Predictive accuracy of peptide–target interaction models in drug discovery: a systematic review and meta-analysis

Peptides represent promising therapeutic agents due to their high specificity, biocompatibility, and capacity to modulate protein–protein interactions. However, the field faces critical challenges: inconsistent evaluation metrics, heterogeneous datasets, and poor reproducibility, which together undermine objective model comparison and benchmarking. To systematically review and quantitatively assess the predictive performance of machine learning (ML)-based peptide–target interaction (PTI) models, with emphasis on commonly reported metrics including area under the curve (AUC), concordance index (CI), and Precision. We conducted a systematic literature search across PubMed, arXiv, and Cochrane databases for studies published through 1 August 2025. Inclusion criteria required original ML-based peptide–target prediction models with quantitative performance metrics. Twenty-three studies met inclusion criteria. Fourteen reported AUC values (pooled estimate: 0.87, 95% Confidence Interval: 0.83–0.90), five reported Concordance Index values (0.90, 95% Confidence Interval: 0.89–0.91), and six reported precision (0.75, 95% Confidence Interval: 0.69–0.81). Top-performing models predominantly employed transformer or graph neural network architectures with structural input features. Critical limitations included inconsistent reporting practices, infrequent external validation, and limited data/code availability. The observation that the twenty-three studies present challenges to a pooled estimate is itself a consequence of the reporting practices we document, and is why we advance STRIDE. ML models demonstrate strong potential for PTI prediction, with leading approaches achieving robust classification and ranking performance. Nevertheless, progress is hindered by non-standardised evaluation metrics, limited transparency, and insufficient reproducibility. We introduce the STRIDE framework (encompassing Standardization, Transparency, Representativeness, Integration, Discovery, and Evidence) to establish rigorous evaluation standards, enhance methodological reproducibility, and support future evaluation of clinical applicability.

High specificity, favorable safety profiles, and biocompatibility distinguish peptides as promising therapeutic agents. Unlike small molecules, they often exhibit superior target selectivity and lower toxicity because their structures resemble endogenous proteins, and they can modulate protein-protein interactions that are difficult for traditional drugs to target (Fosgerau and Hoffmann, 2015). These advantages make peptides attractive candidates for a wide range of therapeutic areas, including oncology, endocrinology, and infectious diseases (Wan et al., 2024a; Torres et al., 2024). Despite these biological strengths, the discovery and optimisation of peptide-based drugs remain resource-intensive. Challenges such as poor oral bioavailability, rapid enzymatic degradation, and structural instability have historically limited clinical translation.

Recent advances in peptide engineering and computational drug discovery have revitalised interest in this modality (Santos-Júnior et al., 2024). Among these, the application of machine learning (ML) to peptide–target interaction (PTI) prediction has gained considerable momentum in both academic and pharmaceutical research (Maasch et al., 2023; Wan et al., 2024b). These models aim to accelerate peptide drug discovery by predicting key interaction parameters, such as binding affinity, pose, and binary binding status, using large-scale computational screens. Notably, novel ML architectures such as graph neural networks (GNNs) and transformer-based models have enabled substantial progress in learning from the complex biochemical and structural properties of peptides and proteins (Wu et al., 2021). By representing molecules as graphs or sequences, these models allow end-to-end learning from raw molecular inputs. In doing so, they reduce reliance on costly and time-consuming high-throughput screening (HTS) experiments (Jumper et al., 2021).

However, the principal challenge confronting systematic evaluation of current PTI modelling efforts, and the central motivating problem of this work, is the absence of standardised reporting practices across the literature. Variability begins with dataset selection and preprocessing: studies draw from heterogeneous sources such as BindingDB, PDBBind, and proprietary assays, each differing in structural resolution, functional annotations, and chemical diversity (Yang et al., 2019), and the problem is compounded when datasets are neither publicly available nor harmonised (de la Fuente-Nunez, 2024). More consequential for meta-analytic synthesis, however, is the inconsistent reporting of evaluation metrics. Regression-based tasks commonly report mean absolute error (MAE), root mean square error (RMSE), Pearson correlation coefficient (PCC), and the concordance index (CI), whereas classification models report area under the ROC curve (AUC), accuracy, F1-score, and Matthews correlation coefficient (MCC) (Chicco and Jurman, 2023). Prospective-hit-rate-benchmarked work, such as RF Diffusion (Watson et al., 2023), represents an important parallel evaluation paradigm, but retrospective-metric studies are required for a quantitative meta-analysis. In practice, only a narrow subset of metrics (AUC, the concordance index, and accuracy) are reported with sufficient consistency across studies to support meta-analytic pooling, and even these are frequently computed under divergent experimental conditions. Pooling such metrics is therefore not an endorsement of their adequacy as evaluation criteria (AUC in particular has well-documented limitations under class imbalance) but a documentation of what the field currently provides. The STRIDE framework, introduced below, is proposed as the remedy to this reporting deficit and as the foundation on which future PTI studies may be evaluated coherently.

A systematic evaluation of peptide–protein interaction prediction models is therefore crucial, especially when considering the wide range of data and methodological approaches currently in use. A meta-analysis aggregating reported performance metrics can reveal broader trends and highlight the strengths and weaknesses across models. A quantitative meta-analysis requires a metric that is consistently and comparably reported across primary studies. The retrospective-metric scope is therefore a methodological necessity, not an oversight. Throughout this review, the “target” in a peptide–target interaction refers to the biomolecular partner whose binding, affinity, or interaction status the model predicts. Across the included studies, targets comprised soluble and membrane proteins, enzymes, cell-surface and nuclear receptors (including G-protein-coupled receptors such as GCGR and GLP-1R), major histocompatibility complex (MHC)/human leukocyte antigen (HLA) molecules, ion channels and transporters, and protein–protein interaction interfaces. Where a study modelled an oligomer–protein system (for example, aptamer–protein interactions), it was considered eligible only when the peptide- or oligomer–protein interaction itself was the quantity being predicted. We use “target” in this operational sense rather than in the narrower pharmacological sense of a validated drug target.

Objectives

This systematic review and meta-analysis evaluates the predictive performance and methodological rigor of peptide–target interaction models published up to 1 August 2025, focusing specifically on the meta-analysis of author-reported values for AUC, Concordance Index, and Precision. These metrics were selected based on their frequency of use, interpretability, and relevance to both clinical and preclinical drug discovery.

This systematic review and meta-analysis was conducted in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 (Page et al., 2021) guidelines to ensure transparency, reproducibility, and methodological rigor throughout the review process. The review focuses specifically on published studies employing machine learning methods to predict peptide–target interactions, with performance evaluated using metrics such as area under the receiver operating characteristic curve (AUC), concordance index (CI), and Precision. These three metrics were selected as the primary outcomes of interest because they are among the most frequently reported measures of predictive performance in the field and provide insights into both ranking and Precision. This systematic review was conducted as the evidence-synthesis phase of a larger eDelphi consensus study.

Search strategy

A comprehensive literature search was conducted in August 2025 using three primary databases: PubMed, Cochrane and arXiv. The search strategy was developed to identify studies relevant to machine learning models applied to peptide–protein interaction prediction in a drug discovery context. For PubMed, the search combined controlled vocabulary and free-text terms relating to peptides, binding affinity, computational modeling, and performance metrics. Boolean operators were applied to maximise sensitivity while maintaining specificity. The results of this search strategy are described in detail in the Results section and the PRISMA diagram. The final query used was:

(“therapeutics” OR “Oncology” OR “peptide” OR “peptides” OR “binding affinity” OR “interaction prediction” OR “drug target” OR “target prediction” OR “binding prediction” OR “docking” OR “bioactivity”) AND (“machine learning” OR “deep learning” OR “neural network” OR “graph neural network” OR “transformer” OR “AI” OR “artificial intelligence” OR “structure-based” OR “prediction model” OR “computational model”) AND (“accuracy” OR “performance” OR “validation” OR “benchmark” OR “mean absolute error” OR “RMSE” OR “Pearson correlation” OR “ranking” OR “AUC” OR “F1 score”) AND (“drug discovery” OR “therapeutic” OR “pharmacology” OR “protein target”).

The PubMed search was limited to English-language articles indexed as journal publications with performance data. The arXiv search was performed using the advanced search interface, targeting titles containing the phrase “drug target binding” and including all cross-listed categories. Abstracts were included in the search, and results were sorted by most recent submission date. No date or subject restrictions were imposed on either database, allowing for the inclusion of a wide range of modeling approaches across different biomedical subfields. Given the computational nature of the topic, we prioritized arXiv to capture state-of-the-art pre-prints. A validation search in broader databases (e.g., Embase) confirmed that saturation was reached via PubMed/Cochrane/arXiv, yielding no additional unique eligible studies. These three databases were selected to balance biomedical, evidence-synthesis, and computational coverage. PubMed provides comprehensive indexing of the peer-reviewed biomedical and pharmacological literature and is the principal source for studies reporting experimentally grounded binding data. Cochrane was included to capture systematic-review and evidence-synthesis records and to align the search with established review methodology. ArXiv was prioritised because a substantial and growing fraction of state-of-the-art machine-learning methods for interaction prediction (particularly Transformer- and graph-neural-network-based architectures) appear first, and sometimes exclusively, as preprints; excluding it would have introduced a systematic lag against the most recent methodological advances.

Eligibility criteria

Studies were included in the review if they presented original research on computational models designed to predict peptide–target binding, bioactivity, or docking interactions. Eligible studies had to use machine learning, deep learning, or related artificial intelligence methods and report at least one of the following performance metrics: AUC, concordance index (CI), or Precision. We included architectures originally designed for small-molecule interactions only where they were explicitly applied to peptide sequences in the primary studies, demonstrating the sequence-agnostic nature of these deep learning frameworks. In addition, studies were required to assess prediction accuracy using a defined test set, external validation, or cross-validation strategy to ensure relevance to generalisable model performance. Peptide bioactivity endpoints defined against a target class, protein–protein and oligomer–protein interfaces where the interaction is the predicted quantity, and small-molecule architectures applied to peptide sequences are all eligible under these criteria, as are hybrid methods combining docking or knowledge-based matching with a learned component.

Studies were excluded if they did not report any of the three-target metrics, or if the performance values were only presented graphically without numerical data. Editorials, reviews, conference abstracts, and commentaries were excluded. Preprints from arXiv were included provided they met the other inclusion criteria and contained sufficient methodological and performance detail. For the purpose of inclusion, a “target” was any protein-class biomolecule (protein, enzyme, receptor, MHC/HLA molecule, channel, transporter, or protein–protein interface) against which peptide binding, affinity, or docking was computationally predicted.

Study selection

Study selection was conducted in two stages. In the first stage, titles and abstracts retrieved from Cochrane, PubMed and arXiv were independently screened by two reviewers (WW and AG) to assess preliminary eligibility. In the second stage, full texts of selected articles were reviewed in detail to confirm that they met all inclusion criteria. Disagreements between reviewers (WW and AG) were resolved through discussion, and when necessary, a third reviewer (HA) was consulted to reach consensus. A detailed screening log was maintained throughout the process to ensure transparency and traceability of inclusion decisions. The PRISMA flow diagram summarising the selection process is presented in Figure 1.

Data extraction

Data were extracted from each included study using a standardised template designed for this review. Extracted variables included bibliographic details (authors, year, journal), study objectives, model architecture (e.g., convolutional neural network, transformer, graph neural network), and input data type (e.g., amino acid sequence, structural features, physicochemical properties). Additional fields included the source and size of the training dataset, data splitting or cross-validation strategy, and negative sampling methodology in classification tasks.

The core focus of data extraction was on author-reported performance metrics. Specifically, values of AUC, concordance index (CI), and Precision were extracted as primary outcomes. When studies reported multiple models or validation folds, the mean value was recorded. Where available, the associated standard deviation, confidence interval, or number of validation folds was also noted. All values were taken as reported by the original authors, and no re-calculation or re-analysis of model outputs was performed.

Two reviewers (WW and AG) independently extracted data from each study, and any discrepancies were resolved by discussion. When data were missing or ambiguous, attempts were made to contact the study authors for clarification. In cases where numerical values were only presented in figures, digital extraction tools were used to estimate values when appropriate.

Risk of bias assessment

The methodological quality and risk of bias of each included study were assessed using a modified version of the Prediction model Risk Of Bias Assessment Tool (PROBAST) (Wolff et al., 2019), adapted for computational modeling studies. This adaptation involved evaluation of key domains such as dataset representativeness, clarity of model training and validation procedures, transparency of performance reporting, and reproducibility. Each study was rated as having low, high, or unclear risk of bias in each domain. The overall risk of bias was derived from the domain-level assessments, with high risk assigned if any critical domain showed high risk. Studies were also assessed for reproducibility indicators, including availability of source code, data accessibility, and use of external validation. The risk of bias assessments were independently performed by two reviewers. Divergences were discussed until consensus was reached. Although the PROBAST tool was not originally developed for use in computational prediction models, it was adapted here to capture the key risks relevant to peptide–target interaction modeling studies, such as overfitting, data leakage, and non-transparent evaluation protocols.

Data synthesis and analysis

We summarised key findings of included studies qualitatively and, where appropriate, quantitatively via meta-analysis. For studies sufficiently homogeneous in terms of reported metrics, we pooled performance measures using random-effects meta-analysis models (DerSimonian-Laird method). Due to the inconsistent reporting of confidence intervals, the n number from each of the primary studies was used for comparative analysis. We chose a random-effects model a priori given the anticipated heterogeneity in study populations, clinical tasks, and AI models. Pooled estimates with 95% confidence intervals (CI) were computed for the primary metrics of interest (Concordance Index, Precision, and AUC). Studies were weighted by the inverse of the sum of this derived variance and τ 2, and Q, τ 2 and I 2 were computed from the same variances. We assessed statistical heterogeneity using Cochran’s Q and the I 2 statistic, with I 2 > 75% indicating substantial heterogeneity. We also report τ 2 as the between-study variance. All meta-analyses were conducted using standard statistical software, and results are displayed in forest plots. We evaluated publication bias qualitatively (e.g., noting if only positive studies were published in certain domains) and with funnel plots for the main outcome. No formal patient or public involvement was applicable in this evidence synthesis, as it relied on previously published studies. The software used to conduct the meta-analysis was Stata Statistical Software Release 15 (StataCorp).

Harmonisation and model selection for pooled analysis

When studies reported results from multiple models, we extracted only the performance of the best-performing model, as defined by the authors, typically based on the highest AUC or CI. This approach avoided overweighting multi-model studies and ensured consistency across the dataset. In cases where model performances were similar, preference was given to models most representative of the study’s final conclusion or those validated externally. Where multiple variants of the same model architecture were presented, the variant incorporating the full input feature set or most advanced learning architecture was selected.

Study selection

The comprehensive literature search identified 4,175 unique records, of which 4,125, remained after deduplication. Title and abstract screening excluded 3,996, records due to irrelevance to peptide–target interaction prediction, absence of machine learning methodologies, or non-reporting of relevant performance metrics. Full-text screening was conducted on 129 articles, and 106 were excluded for reasons including non-use of peptide ligands (n = 37), lack of quantitative evaluation reporting metrics consistently (n = 42), and methodological ineligibility (e.g., rule-based or descriptive-only models, n = 27). Ultimately, 23 studies met the inclusion criteria and were retained for quantitative synthesis, comprising 14 studies reporting AUC, 5 reporting concordance index, and 6 reporting precision. This systematic review and meta-analysis was conducted in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 (Page et al., 2021). A PRISMA flow diagram of the study selection process is shown in Figure 1.

Characteristics of included studies

The 23 included studies span a diverse set of modeling strategies across a broader range of potential studies (Table 1), datasets, and peptide–target systems. Machine learning methodologies included deep neural networks, gradient boosting machines (XGBoost), support vector machines, and ensemble approaches incorporating graph-based or kernel-based methods. Dataset sources ranged from publicly available bioactivity repositories (e.g., PDBbind, BindingDB, DUD-E) to custom-curated and experimental datasets targeting specific protein families or pathologies.

Among the AUC-reporting studies, peptide ligands were assessed for binding to targets across oncology, immunology, and metabolic disorders. In contrast, studies reporting CI were predominantly focused on quantitative affinity prediction, using continuous binding data such as Kd or IC 50 values. Precision-reporting studies often focused on bacteriophage, viral, or host–pathogen interaction modeling, using either sequence-only or hybrid feature sets.

Reporting quality and risk of bias

The methodological quality and risk of bias of each included study were assessed using a modified version of PROBAST (Wolff et al., 2019). Assessment using the adapted PROBAST tool (Figure 2) indicated variable reporting quality. The majority of studies were rated as “unclear” or “high” risk in the domains of reproducibility and validation methodology. Very few performed external test set validation, and most relied on internal cross-validation without stratified splitting. Transparency was also limited: only one-third of studies provided detailed descriptions of hyperparameter tuning, and several lacked clarity in describing negative sample generation. These limitations constrain the generalisability of findings and underscore the need for methodological standardisation in the field.

Meta-analysis of AUC

A total of 14 studies (Figure 3) reported AUC values for binary classification of peptide–protein interactions. The reported AUC scores ranged from 0.6 to 0.98, with an overall AUC of 0.87 (Confidence Interval 95% 0.83–0.90), H value 27.74 and I 2 value 99.9%. Ten models achieved AUC values of at least 0.9 indicating strong discriminatory power in distinguishing binders from non-binders. However, AUC values also varied considerably, with two studies reporting performance below 0.75, possibly due to model simplicity, class imbalance, or limited dataset sizes. The inclusion of a broad AUC range underscores the variability in model generalisability and experimental design. The pooled estimate suggests that ML-based peptide classification models can reach excellent performance levels but are highly dependent on dataset quality and negative sampling strategies. No confidence intervals were reported in most studies, and between-study heterogeneity could not be formally quantified due to insufficient standard error data.

Concordance index (CI)

Five studies (Figure 4) reported concordance index (CI) values, with an overall result of 0.9 (95% Confidence Interval 0.89–0.91), with an H value 9,85 and an I 2 value 99%; CI is used primarily to evaluate regression-based models that predict continuous affinity values. Reported CI values were consistently above 0.9. This narrow distribution of CI values reflects strong internal model calibration for affinity ranking tasks. These studies often used structurally rich input features and large datasets. The relatively high CI values suggest that modern deep learning models can provide reliable affinity ranking, particularly when 3D structural information or advanced sequence embeddings are included.

Precision

Six studies (Figure 5) reported precision scores, resulting overall in a value of 0.75 (95% Confidence Interval 0.69–0.81), with an H value of 13.27 and an I 2 of 99.4%. These values reflect the models’ ability to correctly identify true binders among predicted positives. Importantly, none of the precision-reporting studies provided full confusion matrix data or recall/specificity measures, limiting the interpretability of standalone precision figures.

Overall performance trends

Across all three metrics (AUC, CI, precision), model performance ranged from moderate to excellent, with top-tier models achieving AUCs >0.95 and CIs >0.90. Studies that incorporated structural data, used balanced or stratified datasets, and implemented complex architectures (e.g., transformers or graph-based networks) tended to report higher values than simpler models, though no matched comparison was available. However, reproducibility and reporting standards were inconsistent, and no study reported all three metrics together. Furthermore, very few studies included confidence intervals or standard errors for any metric, highlighting a pervasive gap in statistical reporting.

This systematic review and meta-analysis synthesised the current landscape of machine learning models for PTI prediction, focusing on the most consistently reported performance metrics: AUC, Concordance Index, and Precision. While recent advances in deep learning and structural modeling have yielded high-performance predictions across diverse datasets, our analysis reveals persistent issues related to standardisation, reproducibility, and metric reporting.

Despite differences in architecture, from support vector machines to graph-based neural networks, the consistently high scores suggest that ML models, when trained on well-curated datasets, can reliably distinguish binders from non-binders; those models with improved standardisation of data correlated with increased reliability. However, performance varied substantially across studies, reflecting potential dataset imbalances, limited peptide diversity, and model simplicity. Such findings highlight the importance of rigorous dataset design and sampling strategies (Mizera and Latek, 2021); without balanced or representative datasets, even sophisticated models may yield suboptimal or misleading results (Wang et al., 2022; Yin et al., 2024). Current work typically constructs a balanced dataset by using under-sampling techniques, which removes samples from the majority class to match the size of the minority class (Yin et al., 2024; Rahman et al., 2024). The analysis of regression-based prediction models, evaluated via concordance index, was similarly promising. The narrow range of concordance index, values and their consistently high magnitude suggest that peptide ranking models are reaching maturity, especially when they incorporate structural descriptors or learn contextual embeddings from large protein sequence datasets (Taherzadeh et al., 2018). Only one-fifth of the studies in our review included the concordance index, despite its clear advantages over simpler metrics like Pearson correlation (Yin et al., 2024). Whereas Pearson correlation measures linear agreement between predicted and observed affinities (and can remain high even when a model misorders its top candidates), the concordance index, directly estimates the probability that a randomly chosen pair of compounds is ranked correctly, aligning the metric with the downstream task of prioritising candidates for experimental follow-up. Broader adoption of the concordance index would therefore facilitate more meaningful benchmarking across affinity prediction studies and improve downstream utility for virtual screening.

Precision, though reported less frequently, provided valuable insight into model selectivity. While promising, these figures are difficult to interpret without recall, specificity, or full confusion matrices (Romero-et al., 2022). Moreover, no study reported precision alongside all other standard classification metrics, limiting analysis of sensitivity–specificity trade-offs. Across all metrics, a clear association emerged between model complexity, feature richness, and predictive accuracy. Transformer-based architectures (which use a transformer and evolutionary-based attention mechanisms), Bidirectional Encoder Representations from Transformers (BERT) and Pseudo position specific scoring matrix -Discrete Wavelet Transform (PsePSSM-DWT) approaches to encode peptides (Khan et al., 2025), graph neural networks (Chandra et al., 2023), and hybrid ensemble models (Taherzadeh et al., 2018)) reported higher values than traditional classifiers. These models benefited from multi-modal inputs, including sequence, structure, docking scores, and physicochemical features, allowing them to learn richer representations of peptide–target interactions. Conversely, simpler models, such as logistic regression or traditional random forests (Thornton et al., 2021), tended to underperform, particularly in low-data or imbalanced settings.

Limitations

Despite these technical advances, the review identified limitations in reproducibility and transparency. Only a minority of studies provided access to source code or models. Few included external validation datasets, and none consistently reported confidence intervals for performance metrics. This is a critical shortfall, as confidence intervals provide essential context for understanding metric stability and statistical significance. The absence of such reporting limits the interpretability and comparability of published results. In addition, several studies did not specify key methodological details, such as data splitting strategies, negative sample generation, or hyperparameter tuning procedures. These omissions represent common sources of bias and overfitting in ML studies. Furthermore, heterogeneity in metric reporting remains a significant barrier. While AUC was the most commonly reported metric, Concordance Index and precision were rarely presented in tandem, and no study reported all three. The lack of standardised reporting practices prevents meaningful comparison across studies and undermines efforts to benchmark progress in the field. Drawing from related domains, the implementation of a minimum metric reporting standard, including AUC, Concordance Index, precision, recall, and confidence intervals, would greatly enhance the reliability and utility of future studies. Finally, this review reveals a need for greater emphasis on external validation, particularly on unseen peptide families or protein targets. Generalisation to novel targets is essential for practical applications in peptide drug discovery. In summary, our findings underscore both the potential and limitations of current ML models in PTI prediction. When rigorously trained and validated, modern deep learning and ensemble architectures can achieve high accuracy across both classification and regression tasks. However, the field continues to suffer from inconsistent metric use, limited transparency, and insufficient validation practices. Addressing these gaps through community-driven standards, improved dataset sharing, and robust statistical reporting will be vital to realising the full potential of ML in therapeutic peptide design.

Our search was restricted to PubMed, Cochrane, and arXiv. This carries a risk of omitting relevant peptide–target interaction models, most plausibly those published in engineering-oriented or computer-science venues that are indexed primarily in IEEE Xplore or Scopus rather than in PubMed. Although our validation search against a broader database identified no additional unique eligible studies (indicating that the retrospective, metric-reporting literature we sought was well represented in the sources searched) we cannot exclude the possibility that some eligible models, particularly recent conference proceedings, were not retrieved. Future updates of this review would benefit from incorporating multidisciplinary and engineering databases to further reduce this risk. Although our validation search against a broader database identified no additional unique eligible studies (indicating that the retrospective, metric-reporting literature we sought was well represented in the sources searched) we cannot exclude the possibility that some eligible models, particularly recent conference proceedings, were not retrieved. The observation that the twenty-three studies present challenges to a pooled estimate is itself a consequence of the reporting practices we document, and is why we advance STRIDE.

Future directions

Protein-target interaction (PTI) prediction faces fundamental challenges in evaluation consistency, reproducibility, and clinical translation. We propose the STRIDE framework (Standardization, Transparency, Representativeness, Integration, Discovery, and Evidence) as a methodological foundation for advancing PTI modelling. Drawing on AlphaFold’s transformative impact (Jumper et al., 2021; Thornton et al., 2021), we examine how rigorous benchmarking, honest uncertainty reporting, and community-driven standards can accelerate progress while acknowledging persistent limitations.

STRIDE framework components

S–Standardization: Unified evaluation protocols

Current Challenge: Inconsistent evaluation methods (disparate metrics (AUC, Concordance Index, RMSE, MCC), variable data splits, heterogeneous preprocessing, and inconsistent negative sampling) hinder meaningful performance comparisons (Akdel et al., 2022). Harmonized reporting frameworks essential for regulatory readiness remain absent (Kinch et al., 2021; Kryshtafovych et al., 2019; Wen et al., 2019). Prospective-validation benchmarking is itself contingent on first establishing standardised retrospective reporting; the contribution of this review is upstream of, and a prerequisite for, a future review of prospective validation.

Proposed Approach: Following CASP and CASF-2016 (Su et al., 2019) initiatives, we advocate standardised frameworks like APEX (Wan et al., 2024a) including temporal train/validation/test splits, unified preprocessing with defined clustering thresholds, biologically plausible negative sampling, statistical significance testing with multiple comparison correction (Wen et al., 2019), and comprehensive confidence intervals. AlphaFold’s development benefited from CASP’s standardized evaluation enabling objective comparison and rapid innovation (Jumper et al., 2021; Kryshtafovych et al., 2019). Building on TAPE and PDBbind (Rao et al., 2019; Wang et al., 2005), we explore the potential for a community benchmark suite for binding affinity prediction, classification, and generative tasks. The AlphaFold Database (>200 million structures) demonstrates the potential of standardized resources (AlphaFold Database, 2024; Varadi et al., 2022). Implementation should follow the TRIPOD + AI statement for reporting clinical prediction models using machine learning methods (Johner Institute, 2025; Collins et al., 2024), while adopting REFORMS reporting standards to ensure reproducibility (Kapoor et al., 2024). For studies incorporating large language models, the TRIPOD-LLM guideline provides additional reporting rigor (Gallifant et al., 2025).

T–Transparency: mechanistic interpretability and confidence reporting

Current Challenge: Clinical translation requires transparent reasoning beyond black-box predictions, including calibrated confidence estimates, failure mode identification, and mechanistic understanding (Akdel et al., 2022).

Critical Gaps: PTI models lack (1) rigorous separation of aleatoric uncertainty (experimental noise, biological variability) from epistemic uncertainty (incomplete training data) (Wang T. et al., 2025; Zhou et al., 2026), (2) calibration metrics critical for clinical decision-making requiring prediction intervals, likelihood ratios, and calibration diagnostics (Van Calster et al., 2019; Flores et al., 2025), (3) acknowledgment of trade-offs between deep model performance and interpretability (Bell et al., 2022; Kruschel et al., 2025; Lovo et al., 2025), and (4) clear linkage between ML predictions and biophysical principles, specifically binding free energy landscapes and thermodynamic parameters.

Proposed Approach: We advocate hierarchical interpretability: atomic-level residue attribution via Integrated Gradients (Sundarar et al., 2017) and SHAP (Linardatos et al., 2020; Wang L. et al., 2024), geometric surface fingerprinting via MaSIF (Gainza et al., 2020), and systems-level pathway analysis. AlphaFold’s confidence metrics (pLDDT, PAE) exemplify transparent uncertainty reporting (Jumper et al., 2021) and have been shown to encode aspects of protein flexibility and dynamics (Guo et al., 2022). While early analyses with standard full-MSA AlphaFold2 inference reported only modest correlation between pLDDT changes and experimental ΔΔG values (Pak et al., 2023), more recent binder-design pipelines have demonstrated that AF2 confidence metrics, applied as hard filters on generated candidates, substantially raise experimental hit rates and have in some cases produced order-of-magnitude improvements in binder thermal stability (Sumida et al., 2024). Comprehensive uncertainty quantification must nonetheless address distributional uncertainty (out-of-domain predictions) alongside aleatoric and epistemic components (Ajirak et al., 2025; Bensberg et al., 2025). In parallel, machine learning estimators of binding free energy are increasingly deployed as complementary filters within physics-based and generative design workflows, rather than as wholesale substitutes for physics-based ΔG calculations (Bensberg et al., 2025). Hybrid approaches combining molecular mechanics with machine learning are particularly promising in this layered configuration (Dong et al., 2021; Rube et al., 2022), and recent advances in quantum-classical free energy calculations illustrate how ML can enhance computational efficiency while preserving biophysical grounding (Bensberg et al., 2025). The accuracy–interpretability trade-off, however, remains a central challenge requiring careful evaluation (Atrey et al., 2026), with simpler interpretable models occasionally challenging the assumption that black-box approaches necessarily outperform them (Kruschel et al., 2025).

R–Representativeness: Addressing dataset biases and generalization

Current Challenge: Random splits do indeed fail to capture real-world generalization (Akdel et al., 2022), but the deeper problem is that none of the included benchmarks simulate prospective validation; even a carefully clustered train/validation/test split is still drawn from the same data-generating distribution as the training data.

Critical Gaps: Insufficient evaluation of out-of-distribution robustness to novel peptide sequences and structural shifts (Zhao and Sher, 2018; Dreyer et al., 2025). Negative sampling bias (synthetic versus biological negatives) artificially inflates metrics (Giziński et al., 2024), creating misleading real-world assessments.

Proposed Approach: Current datasets exhibit taxonomic bias (model organism overrepresentation), structural bias (crystallizable fold enrichment), chemical bias (modified peptide underrepresentation), and experimental bias (standardized assay conditions). AlphaFold training inherits these biases: 50%–70% of human proteins are glycosylated, yet AlphaFold 2 excluded post-translational modifications (Bagdonas et al., 2021; Bludau et al., 2022); over 30% of proteomes are intrinsically disordered (Ruff and Pappu, 2021; Tesei et al., 2024). AlphaFold 3 addresses some limitations by incorporating PTMs, ligands, and ions (Abramson et al., 2024). We recommend multi-dimensional evaluation across sequence space (clustering-based splits with varying identity thresholds), structural space (SCOP/CATH fold classifications), chemical space (modified peptides with diverse properties) (Feller and Wilke, 2025; Angelis et al., 2025), and temporal dimension (prospective validation). Approaches include balanced sampling, physics-based augmentation via molecular dynamics (Kong et al., 2026), active learning for underrepresented regions, and leveraging Propedia (Martins et al., 2023) and PepBDB (Wen et al., 2019). Multiple negative sampling strategies with stratified evaluation (Park and Marcotte, 2011) identify method-specific biases (Thrift et al., 2024; Wang X. et al., 2024). Recent work demonstrates that systematically benchmarking peptide-MHC predictors from synthetic to naturally processed epitopes reveals critical performance differences (Zhao and Sher, 2018), with enhanced methodologies substantially improving antigenic peptide discovery (Giziński et al., 2024). Fast and accurate structure prediction via equivariant diffusion models further advances the field (Frühbuß et al., 2025). Geometric deep learning architectures respecting physical symmetries (SE (3)-equivariant mechanisms (Fuchs et al., 2020), E(n)-equivariant GNNs (Satorras et al., 2021)) offer principled generalization (Ying et al., 2019; Jing et al., 2020; Marzella et al., 2024). Studies of antibody out-of-distribution generalization provide valuable frameworks for therapeutic protein design (Dreyer et al., 2025). Incorporating mechanistic sequence information can substantially improve generalization across diverse protein contexts (Shen et al., 2025). AlphaFold’s Invariant Point Attention achieves SE (3) equivariance ensuring prediction validity under rotations and translations (Jumper et al., 2021; Wells, 2022). AlphaFold 3 simplified this through data augmentation (Abramson et al., 2024; Simon, 2024).

I – integration: Multimodal data fusion

Current Challenge: PTI complexity necessitates diverse information integration, yet effective fusion architectures remain underdeveloped.

Critical Gap: Systematic representation benchmarking across sequence embeddings, structure-based features, and multimodal combinations remains absent, preventing identification of optimal encoding strategies (Nguyen and Hy, 2025; Wang L. et al., 2025).

Proposed Approach: Multimodal integration should leverage sequence information via protein language models (ESM-2, ProtT5) capturing evolutionary patterns (Lin et al., 2023; Wang L. et al., 2025), structural data (surface curvature, pocket topology, domain architecture), physicochemical properties (hydrophobicity, electrostatic potential, hydrogen bonding), geometric features via P2Rank (Krivák and Hoksza, 2018) and MaSIF (Gainza et al., 2020), and dynamic information from molecular dynamics via efficient atomistic foundation models (Kong et al., 2026). Recent benchmarking platforms enable systematic assessment of protein language models across diverse tasks (Çevrim et al., 2025), while comprehensive reviews of protein representation learning methods guide optimal selection strategies (Nguyen and Hy, 2025; Wang L. et al., 2025).

AlphaFold’s Evoformer exemplifies sophisticated fusion: 48 blocks shuttling information between MSA representations (evolutionary constraints) and pairwise residue features (structural constraints) (Jumper et al., 2021). Triangle multiplicative and attention updates enforce geometric consistency. Advanced architectures including cross-modal attention and contrastive learning may optimize integration. Joint embeddings of protein sequence and structure enable robust variant effect predictions (Blaabjerg et al., 2024), while synergising sequence and structure representations improves protein variant prediction (Chen and Chatterjee, 2024). Multimodal models integrating diverse data types show enhanced protein function prediction capabilities (Mao et al., 2025). Structure encodings combined with sequence embeddings advance transporter substrate prediction (Denger and Helms, 2025), and tokenized continuous embedding compressions efficiently represent both sequence and structure (Lu et al., 2025). Transfer learning by pre-training on large databases followed by task-specific fine-tuning (Satorras et al., 2021) shows promise. Evolution from AlphaFold 2 through AlphaFold-Multimer to AlphaFold 3 demonstrates continuous refinement (Jumper et al., 2021; Abramson et al., 2024). Integration with emerging methods creates synergistic workflows: AlphaFold 3 for protein-ligand geometries (Abramson et al., 2024), FLOWR. ROOT for SE (3)-equivariant affinity prediction (Cremer et al., 2025), ReaSyn for synthesizability assessment (Lee et al., 2026), and pruned atomistic foundation models for molecular dynamics (Kong et al., 2026). Sequence information should be encoded using high-dimensional embeddings from protein language models (including ESM-2, ProtT5, and Ankh) which capture evolutionary and contextual patterns, and can be combined with structural and physicochemical features in multimodal fusion architectures.

D–Discovery: Advancing molecular representations and architectures

Current Challenge: Feature representations must capture multi-scale interaction complexity while respecting fundamental physical principles.

Proposed Approach: Sophisticated encodings include graph-based hierarchies (atoms, residues, secondary structures, domains) with symmetry-respecting architectures like Geometric Vector Perceptrons (Satorras et al., 2021; Jing et al., 2020). AlphaFold demonstrates co-evolutionary signals (residues mutating together) encode proximity information improving accuracy (Jumper et al., 2021; Skolnick et al., 2021), learning local side chain packing for high-resolution refinement (Skolnick et al., 2021), though not explicitly revealing underlying thermodynamic forces. Machine learning-enhanced approaches can bridge this gap by connecting predictions to quantum-classical binding free energies (Bensberg et al., 2025), with automated resource allocation optimizing computational efficiency (Koby et al., 2025). AlphaFold’s Structure Module employs IPA generating 3D coordinates with SE (3) equivariance through pairwise distance computations cancelling rigid transformations (Jumper et al., 2021; Wells, 2022). End-to-end recycling enables progressive refinement. Biological databases exhibit extreme class imbalance; inadequate negative sampling inflates metrics (Ben-Hur and Noble, 2006; Chen et al., 2019; Zhao and Sher, 2018). AlphaFold-Multimer achieves ∼70% success for heteromeric and ∼72% for homomeric interfaces (Jumper et al., 2021; Zhu et al., 2023). Multiple negative sampling strategies with stratified evaluation (Park and Marcotte, 2011) identify method-specific biases (Thrift et al., 2024; Wang X. et al., 2024). Foundation model approaches combining pre-trained geometric encoders with generative frameworks show promise (Watson et al., 2023). EquiBind (Stärk et al., 2022) and DiffDock (Corso et al., 2022) exemplify geometric deep learning potential. In place of representations that encode only raw spatial coordinates, structure-based models should adopt explicit graph topologies in which atoms and residues are treated as nodes and chemical bonds and inter-atomic or inter-residue spatial distances are treated as edges. Graph neural networks and geometric (SE (3)/E(n)-equivariant) Transformers operating on such 3D structural graphs provide a principled inductive bias for peptide–target geometry and are expected to generalise better across structural shifts than coordinate-based encodings.

E–Evidence: Rigorous benchmarking and community standards

Current Challenge: Progress requires quantitative frameworks enabling systematic evaluation across multiple dimensions.

Critical Gaps: (1) Meta-regression approaches explaining performance heterogeneity through moderator analyses remain absent (Viechtbauer, 2010; Dietrichson et al., 2024; Bartoš et al., 2025), and (2) clear bridges from predictive metrics to downstream hit discovery and clinical utility are lacking (Ehrmann et al., 2023; Flores et al., 2025).

AlphaFold as Exemplar: AlphaFold achieved median GDT scores of 92.4 in CASP14, demonstrating experimental-level accuracy (Jumper et al., 2021; Thornton et al., 2021). It excels at single-chain proteins with evolutionary depth (Jumper et al., 2021), oligomers (Jumper et al., 2021; Zhu et al., 2023), and domain packing (2,180-residue chains) (Jumper et al., 2021). Limitations include intrinsically disordered regions (>30% of proteomes) (Ruff and Pappu, 2021; Tesei et al., 2024; Alderson et al., 2023), conformational ensembles where single-structure predictions miss functional dynamics (Bowman, 2024; Strodel, 2021), ligand-induced states (AlphaFold 3 shows ≥50% improvement) (Hekkelman et al., 2023; Abramson et al., 2024), novel multimers lacking MSA signals (Bryant et al., 2022; Mirabello et al., 2024), and environmental context (pH, lipids, membranes, PTMs) (Bagdonas et al., 2021; Bludau et al., 2022).

Deep mutational scanning via DMS-Fold improves 88% of targets (Drake et al., 2025), though pLDDT changes show limited ΔΔG correlation (Pak et al., 2023). ML methods work well for single-funnelled energy landscapes but struggle with rugged, multi-funnelled landscapes of IDPs and aggregation (Strodel, 2021). AlphaFold learns interaction patterns but does not explicitly model thermodynamic or kinetic physics (Jumper et al., 2021). High predictive performance alone (high AUC) is insufficient for clinical deployment; calibration represents the Achilles heel of predictive analytics (Van Calster et al., 2019), with clinical priorities requiring alignment of evaluation metrics with medical decision-making contexts (Flores et al., 2025). Important considerations when translating machine learning models from development to clinical bedside include model actionability and integration into clinical workflows (Ehrmann et al., 2023; Deng et al., 2025). Machine learning’s impact on clinical decision-making depends critically on these translational factors (Adlung et al., 2021). Progression from AlphaFold 2 through AlphaFold-Multimer to AlphaFold 3 demonstrates refinement guided by comprehensive benchmarking (Jumper et al., 2021; Jumper et al., 2021; Abramson et al., 2024): AlphaFold 2’s 48-block Evoformer with IPA yielded exceptional performance; AlphaFold-Multimer extended to complexes; AlphaFold 3 simplified architecture while expanding molecular coverage. Community Infrastructure: Sustainable progress requires governance (professional oversight, specialized review), technical frameworks (minimum reporting guidelines analogous to CONSORT/MIAME) (Collins et al., 2024; Kapoor et al., 2024), and collaborative infrastructure (shared resources, integration platforms). CASP competitions demonstrate community-driven bench marking’s power (Kryshtafovych et al., 2019). Similar initiatives (Wen et al., 2019; Stärk et al., 2022; Corso et al., 2022) including CASF and CAPRI provide proven frameworks. AlphaFold exemplifies open science through complete code release, trained weights, and the AlphaFold Database (>200 million structures) (Jumper et al., 2021; AlphaFold Database, 2024; Varadi et al., 2022). Community tools like ColabFold (Mirdita et al., 2022), AlphaFill (Hekkelman et al., 2023), and AF_unmasked (Mirabello et al., 2024) enable innovation beyond original implementations. Community evaluation reveals strengths and limitations for missense variants and ligand binding (Akdel et al., 2022), ensuring appropriate application and guiding development.

This systematic review and meta-analysis evaluated the performance and methodological quality of ML models developed for peptide-target interactions, encompassing 23 studies published through August 2025. The analysis reveals substantial progress in applying deep learning, ensemble methods, and graph-based architectures to PTI prediction tasks, yet highlights persistent challenges in evaluation consistency, reproducibility, and translational potential that must be addressed for clinical application. The STRIDE framework provides a comprehensive roadmap for advancing PTI prediction from current proof-of-concept demonstrations toward clinical applicability. Drawing on lessons from AlphaFold’s transformative impact on structural biology (Jumper et al., 2021; Thornton et al., 2021), we demonstrate how rigorous benchmarking, transparent uncertainty reporting, and community-driven standards can accelerate progress while maintaining honest assessment of capabilities.