1 Introduction
Food-derived bioactive peptides (FBPs) are functional segments of natural food proteins. They have gained significant attention in functional foods and clinical nutrition for their strong biocompatibility and diverse regulatory roles [
1]. FBPs show potential for health intervention in six areas: ACE inhibition, antioxidant action, antibacterial properties, anti-glycemic effects, immune regulation, and anti-cancer properties [
2]. However, these fragments are typically hidden within the complex protein matrix and are released only through gastrointestinal digestion or enzymatic degradation
in vitro [
3]. Traditionally, FBP discovery uses a linear approach: enzymatic digestion, LC–MS/MS identification, and
in vitro activity screening [
4]. Given the vast sequence space of 20 amino acids (for a decapeptide, about 10
13 possible sequences), this method faces efficiency and cost bottlenecks. Computational methods are therefore urgently needed [
5].
In recent years, artificial intelligence (AI) and machine learning (ML) have become widely integrated into various branches of materials and engineering sciences as a data-driven, general-purpose methodological paradigm. For instance, discriminative and generative models are employed in the rational design of functional molecules, such as corrosion inhibitors, to predict molecular performance and generate candidate structures tailored to specific target conditions [
10]. In the fields of material aging and structural health monitoring, deep learning enhances the accuracy and efficiency of corrosion and crack detection [
11]. These interdisciplinary applications demonstrate that AI/ML is capable of high-precision prediction and identification of target properties in complex systems, as well as the generation of specific structures within vast candidate spaces. These two capabilities provide valuable insights for two core tasks in food peptide research: bioactivity prediction and sequence design.
In the field of food peptides, the focus of this paper, the integration of artificial intelligence and machine learning technologies has introduced new methodological dimensions to research [
6]. These technologies have expanded from activity prediction to food safety assessments, including toxicity and allergenicity evaluations [
12]. The capabilities of sequence representation and activity prediction have improved from early quantitative structure–activity relationship (QSAR) models that relied on manual feature extraction to automatic feature learning via deep learning architectures and, most recently, protein language models based on large-scale unsupervised pre-training [
7,
8,
13]. However, an objective analysis of the current state of research reveals a significant gap in methodology between statistical prediction at the sequence level and mechanistic understanding at the structural level. Most studies are in the preliminary stage of "computational prediction followed by
in vitro validation," and the gap between computational predictions and
in vivo validation of physiological efficacy significantly hinders the practical application of AI-designed peptides [
9,
14].
Building upon this context, this paper systematically reviewed 467 relevant papers published in the field of food peptides from 2021 to 2026. After screening for relevance, 247 highly relevant studies were included as the main body of the review, consisting of 229 research papers and 18 review papers. Fig. 1 presents the literature search and screening process for this article. It follows the PRISMA guidelines. The Fig. also shows the results of the unsupervised topic analysis, which is based on TF-IDF weighting and hierarchical clustering. The statistics reveal that sequence–activity prediction accounts for 86.6% of the research. Virtual enzymatic digestion prediction makes up 57.1%. Work on peptide structure design only accounts for 2.4%. This uneven distribution is the reason for this article’s focus on examining the task boundaries and evaluation criteria. These findings suggest that although AI has been initially applied in the screening of food-derived bioactive peptides, there is still a lack of systematic discussion on methodological boundaries, evaluation paradigms, and the differences in their application in the food industry across different tasks [
14].
Therefore, this article will focus on three aspects. Firstly, it will distinguish the definitions and technical boundaries of three tasks: "virtual enzymatic prediction", "sequence–activity prediction", and "peptide structure design". Building on this, it will then compare the applicability and transformation constraints of the two AI-assisted design paths—"structure-guided directed release" and "generative
de novo design"—in the food scenario. Finally, the article will analyze existing problems in the current model evaluation, such as sequence similarity leakage, inconsistent data division, and deviation in indicator interpretation. Most existing AI reviews in the field of food peptides are organized by either functional category or algorithmic architecture. Reviews organized by functional category provide individual introductions to various bioactive peptides, such as antioxidants and blood pressure-lowering peptides, while reviews organized by algorithmic architecture focus on summarizing the application progress and performance of various models [
15,
16]. This paper does not aim to provide such a list. Rather, it centers on a comprehensive examination of the three main themes mentioned above: task boundaries, evaluation robustness, and differences in design approaches.
2 Data Resources, Sequence Representations, and AI Model Architectures
The computational study of food-derived bioactive peptides is based on three interdependent technical levels: high-quality data resources for training and evaluating models, characterization methods that convert amino acid sequences into features usable by computers, and model architectures suited for diverse prediction tasks [
5]. To facilitate understanding, this chapter is structured around these three levels. It discusses the suitability of various tools in the context of food peptide research, examines known systematic limitations, and evaluates how these limitations actually impact subsequent modeling practices.
2.1 Databases for FBPs
High-quality datasets are a prerequisite for training and evaluating AI models [
17]. Most current research relies on publicly available bioactive peptide databases [
5]. For instance, BIOPEP-UWM continuously collects experimentally verified multifunctional active peptides from food sources and integrates virtual enzymatic digestion tools, making it the most comprehensive platform in this field [
19]. By contrast, AHTPDB focuses on the systematic organization of antihypertensive peptides [
20], while DFBP collects various food-derived bioactive peptides, including those with antihypertensive and antioxidant properties, thereby providing important data for food peptideomics [
21]. When considering antimicrobial peptides (AMPs), DRAMP [
23] and APD [
25] have constructed the core data infrastructure for this subfield, with a large-scale collection. However, their sequences mainly originate from the defense systems of animals and plants and natural products of microorganisms, and lack in-depth annotations for specific food matrices. Additionally, special databases such as SATPdb supplement data for specific active categories [
26]. Collectively, these resources provide basic data for the computational work in the research of food-derived bioactive peptides. Table 1 summarizes the focus areas, coverage, food source annotations, and cross-database overlap of the major databases mentioned above. Significant heterogeneity in coverage is evident, as is the fact that general-purpose databases generally lack food matrix annotations and sequence overlap between databases is common.
However, from the perspective of machine learning, the existing data resources have three systematic flaws. Firstly, sequence redundancy across databases can lead to a biased evaluation. Since the inclusion criteria of different databases overlap significantly, a random split after merging multiple databases will randomly assign samples to the training and test sets without considering the pairwise similarity of sequences. This can result in sequences that are highly homologous to, or even nearly identical to, training samples ending up in the test set. In this scenario, the test set no longer constitutes a truly independent evaluation set. Instead, the reported performance largely reflects the model’s “memory” of similar sequences rather than its ability to generalise to entirely new ones—a phenomenon known as sequence similarity leakage [
27,
28]. In their AutoPeptideML study, Fernández-Díaz et al. [
29] showed that using clustering based on sequence homology instead of random partitioning led to a significant decline in model performance. This suggests that the performance metrics reported in current literature based on random partitioning are generally overestimated. Secondly, for continuous value prediction tasks, the heterogeneity of
in vitro determination conditions for half of the inhibitory concentrations (such as substrate concentration and enzyme source) is relatively large, which tends to cause label noise [
30]. In response, Du et al. [
31] introduced CleanLab for label cleaning in the pLM4ACE study, thereby providing a feasible strategy to alleviate this issue. Finally, compared to peptides from drug or pathogen sources, the representativeness of samples from food protein sources in the database remains insufficient. As a result, models trained on general libraries face certain distribution deviation risks in the food peptide screening scenario [
5]. Taken together, these problems affect the generalization ability of the model and underscore the need for greater rigor in data cleaning and model evaluation stages.
2.2 Peptide sequence representation methods
Converting the text sequence composed of amino acids into a numerical matrix suitable for computer processing is the first step in building an artificial intelligence prediction model [
32]. Methodologically, sequence characterization methods in food peptides can be divided into three generations, distinguished by the depth of sequence information extraction and dependency on training data size [
15]. The first generation comprises physicochemical property descriptors, such as AAindex, z-scales, and hydrophobicity indicators, which rely on expert knowledge to numerically encode amino acids [
33]. These methods are highly interpretable and remain competitive under the small sample conditions common in food peptide research (typically < 500 sequences), continuing to serve as baseline model features. However, their main limitation is that manually preset features struggle to capture context dependencies and deep evolutionary information between residues [
29]. In contrast, the second generation encompasses statistical encoding methods, such as one-hot encoding, k-mer, and binary encoding, which forgo prior knowledge and offer low computational costs, granting them engineering advantages in early large-scale sequence processing [
34,
35]. Nonetheless, these encodings generate sparse result matrices and show insensitivity to physicochemical similarities and positional dependencies, limiting their performance in capturing sequence semantics [
36]. The emergence of a third-generation, protein language model (pLMs) embeddings—using architectures like ESM-2 and ProtTrans—enables the extraction of evolutionary constraints from massive protein sequence data via self-supervised learning. In transfer learning scenarios with limited food peptide training data (e.g., adapter fine-tuning), pLMs show significant potential [
37]. However, since current pLM pre-training corpora mainly feature long-chain protein sequences, their embedding features for short food-source peptides of 2–10 amino acids lack systematic quantitative evaluation. Thus, it cannot be assumed by default that pLM embeddings provide performance advantages when applied directly to short peptide tasks.
In actual modeling, the choice among the three generations of representation methods should be based on the data scale and the nature of the task. For small sample classification tasks, physicochemical descriptors are often comparable to deep methods [
38]. By contrast, when data volume is sufficient, pLM embeddings offer a significant advantage [
39]. Additionally, statistical encoding has a certain application space in engineering-oriented research due to its simple implementation. Because the choice of representation method directly restricts the upper limit of the downstream model architecture [
40], further discussion on this will be provided in Section 2.3.
2.3 AI model architectures for FBPs research
Building on recent advances, depending on the type of computational tasks, the AI model architectures in the research of food-derived bioactive peptides can be classified into two major categories: discriminative models and generative models. The former is used for tasks such as activity prediction and candidate peptide screening, while the latter is designed for
de novo sequence design [
16].
In discriminative models, traditional machine learning models (such as RF, SVM and XGBoost) are still widely used as baselines due to their robustness to small sample sizes and strong interpretability [
41]. Deep learning models, on the other hand, have gradually evolved along the "local-to-global" feature modelling spectrum: CNNs excel at extracting local sequence patterns, while LSTMs and their variants are well-suited to modelling long-range dependencies. Both have mature applications in predicting food peptide activity [
42–
45]. However, Transformers and attention mechanisms theoretically have advantages in capturing global context, their direct applications are fewer in pharmaceutical peptide research due to the short sequences of food peptides and the scarcity of labelled data [
46]. In recent years, architectures based on protein language models (such as pLM4ACE, various ESM-2 adapter-tuned models and feature-driven ensemble methods like DeepBP) [
31] have used transfer learning to adapt general protein knowledge from pre-trained models for specific food peptide prediction tasks. These approaches offer significant advantages over deep models trained from scratch in scenarios with limited training data and represent the current state of the art in sequence–activity prediction [
47,
48].
Generative models, such as variational autoencoders (VAEs), diffusion models, and autoregressive language models, differ fundamentally from discriminative models. Instead of predicting the activity of existing sequences, generative models autonomously create new sequences with the desired target function within the sequence space [
49,
50]. Since applying generative models to
de novo design of food peptides introduces more complex constraints and evaluation paradigms, this topic will be discussed in detail in Section 4.4. Table 2 provides a side-by-side comparison of the characteristics of the various discriminative and generative architectures mentioned above in food peptide research. The table focuses on three aspects: core strengths, major limitations and applicable scenarios.
When evaluating the performance of the aforementioned architecture, it is important to introduce a methodological perspective. Inconsistencies in data partitioning methods, such as random versus clustering partitioning, and differences in hyperparameter tuning strategies in the food peptide literature, create a lack of reliable horizontal comparison among architectures. In light of this, the existing literature is insufficient to substantiate the assertion that "deep learning models consistently outperform traditional machine learning models in food peptide applications". Due to biases in the datasets and differences in evaluation strategies, it is unclear whether there is a difference in terms of generalisation performance between the two, and the performance advantage of complex architectures in scenarios with low data volumes may be overestimated [
29]. Therefore, the choice of model architecture should depend on the specific task and dataset size, not on the assumption that complex models are always better than simpler ones.
3 AI-Assisted Discovery of FBPs: Applications and Progress
Based on the technical foundation established in Chapter 2, this chapter systematically reviews the application progress of AI in the core workflow for the discovery of bioactive peptides from food sources. This workflow presents a natural sequence of logical steps experimentally, from modeling the enzymatic hydrolysis process of natural food proteins to the prediction of activity for different functional categories. These two steps are methodologically interdependent. The quality of the enzymatic modeling determines the reliability of the candidate peptide library [
51], while the evaluation criteria of the activity prediction model directly affect whether the screening results can guide subsequent experimental decisions[
5]. While reviewing the application progress of each step, this chapter also analyzes the systematic defects existing in the current evaluation practice to define the credibility boundary of sequence-level prediction results and provide a methodological premise for entering the structural-level analysis in Chapter 4.
3.1 Virtual enzymatic hydrolysis modeling and protease specificity prediction
Converting natural macromolecular food proteins into short, biologically active peptides is the prerequisite for discovering active peptides. It is also the first computational step in AI-assisted food peptide research [
52]. This step always focuses on the same core issue: How to accurately predict the cutting probability of specific protein sites by certain enzymes without exhaustive experiments.
Early virtual enzymatic cleavage tools, such as BIOPEP-UWM and ExPASy PeptideCutter, use recognition sequences of known proteases as static rules. They conduct sequential scans on the primary sequence of the substrate [
18,
53]. The limitations of this method do not stem from rule inaccuracy. Instead, these tools operate only on one-dimensional sequence information and cannot use spatial conformational data from the protein’s natural three-dimensional folded state [
54,
55]. Many cleavage sites that match rules at the sequence level are in hydrophobic cores or are obscured by other domains in the real protein structure. Proteases cannot physically access these sites. This core gap between one-dimensional sequence and three-dimensional spatial accessibility limits the migration and generalization capabilities of such tools when analyzing novel food proteins with complex folding [
56].
The introduction of machine learning methods has advanced the prediction of cleavage sites. The field has evolved from isolated sequence pattern matching to context-aware probabilistic prediction [
57,
58]. Evaluating the performance and generalisation capabilities of models across different protease families relies heavily on high-quality benchmark datasets and rigorous evaluation metrics. Currently, model training relies primarily on benchmark datasets constructed from public databases such as MEROPS [
59], as well as platforms such as ProsperousPlus [
60]. In terms of evaluation, the extreme data imbalance between cleavage sites (the minority class) and non-cleavage sites (the majority class) means that the area under the precision-recall curve (AUPRC) and the Matthews correlation coefficient (MCC) are more rigorous evaluation metrics that more accurately reflect model performance than traditional accuracy metrics [
61]. Existing research shows that models exhibit significant differences in generalisation capabilities across different protease families [
62]. For serine proteases or matrix metalloproteinases (MMPs), where data is abundant, deep learning models typically achieve high prediction accuracy. However, models face severe challenges in learning from small datasets for specific protease systems with relatively scarce data, such as plant or microbial enzymes unique to certain food processing applications. Scarce training data easily leads to model overfitting, severely limiting its ability to generalise across families, which has become a core bottleneck in transferring current virtual proteolysis prediction technologies to broader food systems [
63].
At the level of structure-aware modeling, Lu et al. [
64] encoded inter-residue energy terms calculated by Rosetta into the node and edge features of a graph convolutional network. They constructed a specific prediction task for the substrates of hepatitis C virus (HCV) and tobacco etch virus (TEV) proteases. This approach enabled the model to learn how spatially adjacent residues modulate the cleavage probability. It outperformed the traditional sequence statistical baseline under various encoding settings. SinhaRoy et al. [
65] further introduced the E(n)-equivariant graph neural network. This model represents protein structures with full-atom graphs and dynamically updates three-dimensional coordinates during message passing. It achieved a breakthrough over sequence baseline methods for predicting enzyme pH optimum values. The use of attention weights revealed the core contributions of ionizable residues and the vicinity of the active site to the prediction. Guo et al. [
58] combined hierarchical geometric subgraphs with protease interaction networks. They constructed a unified cleavage site prediction framework for 103 proteases, showing significantly higher coverage in many-to-one protease–substrate scenarios. Through
in vitro experiments, they verified three new Caspase-3 substrates and 21 cleavage sites. However, the aforementioned methods primarily rely on static structural snapshots for modelling purposes. Given the highly dynamic nature of enzyme–substrate interactions in physiological or food processing environments, predictions based solely on static structures may not fully capture the conformational dynamics involved in substrate recognition [
66]. In future, exploring the limited integration of molecular dynamics (MD) conformational ensembles into feature characterisation could provide a complementary approach to understanding dynamic recognition mechanisms [
67]. Nevertheless, it should be noted that such mechanistic hypotheses still face significant challenges in terms of theoretical feasibility, managing high computational costs and their ability to actually improve the accuracy of large-scale predictions. These issues require careful validation through systematic and rigorous experimental designs in future studies.
At the multi-enzyme combination optimization level, the introduction of large language models further expands the application boundaries. This shift moves from predicting cleavage sites of a single protease to the intelligent screening of enzyme combinations for specific active peptide release targets. Morena et al. [
68] proposed a computational workflow for tomato by-products. This approach integrates multi-label bioactivity prediction and multi-enzyme combination screening modules. It targets the optimal enzyme configuration for the release efficiency of five types of functional peptides, enabling end-to-end prediction from high-value utilization of food by-products to targeted release of functional peptides. Martin-Alonso et al. [
62] proposed CleaveNet, an end-to-end AI design pipeline for protease substrates. Its predictor achieved continuous value regression for the cleavage scores of 18 matrix metalloproteinases (MMP). The generator could create new sequences based on target cleavage spectrum conditions. However, both approaches still rely to varying degrees on high-throughput experimental data as training bases. This reliance means that actual model implementation heavily depends on the experimental loop. With the deep integration of multi-omics data (proteome, transcriptome, structural data) and large language models [
69], it is expected that in the future, a protease-substrate map across enzyme classes and food substrates will be constructed. Combining this with active learning strategies could minimize experimental costs.
At the research paradigm level, the deep integration of virtual enzymatic digestion and multi-peptideomics is becoming an important trend in this field [
70]. Simply relying on computational predictions for the release peptide sequences has a significant risk of false positives. However, systematically cross-verifying the prediction results with the actual LC–MS/MS identification data can not only evaluate the actual prediction accuracy of the model but also provide reliable candidate sequences confirmed by mass spectrometry for subsequent activity screening [
71].
Despite the continuous advancement of technology, there are still significant gaps in the mechanism. The existing models simplify the simulation of the actual gastrointestinal digestive environment and are relatively weak in modeling the enzymatic hydrolysis kinetic parameters under food processing conditions (such as heat treatment, high pressure). Moreover, the systematic quantitative comparison between the predicted release efficiency and the experimental measured release efficiency is still relatively rare in the literature. This makes it difficult to objectively evaluate the actual prediction accuracy of the model [
72]. These limitations not only exist in the enzymatic hydrolysis modeling stage, but also in the subsequent activity prediction tasks. The systematic defects in the model evaluation design level also restrict the credibility of the computational results.
3.2 Sequence–activity prediction across functional peptide categories
After obtaining the candidate peptide library, predicting activity for specific biological functions is one of the most intensive areas of AI application in food peptide research [
15]. Owing to variations in data volume, biological mechanism complexity, and label standardization across activity categories, prediction strategies are highly tailored, making it difficult to define a single methodological framework. The following sections discuss six representative categories of bioactive peptides [
16]. Table 3 summarises the representative AI models, dataset sizes, evaluation metrics and data splitting methods used in six categories of bioactive peptide prediction task. The vast majority of bioactivity prediction studies reviewed in this paper use random splitting, with very few employing clustering based on sequence homology. Furthermore, there is inconsistency across studies regarding dataset sizes and evaluation metrics. Therefore, the performance metrics listed in the table only reflect the results reported under their respective experimental conditions.
ACE inhibitory peptides are the category with the most abundant data accumulation. Furthermore, the model tasks have expanded from binary classification of activity prediction to continuous value regression of IC
50 [
31]. The method evolution is clear, it has progressed from early QSAR descriptors, through deep learning architectures such as CNN/LSTM, to the current pre-trained models based on ESM-2 fine-tuning, represented by pLM4ACE [
87]. The actual screening efficiency of such models in complex matrix environments must be carefully evaluated. Existing performance claims are based primarily on standardised laboratory datasets, and further validation is required to demonstrate their transferability to industrial applications. Additionally, it is worth noting that molecular docking is more suitable as a tool for verifying the structural mechanism after AI prediction in this task, rather than as a first-round large-scale screening method [
88].
The research on anti-glycemic peptides mainly focuses on molecular mechanisms such as dipeptidyl peptidase IV (DPP-IV) inhibition, α-glucosidase inhibition, and α-amylase inhibition [
89]. Of these, data on DPP-IV inhibitory peptides are relatively abundant. Deep learning models based on atomic-level structural features (such as StructuralDPPIV) have been developed to address the representation bottleneck of traditional sequence descriptors [
75]. In contrast, datasets for α-glucosidase and α-amylase inhibitory peptides remain limited in scale and less standardized in annotation, restricting the development of mechanism-specific models of these two mechanisms [
90].
Antioxidant peptides exert their effects through multiple coexisting mechanisms, including scavenging free radicals, chelating metal ions, and inhibiting enzymes [
91]. Because these mechanisms often occur together, it can be difficult to clearly label each peptide’s function, resulting in label noise. To address this challenge, researchers have applied multi-label and multi-task learning approaches. For example, AnOxPePred integrates different antioxidant mechanisms into a single deep learning model [
77]. Another representative model, AnOxPP uses BiLSTM together with interpretable amino acid descriptors to better explain how peptide sequence relates to activity [
78].
In terms of antimicrobial peptides, there are differences in the sequence space distribution between food-derived antimicrobial peptides and those derived from drugs or pathogens [
92]. As a result, when directly transferring models trained on general databases such as APD or DRAMP to the screening scenarios of food peptides, caution is necessary. In this context, transfer learning methods such as BERT, Mamba, and CNN-LSTM hybrid models have shown preliminary food peptide screening scenarios. However, despite these advances, the performance loss brought about by cross-domain transfer still requires systematic evaluation [
80].
Immunomodulatory peptides are involved in complex biological processes. These include cytokine secretion regulation, macrophage activation, and maintenance of intestinal immune homeostasis. However, the bioactive phenotype strongly depends on the specific experimental system and detection endpoint [
93]. This results in weak label consistency and difficulty in standardizing the definition of positive samples [
84]. As a consequence, intelligent prediction research in this category lags behind ACE inhibitory peptides and antioxidant peptides. This lag is seen both in quantity and method maturity. Most existing work remains at the coarse-grained binary classification level [
82]. To move forward, mechanism label subdivision combined with multi-task learning is a more feasible development direction. In addition, for data scenarios with limited sample size and heterogeneous endpoints, strategies like transfer learning and small-sample learning strategies may be more practical than simply increasing model complexity [
94].
The research on anti-cancer peptides focuses on effects such as inhibition of tumor cell proliferation, induction of apoptosis, and regulation of the tumor microenvironment [
3]. Nevertheless, current studies face issues such as small sample sizes, high heterogeneity of experimental endpoints, and strong reliance on cell lines, which hinder the full evaluation of the model’s generalization ability on independent external data [
85]. Methodologically, most existing works use traditional machine learning or lightweight deep learning models for binary classification predictions. In this regard, attempts based on cost-sensitive or ranking-based frameworks, such as ranking learning, have initially demonstrated more robust screening capabilities under small sample conditions [
95]. Furthermore, when considering food applications, the prediction of anti-cancer peptides also requires taking into account food safety, stability, and actual exposure doses, highlighting that it is inappropriate to simply apply evaluation logic from the field of drug peptides [
3].
Across the six categories of bioactivity prediction tasks described above, model performance varies significantly depending on the task. The prediction accuracy and methodological maturity for ACE-inhibiting and antioxidant peptides are markedly superior to those for immunomodulatory and anticancer peptides [
85]. Upon closer examination of the underlying logic, this performance gap essentially reflects differences in the complexity of the biological mechanisms involved and the degree to which the experimental endpoints are standardised. For instance, ACE-inhibiting peptides usually have a single, well-defined molecular target (e.g., the active site of the ACE enzyme). Their
in vitro enzymatic reaction systems are highly standardised and the physicochemical mechanisms of targeted binding are relatively clear. This makes it possible to generate large-scale, low-noise, continuous-value labels (such as IC
50) [
96]. In contrast, immunomodulatory and anticancer activities often involve multi-target, network-level dynamic regulation (e.g., cytokine secretion pathways and immune evasion mechanisms in the tumour microenvironment) [
97]. Their phenotypic activity depends heavily on specific live cell lines and three-dimensional microenvironments, resulting in labels that essentially encapsulate multidimensional biological effects [
95]. Current deep learning models based solely on one-dimensional amino acid sequences clearly struggle to capture the full structure–activity relationship when faced with such highly complex network mechanisms, and their limited generalisation ability is therefore understandable. It should be noted that the differences in predictive performance across the aforementioned categories are influenced not only by the complexity of biological mechanisms but also by several systemic factors at the data level: Class imbalance caused by a scarcity of positive samples, sequence redundancy resulting from undeduplicated data, label noise arising from multiple mechanisms and endpoints, and experimental heterogeneity across different experimental systems—all of which collectively undermine the model’s reliability and comparability for underrepresented categories. The specific impact of these factors on the credibility of reported performance will be discussed in detail in Section 5.1.
4 AI-Driven Structural Analysis and Rational Design of FBPs
There is an essential methodological gap between statistical prediction at the sequence level and mechanism understanding at the structural level [
100]. Bridging this gap is a prerequisite for enhancing the credibility of prediction results and is necessary for moving from passive screening to active design [
101]. To address this core objective, this chapter follows two progressive technical logics. First, Section 4.1 discusses how molecular docking and molecular dynamics simulations can be used as validation tools for transforming sequence-level prediction results into structural-level mechanism explanations, and the applicable boundaries in the field of food peptides research. Next, Section 4.2 systematically reviews the progress and adaptability limitations of AI-driven structural analysis methods at the levels of parent proteins and candidate short peptides. Based on this, Sections 4.3 and 4.4 respectively elaborate on the relevant research and limitations of the two technical routes—structure-guided directed release and generative
de novo design—in the context of food peptides.
4.1 Molecular docking and molecular dynamics simulations
In the AI-assisted process of discovering food peptides, molecular docking and molecular dynamics simulation play different levels of roles in mechanism verification. Specifically, molecular docking is used to answer the qualitative question of whether the candidate peptides can form a reliable binding interface with the target. In contrast, molecular dynamics simulation evaluates the dynamic stability of the complex under physiological conditions. Together, these methods constitute the key bridge for transforming sequence-level activity prediction results into structural-level mechanism explanations [
88].
The core value of molecular docking, which uses tools such as AutoDock Vina, HPEPDOCK and GalaxyPepDock, lies in qualitatively and semi-quantitatively validating whether candidate peptides can form credible hydrogen bond networks and hydrophobic interactions with the active pockets of target enzymes, such as ACE, DPP-IV and xanthine oxidase (XO) [
102]. However, the physical approximation flaws in current mainstream docking scoring functions mean they struggle to handle flexible short peptides with a large number of rotatable bonds. This results in limited quantitative correlation between docking scores and the IC
50 of the peptide–enzyme system. Firstly, scoring functions often struggle to accurately assess the conformational entropy penalty during peptide binding. Free food-derived short peptides exhibit a highly dynamic conformational ensemble in solution. Binding to a receptor requires freezing a large number of conformational degrees of freedom and this substantial loss of entropy is often underestimated by traditional docking algorithms [
104]. Secondly, implicit solvent models struggle to accurately describe the critical hydrogen-bond networks mediated by water molecules at the peptide–receptor binding interface [
105]. Thirdly, peptide binding is often accompanied by a significant induced-fit effect, yet most high-throughput docking tools treat the receptor as rigid by default or allow only very limited side-chain flexibility. This results in an inability to capture the true global energy-minimum binding conformation [
106]. In their study of bovine DPP-IV inhibitory peptides, Nongonierma et al. [
103] observed that the peptide segment with the optimal docking energy exhibited only moderate activity (IC
50 = 216 μmol/L), whereas potent inhibitory peptides frequently received low docking scores. Therefore, equating docking scores directly with activity strength is risky, and molecular dynamics simulations must be employed for validation.
Molecular dynamics (MD) simulations introduce the dimension of time based on docking, by simulating the conformational evolution of peptide-receptor complexes under physiological conditions, to evaluate the dynamic stability and dissociation tendency of the binding interface [
107–
108]. When evaluating the dynamics of a system, it is crucial to distinguish clearly between the applicability limits of conventional molecular dynamics (cMD) and enhanced sampling techniques. While cMD can provide detailed information on conformational evolution with high spatiotemporal resolution, significant sampling bottlenecks are encountered when simulating flexible short peptide systems in food. The high flexibility of peptides, coupled with the fact that receptor binding pocket conformational rearrangements (e.g., opening and closing motions) are typically accompanied by high energy barriers, means that cMD is highly prone to becoming trapped in local energy minima within the typical nanosecond to microsecond simulation timeframes. This results in insufficient conformational sampling, and overcoming these barriers would require prohibitively high computational costs [
109]. In contrast, enhanced sampling techniques such as Gaussian-accelerated molecular dynamics (GaMD), metadynamics or umbrella sampling can effectively reduce energy barriers by adding a bias potential to the true energy surface or advancing along a reaction coordinate [
110]. With comparable or even lower computational costs, these techniques can explore a broader conformational space, thereby improving the quality of sampling for slow conformational transition [
111]. This capability is particularly critical in the practical application of food peptide research. For instance, Li et al. [
112] performed a 1 μs GaMD simulation to overcome the timescale limitations of conventional MD and successfully revealed how the soybean-derived dipeptide IY achieves highly efficient inhibition by contracting the ACE active site and stabilising key hinge regions. Similarly, Liu et al. [
113] elucidated the complex inhibition mechanism of a silkworm pupa-derived peptide against DPP-IV, which disrupts the
α-helix structure, through a 500 ns GaMD simulation. Nevertheless, a review of the current literature reveals two prevalent methodological shortcomings. Firstly, the simulation times reported in many conventional cMD studies are insufficient to adequately sample complex peptide–protein systems [
109]. Secondly, the rationale for force field selection and reporting of key solvation parameters is generally inadequate, which directly affects the reproducibility of results [
114].
4.2 AI-driven peptide structure analysis and conformational modeling
The activity prediction at the sequence level essentially establishes statistical correlations. However, converting these prediction results into structural hypotheses that can guide experiments requires the introduction of three-dimensional analysis tools [
101]. In food peptide research, this requirement is manifested at two interrelated but technically distinct and unevenly challenging levels. First, there is the three-dimensional folding structure of the parent protein and the target enzyme. Second, there are the conformational characteristics of the candidate short peptides. Notably, recent breakthroughs in deep learning for structural biology have provided differentiated tool support for addressing these two levels.
In terms of the structural analysis of parent proteins and target enzymes, deep learning-based structure prediction tools such as AlphaFold2 and AlphaFold3 can now provide atomic coordinates with a high degree of confidence [
15,
117]. Tools such as AlphaFold2 and AlphaFold3 can generate reliable structural predictions for target enzymes with stable folding cores (e.g., ACE and DPP-IV). AlphaFold3 can also model peptide-target complexes, providing a solid basis for the structural analysis of parent proteins and their targets [
118]. In food protein research, this level of structure prediction has practical applications. For example, researchers use structures predicted by AlphaFold to conduct solvent accessible surface area (SASA) analysis. Identifying exposed regions in the parent protein that have both activity potential and spatial accessibility is the core computational premise of the structure-guided directed release strategy described in Section 4.3 [
5]. ESMFold generates structure predictions based on protein language model embeddings rather than multiple sequence alignments. ESMFold is about 60 times faster than AlphaFold2 [
8]. It has significant engineering advantages when doing batch structure analysis of large-scale food protein databases. It has also been used in the structural perception graph neural network modeling of antimicrobial peptides [
119].
However, when the prediction target shifts to free food-derived short peptides with lengths of only 2–15 amino acids, the adaptability of the existing tools becomes significantly inadequate. The main issue is not the accuracy of the tools themselves. Rather, it is the physical nature of short peptides. These peptides typically lack stable single conformations in solution and exist in a highly dynamic conformational ensemble. Their biological activity is determined largely by this conformational distribution, not by a single static structure [
120]. McDonald et al. [
121] conducted a systematic multi-peptide structure benchmark test on AlphaFold2. They found that the model performs better in predicting rigid ring peptides with constraints like disulfide bonds. For highly flexible linear short peptides, the prediction error rises significantly. Johansson-Åkhe et al. [
122] noted that when evaluating the prediction ability of deep learning models for the bound state of short peptides, both AlphaFold2 and protein language models usually output a single low-energy conformation. This makes it difficult to capture dynamic processes such as conformational selection or induced fit. This limitation is also reflected in the model’s self-reported confidence metrics, which suggest that the reliability of short-peptide structure predictions cannot be judged based solely on confidence levels. The AlphaFold series quantifies prediction reliability using pLDDT (Predicted Local Distance Test) and PAE (Predicted Alignment Error). However, the reliability of these metrics has primarily been validated for structures with stable folding cores [
118]. When applied directly to flexible short peptides of 2–15 amino acids, a low pLDDT value does not necessarily indicate a prediction error, it can also reflect the fact that the sequence does not inherently possess a single stable conformation in solution [
123]. In such cases, the model output does not represent a sample from its conformational Boltzmann distribution and cannot characterise true conformational diversity with a single structure [
124]. Conversely, a high pLDDT value does not guarantee that the obtained conformation is the functional state [
66]. Therefore, directly screening or ranking short peptide structural hypotheses based on pLDDT/PAE carries a systemic risk of misguidance. Furthermore, since flexible short peptides lack a unique reference structure, traditional accuracy metrics based on a single reference conformation are limited in their ability to characterise the quality of such predictions. This is also the root cause of the significantly amplified evaluation errors observed in the McDonald et al. benchmark for highly flexible linear short peptides [
125]. Therefore, when using single-conformation prediction results to explain the mechanisms of flexible short peptides, caution should be exercised: for linear short peptides lacking a stable folding core, the single structures output by tools such as AlphaFold, which are essentially oriented, should be regarded as representative snapshots within their conformational ensemble rather than the true functional conformation in solution [
124]. Therefore, when discussing the mechanisms of action of such short peptides, it is advisable to use qualifying terms such as "putative conformation" to avoid equating single-structure predictions directly with functional conformations.
In terms of methodology, when it comes to acquiring short-peptide conformational ensembles, structure prediction based on protein language models/deep learning and physics-based conformational sampling are two complementary, rather than mutually exclusive, approaches. They have fundamental differences in terms of their principles, computational costs and applicability boundaries. Predictive methods such as AlphaFold2 and ESMFold rely on co-evolutionary information or language model embeddings to rapidly generate a single high-scoring conformation at low computational cost. This makes them suitable for the large-scale, batch modelling of parent proteins and target enzymes [
126]. However, their training objectives mean they favour outputting a single most-likely folding state, which makes it difficult to characterise the conformational distribution of short peptides. Conversely, molecular dynamics and enhanced sampling methods (e.g., GaMD and meta-dynamics, see Section 4.1 for details) explicitly describe conformational evolution using physical force fields. While these methods can, in principle, approximate a true thermodynamic ensemble, they face inherent limitations such as high computational costs and insufficient sampling in systems with high energy barriers. Recently emerging generative deep learning methods attempt to combine the advantages of both approaches by training on physical simulation data to generate conformations with ensemble diversity at a speed approaching that of predictive methods. This partially replaces expensive physical sampling [
66]. In the context of food-derived short peptides, this fusion approach is highly significant. While strategies such as introducing noise perturbations or reducing the depth of multiple sequence alignments (MSAs) in AlphaFold can increase the sampling space to some extent, they still fail to reconstruct the true thermodynamic ensemble distribution of peptides in solution. Consequently, generative methods specifically designed for peptide conformational diversity have emerged. PepFlow uses the diffusion framework and hypernetwork architecture to directly sample the full atomic conformation space of peptide segments. It surpasses the prediction accuracy of AlphaFold2 and ESMFold at various peptide lengths on the test set and can effectively reproduce the conformational ensemble revealed by nuclear magnetic resonance (NMR) experiments [
127]. This methodological shift—from predicting a single static conformation to sampling conformational distributions—has direct significance for food peptide research. Using conformational ensembles instead of a single structure as input for molecular docking is expected to significantly improve the reliability of peptide-target binding mode predictions [
128].
4.3 Structure-guided targeted release of FBPs
Unlike generative
de novo design, this route yields product sequences entirely from enzymatic hydrolysis fragments of natural food proteins. Under most regulatory frameworks, it can be classified as a protein hydrolysate category. The commercial transformation path is relatively clear [
9,
129–
130]. The differences in regulatory classification and evidence level between the two routes are detailed in Section 4.5. This boundary defines the core design logic of this route. Based on the known three-dimensional structure of the food parent protein, we select proteinase types and enzymatic hydrolysis conditions rationally. This approach minimizes steric barriers and precisely enhances the release efficiency of specific active peptide fragments [
131]. Fig. 2 compares the processes, algorithm tools, and food industry constraints of the two AI-assisted design routes: the structure-guided directed release and the generative
de novo design.
Implementing this approach relies on a multi-step computational-experimental workflow [
135]. Centred on the "sequence–structure–function" axis, the workflow begins with the three-dimensional structure of the parent protein, which is predicted by AlphaFold. SASA accessibility analysis is then used to identify exposed regions that are spatially accessible to proteases. The candidate pool is progressively narrowed through cleavage site prediction [
60] and multi-constraint virtual screening, and the process is validated through directed proteolysis and omics experiments [
132]. This workflow’s core value lies in its use of spatial accessibility information at the structural level, which overcomes the limitations of traditional virtual proteolysis tools that rely solely on one-dimensional sequence rule [
131]. This paradigm has been validated across multiple food systems. An et al. [
133] utilised the Transformer-based Protein T5 model to predict cleavage sites in human casein, achieving the efficient discovery of bioactive peptides by combining multi-constraint virtual screening with peptidomics. Concurrently, Mao et al. [
134] applied this closed-loop workflow to the fermented tofu system. By combining high-throughput peptidomics with machine learning, they were able to characterise the hydrolysis products of soy protein comprehensively. Using molecular docking and virtual screening targeting T1R1/T1R3 receptors, they identified five novel flavour peptides from over 600 candidate sequences. These were then validated via sensory evaluation. These studies demonstrate that combining computational virtual screening with omics-based validation is an effective strategy for developing functional food peptides. Fig. 3 illustrates the complete workflow for the structure-guided discovery of bioactive food peptides.
However, scaling up this approach to industrial levels presents three significant implementation challenges [
137–
139]. Firstly, there is the issue of computational complexity and distortion of physical states. Parent proteins in natural food matrices (such as casein micelles or soy globulin) usually exist as supramolecular polymers or aggregates, which are often denatured by heat during processing. Idealised single-monomer structures predicted by tools such as AlphaFold struggle to reflect the actual aggregated state of the substrate in the enzymatic reaction vessel. Obtaining reliable structures for multimers or aggregates, whether using AlphaFold-Multimer or molecular dynamics simulations, incurs significantly higher computational costs [
121]. Secondly, there are dual limitations in cleavage prediction with regard to 3D structure and specificity. Although 3D structural analysis has been incorporated into the initial stage of the workflow, the underlying algorithms of commonly used cleavage prediction tools (such as ProsperousPlus) continue to rely heavily on one-dimensional sequence pattern recognition. This "dimension reduction" makes it difficult to dynamically simulate steric hindrance and induced fit during protease–substrate binding. This leads to a large number of sites that match the sequence but are spatially masked being misclassified as positive. Furthermore, existing models rely heavily on a small number of well-characterised proteases in the training data to describe protease–substrate specificity. Consequently, their ability to generalise specificity predictions for niche proteases, which are commonly used in the food industry but lack sufficient annotation, is significantly limited [
56,
58]. Thirdly, there is an engineering gap in multimodal data integration. Cross-validation is required between computationally derived theoretical cleavage maps and experimental, high-throughput LC–MS/MS proteomics data. However, the current lack of standardised bioinformatics pipelines to achieve automated, quantitative matching of these two heterogeneous data types means that "computational screening" and "experimental validation" are fragmented into two independent steps. This makes it difficult to form an efficient, closed-loop data system. Integrating structure prediction, dynamic cleavage modelling and peptidomics alignment into an automated end-to-end workflow is an important obstacle to transitioning this approach to industrial applications [
140].
4.4 Generative models for de novo design of FBPs
Unlike the targeted release route described in Section 4.3, the candidate peptides designed from scratch do not require continuous enzymatic digestion of fragments derived from natural food proteins. Instead, they are new sequences generated autonomously by the model in the sequence space [
141]. This route offers greater freedom in sequence design. However, it also brings higher preparation costs and regulatory demonstration requirements. As a result, there is a relatively higher industrialization threshold. In the food field, related research is currently still in its early stages. Most of the work remains at the level of preliminary
in vitro validation. There is a significant gap compared to the maturity of the fields of drug peptides and antimicrobial peptides [
142].
The core constraints faced by
de novo design in food applications stem from the need to optimize more than just activity. The generated target must also meet several food-specific conditions, such as gastrointestinal stability, sensory acceptability, low toxicity, and low allergenicity [
143]. These constraints are not usually mandatory in pharmaceutical peptides. As a result, the generation framework from the pharmaceutical field cannot be directly applied. Food constraints must instead be clearly included in the generation process goals [
144–
145].
There are currently three main types of generative framework available for the
de novo design of food peptides. These frameworks differ significantly in terms of how their underlying mathematical logic is integrated with food-specific constraints. The first is the Conditional Variational Autoencoder (CVAE), which works by compressing and mapping discrete peptide sequences into a continuous mathematical latent space. It then guides the generation of new sequences by inputting specific labels (such as "high ACE inhibitory activity") during decoding. Studies such as PepVAE have demonstrated the potential of this approach for producing antimicrobial peptides [
146]. However, CVAE faces severe "data-hungry" limitations in the food context. Designing food peptides often requires hard constraints to be satisfied across multiple dimensions simultaneously, such as high activity, non-toxicity and low allergenicity. This means that CVAE training data must possess extremely comprehensive multi-label annotations. However, given the current scale of food peptide databases and the prevalence of missing labels, generation frameworks that rely heavily on joint distribution data are highly susceptible to model collapse. A second approach uses classifier-guided diffusion models (such as discrete diffusion architectures like D3PM) [
147]. The principle of diffusion models is similar to "denoising sculpting". First, valid sequences are disrupted into completely random amino acid jumbles (adding noise), and then a neural network is trained to gradually undo these disruptions and ultimately "extract" meaningful new peptides from the random noise. In terms of integrating food-specific constraints, this architecture demonstrates great flexibility. It allows for the use of a "classifier guidance" mechanism, meaning there is no need to retrain a large generative model. Instead, independent predictors of bitterness and toxicity (such as ToxinPred) act as external "navigators", penalising sequences that deviate from food safety goals at each stage of denoising. This "plug-and-play" approach to embedding constraints is well suited to the highly fragmented landscape of predictive tools in the food industry. Strong conditional generation capabilities have already been demonstrated in image and protein design within continuous spaces. The primary limitation of this approach is that, when performing continuous gradient approximations for discrete sequences of fewer than 20 amino acids, rounding errors are mathematically likely to occur. This results in generated sequences that lack stability in their physicochemical properties [
148]. The third approach is the constrained autoregressive language model. Models such as ProtGPT2 are representative of this approach, which is similar to the "next-word prediction" feature in smartphone keyboards. Here, the probability of the next amino acid is predicted sequentially from left to right based on the generated amino acid prefix. This architecture effectively inherits the powerful sequence-based linguistic patterns of large language models [
149]. Compared to the first two types of architecture, this method naturally leverages the knowledge of pre-trained protein language models. However, the real-time introduction of food-specific constraints remains challenging. Gastrointestinal stability (resistance to proteolytic cleavage), for example, is often a global property determined by the three-dimensional folding of the entire polypeptide. When autoregressive models concatenate amino acids sequentially from left to right, they cannot predict the global structural trajectory of the sequence’s terminus. Consequently, it is extremely difficult to introduce such global hard constraints, which depend heavily on spatial conformation, in real time with precision during the decoding phase. Currently, the only viable approach is a compromise involving post-generation filtering, but this drastically reduces the effective hit rate of the candidate library [
150]. Table 4 summarises representative studies on generative peptide design in the pharmaceutical field, covering the core mechanisms and experimental results of the three aforementioned architectural categories. These cases demonstrate that generative methods have established a relatively comprehensive evidence chain in pharmaceutical applications, ranging from algorithm design to
in vivo validation. While the architectural choices in these studies can serve as a reference for the food sector, food-specific constraints still require explicit modelling within the generative objective function.
Overall, all three types of architectures remain in the stage of method exploration, and there is a lack of systematic experimental verification in the field of food peptides. The fundamental factors currently restricting the development of this approach are twofold. First, the insufficient scale of food-specific annotation data makes it difficult to support large-scale conditional generation training. Second, the lack of high-quality peptide-target complex structure databases, such as those found in the pharmaceutical field (e.g., ChEMBL or PDB), limits the feasibility of structure-guided generation [
156]. In comparison with the targeted release route discussed in Section 4.3, it is important to note that while
de novo design theoretically offers a larger sequence exploration space, in the food scenario, its practical transformation still faces more unresolved methodological challenges. Therefore,
de novo design is more suitable as a medium-term development direction rather than a near-term industrialization path.
4.5 Regulatory classification and level of evidence
When translating AI predictions into practical applications, it is important to understand the regulatory context of the two design approaches, as this determines the level of evidence required and the path to commercialisation. For the structure-guided targeted release approach, the product sequence is naturally derived from food matrices, so its regulatory status typically aligns with that of traditional protein hydrolysates or functional food ingredients. Under most regulatory frameworks, the product can enter the functional food market via the U.S. Food and Drug Administration (FDA)’s generally recognised as safe (GRAS) notification process or the European Food Safety Authority (EFSA)’s Novel Food procedure. This means that a pharmaceutical-grade submission as a new chemical entity is not required [
129–
130]. The focus of commercialisation under this approach lies in providing detailed enzymatic processing parameters, purity standards and batch consistency data. The level of evidence should focus on validating the product’s safety and efficacy based on human cohorts, emphasising its bioequivalence and the reproducibility of industrial production.
In contrast, the generative
de novo design route typically yields non-natural sequences. The innovative nature of these sequences means that their regulatory classification is closer to novel foods or high-value additives. In certain cases involving high activity, the substantiation requirements may be even more stringent. Consequently, regulatory authorities typically require more rigorous toxicological characterisation, allergen screening and dietary exposure assessments [
157]. For such products, it is advisable to establish a tiered evidence chain comprising "laboratory safety screening, organoid functional validation, and systematic safety evaluation". In short, the former translation strategy leans towards "process optimisation and efficacy enhancement within the existing regulatory framework", while the latter requires "systematic safety validation and regulatory filing starting from the sequence level". Researchers should select a suitable regulatory pathway based on their vision for the product at the initial design stage, rather than merely pursuing the peak activity predicted by models.
5 Challenges and Future Directions
The preceding chapters have outlined a methodological framework for AI-assisted food peptide research, covering technical foundations, application progress, structural validation, and rational design. This framework also makes clear that the field still faces several barriers to practical translation. Data quality and evaluation standards remain insufficient, limiting model credibility. The gap between prediction and validation continues to reduce the translational efficiency of computational results. In addition, the two design routes discussed above still face unresolved methodological challenges [
158]. This chapter will set aside excessive theoretical speculation and focus instead on the fundamental data deficiencies and practical industrial obstacles currently hindering the field’s development, and on that basis, propose a more prudent path forward.
5.1 Limitations of reported performance
Although current artificial intelligence models have achieved high performance metrics when predicting the bioactivity of food-derived bioactive peptides, claims regarding model performance in existing studies tend to be overly optimistic and their true generalisation capabilities should be viewed with caution. The performance metrics summarised in this section and Table 3 are largely derived from experimental designs involving random splitting, single datasets and a lack of independent external validation. Therefore, they should be understood as optimistic upper-bound estimates under the specific, controlled conditions of each study rather than as an indication of the performance that can be expected from the models in real-world food protein screening scenarios. Specifically, this systematic overestimation primarily stems from three overlapping factors.
First, evaluation bias caused by sequence similarity leakage. As discussed in Sections 2.1 and 3.2, highly homologous derived sequences are widespread in public databases. Yet most studies in the current literature still employ random splitting, which causes homologous sequences to span both the training and test sets. This means that the model’s high accuracy on the test set is largely due to "memorisation" of similar sequences in the training set, rather than the generalisation of structure-activity relationships. Multiple comparative studies have shown that, once strict evaluation is conducted using clustering based on sequence homology, model performance typically drops significantly. This confirms that the performance reported in existing studies is largely due to dataset bias.
Secondly, there is a systematic lack of independent external validation. Current studies generally rely on internal cross-validation using a single database or data from the same laboratory, while validation across different sources is relatively rare. However, cross-validation and independent external validation are not equivalent. The former cannot reveal a decline in generalisation when the model encounters new sequences under different processing conditions or from different biological sources, such as whey protein, fish collagen and soy protein. Due to the diversity of food matrices and the heterogeneity of extraction techniques, model performance often fluctuates significantly across different food protein sources. Consequently, existing metrics primarily reflect the degree of fit under specific, controlled conditions rather than true cross-source generalisation and industrial transferability.
Thirdly, there are the dual challenges of metric selection and label noise. Firstly, most bioactive peptide screening scenarios exhibit significant data imbalance, with far fewer positive samples than negative ones. If accuracy or ROC-AUC are used as the sole evaluation metrics, they may mask deficiencies in the model’s recognition of critical minority classes. Therefore, metrics that are more sensitive to imbalance, such as the area under the precision-recall curve (AUPRC) and the Matthews correlation coefficient (MCC), are preferable. On the other hand, for continuous regression tasks, differences in assay conditions, such as enzyme sources and substrate concentrations, across laboratories can introduce systematic shifts in labels such as IC50 by orders of magnitude when data is pooled from different sources, thereby undermining the reliability of regression models. In classification tasks, artificially setting fixed thresholds forces the continuous dose–response relationship into a binary format. This inevitably introduces significant label noise at the threshold boundaries, thereby inflating the false positive rate in virtual screening.
In summary, the combined effects of sequence similarity leakage, the absence of independent external validation and biases at the metric/label level constitute the primary threats to the credibility of current literature-based performance metrics. Before proceeding to the structural analysis and rational design of bioactive peptides, a unified evaluation standard must be established that encompasses rigorous deduplication, independent external validation and evaluation metrics adapted to imbalanced data. This is a prerequisite for driving credible development in this field.
5.2 Data infrastructure, evaluation standards, and interpretability challenges of AI models
The core obstacles restricting the reliable application of AI models in the field of food-derived bioactive peptides are sequentially derived from three interrelated aspects. The weakness of the data infrastructure determines the upper limit of the model’s performance, the systematic flaws in the evaluation design make it difficult for the reported performance to reflect the true generalization ability, and the limitations of the interpretability tools further impede the transformation of the model’s output into experimental decisions [
159]. Any form of systematic deficiency will hinder the process of AI tools moving from academic verification to practical application.
At the data foundation level, the field of food-derived bioactive peptides has not yet established a data infrastructure similar to ChEMBL in the drug discovery field, which features unified experimental condition annotations, auditability, and cross-database comparability [
160]. The high level of noise commonly encountered during cross-database merging, as well as the limited amount of data available for specific food matrices, are the underlying issues that restrict the maximum performance of the model.
In terms of interpretability, tools such as SHAP and attention weight analysis are generally limited to the post-hoc attribution of existing prediction results, and have yet to be effectively translated into forward-looking guidance for experimental design. Post-hoc attribution informs researchers of the residues that the model considers important, but it cannot answer the core experimental question of how to modify the sequence to enhance activity. Furthermore, SHAP scores and attention weights are sensitive to background datasets, model random seeds and hyperparameter settings. The same model may produce inconsistent residue importance rankings under different configurations. Therefore, when used to assist in sequence design, any uncertainty in the attribution results (e.g., stability of importance rankings across multiple runs or sets of background samples) should be reported rather than treating a single attribution as a definitive design rule. Future development of interpretable methods should shift towards providing actionable recommendations for optimising sequences, enabling model outputs to directly inform the rational modification of candidate peptides.
The practical value of artificial intelligence in the field of bioactive peptides of food origin can be enhanced not only through iterative improvements in algorithmic performance, but also through the scientific community’s commitment to reproducible research and transparent methodology. The inconsistencies in evaluation methods found in the current literature undermine the credibility of computational predictions. Therefore, to promote robust development in this field, future research should standardise these key elements in reporting guidelines. Firstly, the version and original source of the datasets used should be clearly disclosed. Redundancies arising from merging multiple databases should be avoided and clustering strategies based on sequence similarity rather than random partitioning should be explicitly defined to eliminate performance leakage caused by homologous sequences. Secondly, the selection of hyperparameters, the details of the network architecture, and the random seed settings used during model training must be fully documented to ensure reproducibility of results. Finally, as computational paradigms become increasingly complex, researchers should actively promote open-source code accessibility by providing standalone repositories containing detailed configuration environments and pre-processing scripts. This methodological transparency mechanism, characterised by "data traceability, strategy auditability, and code reproducibility", is a prerequisite for transitioning research on food-derived bioactive peptides from preliminary academic exploration to rigorous industrial reference standards.
These requirements for data source documentation, model version management and auditable traceability align with emerging lifecycle governance practices in other high-risk AI application domains, such as clinical diagnostic laboratories. The latter emphasise clarifying accountability through MLDevOps processes, human-supervised thresholds and enforceable audit and version control. As the AI-assisted design of bioactive food peptides moves towards industrialisation, the adoption of such governance frameworks and the establishment of model traceability and accountability mechanisms tailored to food applications will be critical steps in ensuring their reliable translation into practical use.
5.3 Challenges in bridging computational prediction and experimental validation
Based on the analysis of current literature in the field of food peptides, there is a clear discontinuity in the translational chain from computational prediction to experimental verification. A large number of studies stop at the stage of
in vitro enzyme inhibition or preliminary validation at the cellular level, and very few can further extend to animal experiments or even clinical nutritional intervention evidence in humans [
41,
166]. This discontinuity has several objective causes. The oral bioavailability of food peptides is constrained by multiple factors such as gastrointestinal degradation, epithelial transport efficiency, and systemic distribution. There is a structural gap between
in vitro activity data and
in vivo efficacy that is difficult to directly extrapolate [
167]. More importantly, if AI-predicted food-derived bioactive peptides are to enter the health claim evidence system, they need to meet the clear requirements of regulatory agencies for levels of evidence. This poses higher requirements for the current research paradigm that mainly relies on
in vitro validation [
157,
168–
169]. The specific regulatory classifications corresponding to the different design routes have been discussed in Section 4.5.
To narrow this gap, intestinal organoids and intestinal organ chip systems are being regarded as an intermediate validation platform between
in vitro calculations and
in vivo animal experiments [
170–
171]. Compared with the traditional Caco-2 monolayer cell model, these two platforms can more realistically reproduce the multicellular structure of intestinal epithelium. They also simulate dynamic fluid shear force and the mucus layer barrier. They have higher physiological relevance in evaluating the stability of polypeptides in the gastrointestinal tract and the efficiency of epithelial transport [
167,
172]. Currently, the systematic application of these platforms in the field of food peptides is still in its infancy. But their potential as transitional validation tools between
in vitro calculations and
in vivo animal experiments deserves close attention. Establishing a hierarchical validation path from computational prediction to organoid or organ chip verification, and then to animal experiments, is expected to increase the hit rate of candidate peptides entering subsequent
in vivo studies. This progression also helps control experimental costs [
173].
The key to overcoming methodological differences between computational modules is to integrate various prediction tools into a closed-loop workflow centred on the "sequence–structure–function" axis. The core logic involves replacing isolated prediction models with explicit constraint propagation. As outlined in Section 4.3, this workflow incorporates all the necessary technical components, including structural prediction, SASA accessibility analysis, cleavage site prediction, multi-constraint screening, and experimental validation. However, the current workflow relies primarily on unidirectional, manual connections between these components, which makes it difficult for prediction results to provide feedback constraints on upstream step. A true closed-loop system requires downstream experimental validation data (e.g., LC–MS/MS proteomics results) and multi-constraint scores to be fed back into upstream structural and cleavage model. This forms an iterative optimisation cycle of "prediction–validation–re-prediction", rather than the current linear paradigm of "broad-range screening–one-time validation".
5.4 Emerging research directions
The challenges highlighted in the previous two sections operate at different levels. Data and evaluation issues represent deficiencies at the infrastructure level, while the disconnect between prediction and validation constitutes a structural barrier at the level of the translation pathway. This section discusses two interrelated emerging approaches to address these two categories of issues: enhancing the adaptability of computational toolchains for food-related scenarios (which addresses deficiencies at the infrastructure level directly) and establishing a closed-loop design paradigm from start to finish (which bridges the gap between prediction and validation in the translation process). These two approaches are methodologically interdependent. Deep adaptation of the toolchain is a prerequisite for the closed-loop system to function, and the data accumulated through closed-loop feedback will continuously optimise the representational quality of the toolchain.
The adaptability of Direction One is improved throughout the entire process, from sequence representation and structural modelling to industrial production [
174]. At the representation stage, current protein language models (such as ESM-2 and ProtTrans) are trained on general protein databases. The unique sequence characteristics of food proteins, like the Pro-Hyp-Gly repeating unit in collagen and the proline-rich region in cereal proteins, are underrepresented. This underrepresentation leads to systematic bias when these models are transferred to food peptide tasks in zero-shot settings [
8,
175]. There are two complementary improvement paths: One is improving domain adaptability through secondary pre-training for food protein corpora, the other is introducing small-sample learning strategies based on meta-learning for data-poor categories, such as immunomodulatory peptides and anticancer peptides [
176]. At the modeling stage, most prediction models use only one-dimensional sequence information. However, bioactivity is fundamentally shaped by three-dimensional conformation and its spatial complementarity with the target [
177]. Multimodal joint modeling of sequence and structure has shown better generalization in antimicrobial peptides. Introducing structural information can ease excessive reliance on the training set, especially in scenarios with low sequence similarity [
178]. AlphaFold 3 enables joint prediction of protein-ligand complexes and offers new structural modeling possibilities for peptide-target complexes [
117]. However, as discussed in Section 4.2, conformational diversity in food-source short peptides remains unresolved. The advantage of multimodal integration depends largely on improving short peptide structure prediction. At the preparation stage, AI-assisted enzyme engineering can support both structure-guided directed release and
de novo sequence design. This approach forms a promising interdisciplinary link between computational design and industrial production [
179–
180]. Rationally modifying natural proteases with ProteinMPNN and other design algorithms may improve specificity for certain specific cleavage sites. [
181]. Perhaps more importantly, the non-natural candidate sequences generated by
de novo design are theoretically expected to be produced by customised proteases through enzymatic methods, thus providing a potential route for the large-scale production of designed products that is worth exploring. However, their practical feasibility still needs to be verified through experimentation [
177]. Advancements at the levels of representation, modeling, and preparation together form a complete technical route for deeper adaptation of AI methods to food peptide research.
A second approach is an end-to-end, closed-loop design that aims to integrate the three stages of computational generation,
in vitro validation and industrial application. A key challenge in the current state of AI research on food peptides is the absence of bidirectional feedback between these stages, most studies only consider single
in vitro activity assays. One promising approach is to incorporate gastrointestinal stability prediction, bioavailability estimation and manufacturing cost constraints as explicit, multi-objective optimisation criteria into the generative model’s objective function, rather than applying them as a post-hoc filtering layer once generation is complete [
143,
145]. In the pharmaceutical field, this logic has been initially practiced in the research of antimicrobial peptides [
183–
184]. Nevertheless, the additional constraints in the food scenario, such as sensory acceptability, food matrix stability and regulatory classification, make it inappropriate to directly transfer the closed-loop framework from the pharmaceutical field. The end-to-end closed-loop design for food peptides requires the establishment of a food-scenario-specific multi-constraint collaborative optimization framework at the methodological level [
168,
185], which is the core challenge that both design routes are currently facing for practical translation and a methodological direction that can be systematically explored in the future. It should be emphasised that the practical implementation of the above approaches is limited by three obstacles: The computational costs associated with high-precision structure prediction and dynamic simulation (see Sections 4.1 and 4.3 for details). The regulatory acceptance and evidence requirements for health claims (see Section 4.5). Thirdly, there are scalability bottlenecks from computational screening to experimental validation (see Section 5.3). These obstacles cannot be overcome by a single algorithmic breakthrough, but rather require collaborative efforts across computational, experimental and industrial domains.
6 Conclusions
This paper focuses on the AI-assisted discovery and rational design of food-derived bioactive peptides, establishing a systematic analytical framework spanning technical foundations, application progress, structural validation and design approaches. Unlike existing reviews, which merely list individual functional categories or algorithms, this paper provides a comprehensive examination across three levels: defining task boundaries, critically analysing evaluation practices and assessing the two design approaches in food-related contexts. The central argument of the paper is that current AI implementation bottlenecks are not primarily caused by the algorithms themselves, but by two non-algorithmic issues: evaluation credibility and translation pathways.
At the technical foundation level, the existing data resources, sequence representation methods, and model architectures have initially formed a tool system. This system supports computational research on food peptides. However, sequence redundancy, label heterogeneity, and insufficient representativeness of food protein sources at the database level are the main data-related issues. These issues restrict the credibility of the model performance [
21,
161]. The selection of sequence representation methods should be based on data scale and task nature, rather than assuming that complex methods are necessarily superior to simpler ones [
38]. The tasks of virtual enzymatic hydrolysis prediction, sequence–activity prediction, and peptide structure design have essential differences. Their methodological positioning, data requirements, and evaluation criteria differ and should not be assessed using a single undifferentiated evaluation framework [
186].
At the level of application progress, AI has demonstrated practical potential in aspects such as virtual enzymatic hydrolysis modeling and peptide activity prediction [
187]. The leakage of sequence similarity caused by random data splitting, the lack of systematic independent external validation, and the limitations of post hoc attribution of interpretability tools collectively constitute the main threats to the reliability of performance metrics in current literature [
161]. These systematic weaknesses in evaluation affect the entire chain from model selection to industrial deployment, and are a methodological issue that needs to be prioritized in the current field. Adopting sequence-cluster-based evaluation protocols and establishing standardized benchmarks covering multiple food protein sources do not require new algorithmic breakthroughs and therefore represent a relatively feasible short-term path for improvement [
188].
At the level of rational design, structure-guided directed release and generative
de novo design differ fundamentally in product origin, design flexibility, regulatory pathway, and industrialization threshold. These differences require route-specific assessment. The main technical components required for the structure-guided approach have been developed: receptor structure prediction and protease cleavage site prediction tools based on deep learning are becoming more advanced, and the multi-constraint virtual screening framework has been put into practice [
68,
156]. However, as discussed in Section 4.3, these components are not yet standardised or reusable in automated workflows. Generative
de novo design is different. It provides a wider theoretical space for sequence exploration, but its translation in food peptide research is still limited by the shortage of food-specific annotated data [
189], the difficulty of explicitly encoding food-related constraints, and the costs associated with preparation and regulatory substantiation. For this reason, generative
de novo design is likely to remain a medium-term research direction rather than a near-term industrialization pathway [
141,
190].
Overall, at present, artificial intelligence is more suitable as an efficient auxiliary tool for candidate peptide discovery and prioritization in food peptide research. It is not yet capable of replacing systematic experimental validation. The structural gap between computational prediction and in vivo physiological efficacy remains the important obstacle restricting the transformation of this field. In the future, if we can continue to make progress in the development of dedicated food-protein-specific representation models, the deep adaptation of sequence–structure multimodal modeling, and the explicit embedding of food-specific constraints into the end-to-end closed-loop design framework, AI is expected to play a more stable and more interpretable substantive role in the discovery and rational design of food-derived bioactive peptides.
The Author(s) 2026. This article is published by Higher Education Press.
PDF
(5455KB)
