements for true positives and downward movements for true negatives:

(11)

where and denote true positive substrates and true negative non-substrates, respectively. and represent the fractions of samples that successfully transition to a higher () or lower () predicted probability tier when switching from the baseline to KSDiffusion.

Complementarily, the IDI measures the overall improvement in the separation between positive and negative prediction distributions:

(12)

where the bars denote average predicted probabilities over positive and negative samples for the new and baseline models, respectively. To provide a representative evaluation across different data regimes, we conduct this analysis on three kinase groups: TKL (low-data), Atypical (medium-data), and CMGC (large-data).

As summarized in Table 5, KSDiffusion consistently achieves positive improvements across both metrics and all data regimes. For example, under the fixed threshold of 0.5, the categorical NRI reaches 0.1281, 0.0917, and 0.1776 for the low-data (TKL), medium-data (Atypical), and large-data (CMGC) groups, respectively, indicating systematic improvements in discrete classification decisions. The IDI results (0.1230, 0.0600, and 0.1519) demonstrate a consistent increase in the separation between positive and negative prediction distributions.

Table 5.

Practical utility evaluation of KSDiffusion using cNRI and IDI, where positive values indicate a net improvement over the standard Weighted BCE model baseline.

Data regime	Kinase group	NRI	IDI
Low-data	TKL	0.1281	0.1230
Medium-data	Atypical	0.0917	0.0600
Large-data	CMGC	0.1776	0.1519

Overall, although the absolute magnitude of improvement varies across data regimes, the consistent positive trends across NRI and IDI indicate that diffusion-based augmentation not only improves global discrimination performance but also enhances the reliability and practical usability of individual predictions.

Analysis of diffusion-based augmentation strategies

To further investigate the behavior of the proposed diffusion-based data augmentation strategy beyond overall performance comparisons and ablation analyses, we conduct a set of controlled experiments focusing on (i) the effect of different generation ratios under varying data regimes and (ii) the generalization capability of the conditional diffusion module when integrated into an alternative backbone architecture. These experiments aim to provide a more fine-grained empirical understanding of when and how diffusion-based augmentation contributes to kinase–substrate relationship prediction.

Effect of diffusion-based generation ratios

We systematically evaluate the impact of diffusion-based positive sample generation under three augmentation ratios (10%, 20%, and 30%) across low-, medium-, and large-data kinase groups. The selection of augmentation ratios (10%, 20%, and 30%) is guided by a practical trade-off commonly observed in imbalanced learning: balancing the expansion of minority class support against the risk of distributional drift introduced by synthetic samples. In particular, a minimum level of augmentation is required to meaningfully enrich the minority class representation, while excessive augmentation may reduce the influence of real samples and introduce bias in the learned decision boundary. The chosen range reflects this balance and is further validated empirically in our experiments. The corresponding results are summarized in Table 6, together with the quality scores of the generated samples.

Table 6.

Experimental results on the generation ratio of diffusion models across different data regimes.

Method	AUC	PR	F1	ACC	FPR	Quality
Low-data regime
No Gen	0.6996	0.7312	0.6901	0.5600	0.8600	–
10% Ratio	0.7856	0.8090	0.7213	0.6600	0.5600	0.2842
20% Ratio	0.7440	0.7122	0.7101	0.6000	0.7800	0.2897
30% Ratio	0.7840	0.7856	0.7143	0.6400	0.6200	0.3008
Medium-data regime
No Gen	0.8994	0.9052	0.8078	0.7924	0.2881	–
10% Ratio	0.9017	0.9155	0.8230	0.8178	0.2119	0.6151
20% Ratio	0.9018	0.9160	0.8261	0.8305	0.1441	0.6269
30% Ratio	0.8950	0.9071	0.8077	0.8305	0.0508	0.6198
Large-data regime
No Gen	0.8527	0.8661	0.7752	0.7810	0.1928	–
10% Ratio	0.8593	0.8697	0.7767	0.7525	0.3562	0.7679
20% Ratio	0.8757	0.8847	0.7950	0.7868	0.2533	0.7745
30% Ratio	0.8594	0.8683	0.7843	0.7761	0.2614	0.7748

Best results are highlighted in bold.

For low-data kinase groups, introducing diffusion-generated samples leads to a substantial performance improvement compared with the non-augmented baseline. In particular, a 10% augmentation ratio increases AUC from 0.6996 to 0.7856, representing the largest gain among all tested ratios. Increasing the augmentation ratio beyond this level does not yield further improvements; at a 20% ratio, AUC decreases to 0.744, despite a slight increase in the quality score of generated samples. This indicates that, under extreme data scarcity, moderate augmentation is sufficient to provide complementary information, whereas excessive synthetic samples may dilute the limited real-data distribution.

For medium-data kinase groups, diffusion-based augmentation consistently improves predictive performance across all evaluation metrics. A 20% augmentation ratio achieves the best overall results, increasing AUC from 0.8994 to 0.9018 and reducing the FPR from 0.2881 to 0.1441. Compared with lower ratios, this setting offers a favorable balance between sample diversity and distributional consistency, leading to simultaneous improvements in AUC, PR-AUC, and F1 score.

For large-data kinase groups, the generated samples exhibit the highest quality scores (0.77). However, the marginal performance gains become less pronounced. While a 20% augmentation ratio still yields a modest AUC improvement (from 0.8527 to 0.8757), higher ratios do not provide additional benefits and may introduce unnecessary computational overhead. These results suggest that when sufficient real training data are available, diffusion-based augmentation should be applied conservatively.

Overall, the results in Table 6 demonstrate that the effectiveness of diffusion-based augmentation depends jointly on the augmentation ratio and the scale of the original dataset, motivating adaptive rather than fixed generation strategies.

Cross-model generalization of the conditional diffusion module

To examine whether the proposed conditional diffusion module is tightly coupled to the KS-DiT architecture, we further evaluate its generalization capability by integrating the diffusion-based generator into the Phosformer framework. Performance comparisons under different augmentation ratios are reported in Table 7.

Table 7.

Experimental results on the generalization capabilities of generative modules (integrated with Phosformer backbone).

Method	AUC	PR	F1
Low-data regime
No Gen	0.6058	0.5984	0.6687
10% Ratio	0.6224	0.6077	0.6745
20% Ratio	0.6257	0.6212	0.6727
30% Ratio	0.6252	0.6158	0.6789
Medium-data regime
No Gen	0.8064	0.7656	0.7871
10% Ratio	0.8598	0.8627	0.7901
20% Ratio	0.8676	0.8779	0.7975
30% Ratio	0.8691	0.8813	0.8077
Large-data regime
No Gen	0.9102	0.9014	0.8591
10% Ratio	0.9065	0.8947	0.8583
20% Ratio	0.9099	0.8928	0.8626
30% Ratio	0.9057	0.8687	0.8599

In low-data kinase groups, the introduction of diffusion-generated samples consistently improves predictive performance. At a 20% augmentation ratio, AUC increases from 0.6058 to 0.6257, and PR-AUC improves from 0.5984 to 0.6212. These gains indicate that the generated samples provide informative supervisory signals even when transferred to a different classifier architecture.

In medium-data kinase groups, the performance improvements become more pronounced. AUC increases from 0.8064 to 0.8691 as the augmentation ratio increases, accompanied by a corresponding improvement in F1 score from 0.7871 to 0.8077. These results suggest that diffusion-based generation effectively enhances sample diversity in regimes where real data are sufficient to constrain the generative distribution while still benefiting from additional variability.

For large-data kinase groups, the inclusion of generated samples does not lead to significant performance improvements, but importantly, it does not degrade performance either. Across all augmentation ratios, AUC remains stable at 0.91, indicating that the diffusion module maintains robustness when applied to data-abundant settings.

Taken together, the results in Table 7 demonstrate that the proposed conditional diffusion module functions as a model-agnostic data augmentation component, capable of improving performance in data-limited scenarios while preserving stability when integrated into alternative backbone architectures.

Discussion

As demonstrated in section Experiments and result, the KSDiffusion framework consistently improves kinase-specific phosphorylation site prediction, especially under data-limited conditions. Beyond merely reporting performance gains, we must explore the underlying reasons for these behaviors. Furthermore, it is crucial to identify when diffusion-based augmentation is most effective and to consider the broader implications of applying generative models in biological sequence analysis. Therefore, this section synthesizes our empirical findings to discuss these aspects, rather than introducing new experimental evidence.

Representation-level interpretation and structural validation

Representation-level interpretation

KSDiffusion operates entirely in the continuous ESM-2 embedding space, where synthetic samples are generated as latent representations rather than discrete sequences. From a biological and modeling perspective, a diffusion-generated “positive sample” does not correspond to a single realizable peptide sequence. Instead, it represents a distribution-level embedding that captures shared biochemical and contextual characteristics of kinase-specific substrates. Because ESM-2 embeddings encode high-order dependencies, including evolutionary constraints and physicochemical properties, each generated embedding can be viewed as a plausible point within the latent substrate space associated with a given kinase. In this sense, the diffusion model approximates a functional region of substrate representations, rather than generating discrete molecular instances.

This design reflects the primary objective of KSDiffusion: to enhance discriminative learning under data scarcity by enriching the representation space with kinase-consistent samples. Mapping these embeddings back to explicit peptide sequences is non-trivial and is not required for the downstream prediction task, representing a trade-off between representation flexibility and sequence-level interpretability.

Structural validation in embedding space

To further examine the structural properties and biological plausibility of the generated embeddings, we analyze their organization in the learned representation space from both macroscopic and microscopic perspectives (Fig. 5).

Figure 5.

Visual validation of diffusion-generated synthetic embeddings in the continuous ESM-2 latent space across different data availability regimes. (a–c) t-SNE visualizations (following initial PCA dimensionality reduction) demonstrating the macroscopic distribution manifold. The generated synthetic positive embeddings (star markers) accurately anchor onto the region of experimentally validated real positive substrates (circular markers) while maintaining distinct separation from the real negative background. (d–f) -Nearest Neighbor (NN) neighborhood heatmaps quantifying the microscopic neighborhood composition. With a fixed neighborhood size of , synthetic data predominantly share local neighborhoods with real positive samples, showing near-zero density overlap with real negative samples. This quantitative topological analysis rigorously validates that the sequence-free generative augmentation produces biologically plausible and distinct substrate representations rather than statistical noise.

Macroscopically, PCA-accelerated t-SNE visualizations (Fig. 5a–c) reveal distinct distributional patterns across different data regimes. In the medium- and large-data regimes, synthetic positive embeddings closely align with the distribution manifold of experimentally validated kinase substrates, while remaining clearly separated from real negative samples. This indicates that the diffusion model successfully captures the global structure of kinase-specific substrate representations. In contrast, under the low-data regime, the separation among real positives, real negatives, and synthetic samples becomes less pronounced, with noticeable overlap in the embedding space. This reflects the inherent difficulty of accurately estimating the underlying data manifold when only limited annotated samples are available.

Microscopically, we further quantify local topological relationships using NN neighborhood enrichment analysis with a fixed neighborhood size of (Fig. 5d–f). The resulting heatmaps are structured analogously to a mixing matrix: each row corresponds to the anchor sample type (real positive, real negative, or synthetic), while each column represents the composition of its local neighborhood. Each entry reports both the absolute neighbor count and a normalized enrichment ratio relative to random expectation.

The results show that synthetic samples are consistently enriched with real positive neighbors ( higher than random expectation), while their association with real negative samples remains substantially lower. This demonstrates that the generated embeddings are not randomly distributed but are instead preferentially localized within biologically meaningful regions corresponding to true kinase substrates.

Notably, this alignment is not fully symmetric: while synthetic samples are strongly associated with real positive samples, which contain relatively fewer synthetic neighbors. This asymmetry arises from both sample size imbalance and the role of diffusion in expanding the support of the positive distribution rather than replicating existing samples. As a result, the generated embeddings provide broader coverage of the latent space, contributing diversity while preserving biological consistency.

In the low-data regime, both visualization and neighborhood analysis indicate weaker global separability. This can be attributed to the limited number of annotated positive samples, which leads to an incomplete estimation of the underlying data manifold. Nevertheless, the NN analysis still reveals a meaningful local structure: synthetic samples remain more strongly associated with real positive samples than with real negative samples. These observations highlight the intrinsic challenges of learning under extreme data scarcity (as discussed in the Impact of core components in low-data regimes section), where both real and generated samples are constrained by the limited and noisy estimation of the underlying distribution. This also explains why generative augmentation yields the most substantial performance gains in low-data regimes, despite reduced global separability in the embedding space.

Limitations and scope

Operating purely in embedding space imposes important limitations. In particular, the model does not directly generate peptide sequences and therefore cannot support sequence-level validation or experimental design. Consequently, the utility of KSDiffusion is primarily confined to improving predictive inference, rather than enabling de novo substrate design. Future work may explore hybrid approaches that integrate embedding-level generation with sequence reconstruction to enhance biological interpretability and practical applicability. In addition, while the current study focuses on kinase-specific phosphorylation, extending this embedding-level generative framework to other PTMs remains an important direction for future investigation.

Why diffusion-based augmentation is most effective under data scarcity

Our empirical results reveal a clear trend: diffusion-based augmentation substantially improves performance in low- and medium-data regimes. However, this marginal benefit naturally decreases as real data becomes abundant. This behavior stems from the interaction between data diversity, distributional coverage, and model capacity.

In low-data settings, kinase-specific phosphorylation patterns are sparsely sampled. This sparsity leads to incomplete coverage of the underlying sequence-function space. The conditional diffusion model alleviates this limitation by generating synthetic samples that interpolate between observed phosphorylation sites. Crucially, these samples remain constrained by kinase-specific conditioning. As a result, the classifier learns a smoother, more continuous decision boundary, which significantly improves generalization. However, if the proportion of generated samples becomes too large, the limited real data can no longer sufficiently anchor the training distribution. This destabilization explains the performance degradation observed at higher augmentation ratios.

In medium-data regimes, abundant real samples provide stronger distributional constraints. Here, diffusion-generated samples act primarily as diversity enhancers rather than distributional substitutes. Under these conditions, moderate augmentation ratios enrich the training set with biologically plausible variations. They preserve the core characteristics of the original data, leading to stable and consistent improvements across evaluation metrics.

By contrast, in large-data regimes, the original dataset already offers broad and representative coverage of kinase-specific phosphorylation patterns. Consequently, the additional information provided by synthetic samples becomes redundant. The role of diffusion-based augmentation thereby shifts from performance enhancement to regularization. This shift explains why performance gains become marginal yet remain stable, avoiding overfitting or degradation.

Architectural independence and robustness of the conditional diffusion module

Our experiments also reveal the strong architectural independence of the proposed conditional diffusion module. As demonstrated in section Cross-model generalization of the conditional diffusion module, diffusion-generated samples consistently improve or preserve predictive performance, even when integrated into alternative backbone architectures.

This behavior suggests that the diffusion module primarily captures data-level regularities rather than model-specific inductive biases. Operating on protein sequence embeddings and conditioning on kinase identity allows the generator to learn transferable representations. Various downstream classifiers can then leverage these captured kinase–substrate relationships. Consequently, the benefits of diffusion-based augmentation extend beyond the KS-DiT architecture to other sequence modeling frameworks.

Furthermore, the absence of performance degradation in data-rich regimes highlights the approach’s robustness. Even when synthetic samples contribute marginally, the diffusion module introduces no instability. This indicates it can serve safely as a plug-and-play augmentation component without extensive hyperparameter tuning.

Collectively, these findings suggest that the conditional diffusion module serves as an effective augmentation mechanism for tasks characterized by heterogeneous data availability. Its architectural independence allows it to be readily integrated into different prediction pipelines, particularly in scenarios where data scarcity varies across functional groups.

Implications and practical guidelines

Our analyses yield several practical guidelines. First, diffusion-based augmentation is most beneficial under data scarcity, where limited training samples constrain model generalization. In these scenarios, moderate augmentation ratios maximize performance gains while avoiding distributional bias.

Second, the conditional diffusion module readily integrates into existing kinase–substrate prediction pipelines without architectural modifications. Serving as a flexible augmentation component rather than a task-specific redesign, this flexibility lowers the barrier to adoption and facilitates application across diverse predictive frameworks.

Finally, our findings highlight that generative augmentation is a data-regime-aware strategy, not a universally optimal solution. Although diffusion models powerfully enrich sparse biological datasets, their contribution remains inherently bounded by the quality and quantity of available empirical data. Recognizing these boundaries is essential for the responsible and effective use of generative models in computational biology.

Conclusion

This study addresses the persistent challenge of data scarcity in kinase-specific phosphorylation site prediction by proposing KSDiffusion, a unified framework that integrates protein language representations with conditional diffusion-based data augmentation. By leveraging contextual sequence embeddings from ESM-2 and a conditional generative mechanism, the proposed approach effectively enhances model learning in regimes where kinase-specific annotations are limited and highly imbalanced.

Comprehensive experiments across kinase groups with varying data scales demonstrate that KSDiffusion consistently achieves competitive or superior performance compared with representative baseline methods. The performance gains are most pronounced in low-data scenarios, where diffusion-based augmentation substantially improves generalization, while remaining stable and non-degradative in data-rich regimes. These results highlight a clear data-scale-dependent behavior, indicating that conditional diffusion serves as a targeted and effective strategy for alleviating sample scarcity rather than a uniform substitute for large empirical datasets.

While the current framework is developed for kinase–substrate relationship prediction, the architectural independence of the generative module enables it to be naturally extended as a plug-and-play augmentation component for other data-scarce biological prediction tasks. Future work will focus on exploring its application to diverse PTMs, such as ubiquitination and methylation, as well as incorporating structural and dynamic interaction information to further enhance biological fidelity.

Key Points

• Problem-driven formulation: We investigate kinase-specific prediction under realistic data-limited and highly imbalanced conditions.

• Task-aware framework: We propose KSDiffusion, integrating protein language models with conditional diffusion to generate kinase-consistent synthetic samples.

• Effective integration: KSDiffusion improves generalization for data-scarce kinases while maintaining competitive performance on data-rich groups.

Conflicts of interest

The authors declare no competing interests.

Funding

This work was supported by the National Natural Science Foundation of China (grant number 62272192).

Data availability

The source code and data are available at https://github.com/aaalichking/ksdiffusion.