PERSPECTIVE article

Front. Bioinform., 20 November 2026

Sec. Protein Bioinformatics

Volume 6 - 2026 | https://doi.org/10.3389/fbinf.2026.2010708

Epistemic blind spots in protein function prediction: uncertainty, homology bias and failure modes in deep learning models

  • 1. Tecnologico de Monterrey, Escuela de Ingeniería y Ciencias, Querétaro, Mexico

  • 2. Instituto de Ciencias Aplicadas y Tecnología (ICAT), Universidad Nacional Autónoma de México, Mexico City, Mexico

  • 3. Instituto de Investigaciones en Matemáticas Aplicadas y en Sistemas, Unidad Académica del Estado de Yucatán, Universidad Nacional Autónoma de México, Mérida, Yucatán, Mexico

Abstract

Protein function prediction has advanced from homology-based annotation transfer to deep learning, protein language models, and multimodal foundation models. Yet improved benchmark performance does not resolve a more basic problem: these systems learn from biological knowledge that is incomplete, heterogeneous in evidential support, and strongly structured by evolutionary relationships. We organize this problem as two sequentially dependent epistemological levels. The first concerns biological knowledge: what functional states become observable, with what evidential support and evolutionary structure. The second concerns what models can legitimately infer from that supervision, including representation sufficiency, uncertainty, generalization, and the scope of a universal computational blueprint for protein function inference. Because limitations at the first level enter the learning problem before model fitting, benchmark performance and model uncertainty cannot by themselves establish certainty about latent biological function. Trustworthy protein AI therefore requires explicit attention to applicability domains, distribution shift, uncertainty, and the limits of inference from available biological knowledge.

1 Introduction

Protein function prediction (PFP) has undergone four major methodological transitions over the past 3 decades: homology-based annotation transfer, benchmark-driven machine learning, deep supervised prediction, and, most recently, protein language models (PLMs) and multimodal foundation models. Each transition has expanded functional inference from explicit sequence similarity and engineered features to learned sequence representations increasingly complemented by structural, network, and ontology-derived information (Altschul et al., 1990; Radivojac et al., 2013; Kulmanov et al., 2018; Rives et al., 2021; Elnaggar et al., 2022; Bileschi et al., 2022). These advances have substantially improved the accuracy and scope of computational protein annotation.

However, as predictive performance has improved, the field has also become confronted with questions that benchmark performance alone cannot answer. Many of these questions are epistemological rather than algorithmic, arising from the interaction between incomplete biological knowledge and increasingly powerful learning algorithms. Protein function annotation is inherently an open-world problem: annotations are incomplete, the absence of annotation does not imply the absence of function, and negative instances are poorly defined (Youngs et al., 2014; Jiang et al., 2014). These limitations become particularly important under distribution shift, when predictors encounter novel families or divergent proteins that differ substantially from their training data.

Because homology pervades the inference and propagation of functional annotations, an equally important question is whether deep learning truly overcomes homology bias or merely learns more sophisticated representations of evolutionary relatedness (Rives et al., 2021; Elnaggar et al., 2022). This success sharpens rather than resolves the underlying question: how much of a model’s predictive performance derives from evolutionary and other statistical regularities present in protein sequences, and to what extent does it support functional generalization beyond familiar regions of protein space? Increasingly powerful models may still fail silently on proteins that depart from their training distribution, which makes uncertainty estimation, calibration, and selective abstention important for determining when their predictions can be trusted.

Earlier work has addressed open-world annotation and negative selection (Dessimoz et al., 2013; Youngs et al., 2013; 2014), incomplete knowledge in evaluation (Jiang et al., 2014), evolutionary relatedness in PLM-based inference (Kabir et al., 2024; Kilinc et al., 2025), and predictive uncertainty under distribution shift (Guo et al., 2017; Ovadia et al., 2019). We connect these as two sequentially dependent epistemological levels. The first concerns what biological function becomes observable, with what evidence and evolutionary structure; the second, what models can infer from that supervision. Our contribution is to trace how limitations at the first level carry through dataset construction, training, evaluation, and prediction, restricting the claims that benchmark results can support.

This dependency also points to characteristic failure modes. Incomplete or unevenly supported annotations can distort supervision, including through mislabeled operational negatives; evolutionary overlap can inflate apparent generalization; and distribution shift can degrade performance and calibration. Better models may reduce downstream error, but do not show that uncertainty in supervision has been resolved. Predictions should therefore be interpreted relative to the evidence, assumptions, and applicability domain supporting them. Supplementary Material S1 formalizes this relationship.

2 The epistemology of biological knowledge

2.1 The open-world nature of protein function annotation

A central epistemic limitation of protein function prediction is the distinction between biological function and observed annotation. Annotation resources capture only a subset of a protein’s functional repertoire, including functions that may emerge under specific cellular or environmental conditions. Under the open-world assumption (OWA), absence of annotation therefore does not imply absence of the corresponding function (Dessimoz et al., 2013).

This incompleteness directly complicates supervised learning because negative instances are difficult to define. An unannotated protein–function relationship may represent either a function that is truly absent or one that has not yet been observed or characterized, while evidence demonstrating functional absence is considerably scarcer than evidence supporting functional presence (Youngs et al., 2014). The unannotated space therefore cannot generally be interpreted as a reliable negative class.

Binary and multilabel classifiers nevertheless require operational negatives, commonly constructed heuristically from unlabeled data. These choices can materially alter supervision. By restricting negative examples to genes annotated within the same GO branch, Youngs et al. (2013) reduced verifiable positives incorrectly assigned to the negative class by 91.5% in yeast and 22.5% in mouse. The problem is therefore not merely class imbalance, but class indeterminacy: unlabeled data contain an unknown mixture of truly negative and yet undiscovered positive relationships.

This has motivated positive–unlabeled (PU) formulations of protein function prediction. PU-GO, for example, formulated the task as positive–unlabeled ranking and reported improved robustness on similarity based and time based benchmarks (Zhapa-Camacho et al., 2024). From an epistemic perspective, negative labels derived from unlabeled data are assumptions introduced by the learning framework rather than direct observations of biological absence. The annotation space should therefore be understood as a partially observed functional state, not a complete positive–negative representation of biological reality.

2.2 The uncertainty embedded in biological annotations

Even positively annotated protein–function relationships do not necessarily have equivalent epistemic status. Functional annotations are supported by heterogeneous evidence, including experimental assays, phylogenetic inference, sequence or structural similarity, computational analysis, curator interpretation, and automated annotation. The Gene Ontology represents these differences through evidence codes grouped into experimental, phylogenetic, computational, author, curatorial, and automatically generated categories (The Gene Ontology Consortium, 2021). A positive annotation should therefore be understood not simply as a binary statement of biological truth, but as a functional assertion supported by particular evidence.

Transforming annotations into categorical labels for machine learning can obscure differences in provenance and evidential support. Errors may also arise from incorrect functional transfer or excessive specificity. In 7,255 sequences from 37 well-characterized enzyme families, Schnoes et al. (2009) found substantial variation in misannotation across databases and enzyme superfamilies; in NCBI non-redundant protein database (NR), 85% of identified misannotations correspond to overannotation. Although these estimates reflect specific families and historical database versions, they demonstrate that functional labels can contain systematic uncertainty rather than purely random noise.

The problem is broader than conventional label noise. Annotations differ not only in whether they are correct, but also in how strongly and directly the available evidence supports them. Ignoring these differences during training can leave models with heterogeneous supervision while their predictions appear uniformly certain. Confidence in a prediction must therefore be distinguished from confidence in the biological knowledge used to supervise it. Model uncertainty and annotation uncertainty are related, but they are not the same epistemic quantity.

2.3 Can deep learning escape homology?

Homology occupies a distinctive position in protein function prediction. It is not an epistemic defect: evolutionary relationships provide biological evidence, and homology-based transfer remains a foundation of computational annotation (Loewenstein et al., 2009). The relevant question is not whether deep learning eliminates homology, but what form of generalization an evaluation design supports.

PLMs support remote homology detection at low sequence identity (Liu et al., 2024; Kilinc et al., 2025). In a zero-shot remote homology task, not direct functional annotation, Kabir et al. (2024) reported that ProtT5 AUPRC decreased from approximately 56%–35% as the maximum sequence identity threshold decreased from 95% to 10%. Although this is not a function-prediction result, it is relevant because evolutionary relationships are an important source of evidence for function transfer, and their detectability changes as sequence identity decreases.

Three levels of generalization should be distinguished. First, sequence-identity filtering tests generalization beyond the similarity criterion used to construct the benchmark, not evolutionary independence. Second, a low-identity test protein may still belong to a family or superfamily represented in supervised training and remain detectable by profile-based methods; performance then shows transfer across sequence divergence, but within detectable evolutionary relationships. Third, a stricter design excludes the same family or superfamily and screens for residual relationships with profile-based or structure-aware methods. If none is detected, performance supports generalization beyond the detection capacity of those methods, not proof that no evolutionary relationship exists. Homology bias arises when unaccounted evolutionary overlap supports a stronger generalization claim than the evaluation design warrants.

Family-level separation reduces close evolutionary overlap but depends on family definitions and can leave deeper relationships. Structural comparison can reveal relationships missed by sequence similarity, but structural similarity does not by itself establish common ancestry. Temporal separation approximates prospective evaluation and avoids training on annotations that appeared after the cutoff, but it does not remove evolutionary relatedness. For PLMs, downstream supervised overlap must also be distinguished from self-supervised pretraining exposure. Sequence exposure during pretraining is not equivalent to functional-label leakage, but an evaluation protein or close relative may not be unseen at the representation-learning stage. Evaluation should state the controls used and the claims they support.

Figure 1 summarizes these limitations. For example, treating an unannotated protein–function pair as a negative can make a prediction later supported by evidence count as a false positive, although biological absence was never established. Supplementary Section 1.9 develops this case, and Supplementary Table S1 maps this and four further failure modes to their controls and residual limitations.

FIGURE 1

3 The epistemology of protein function learning

3.1 What are deep learning models actually learning?

The increasing predictive capacity of deep learning raises a broader epistemic question: what information do these models learn about protein function? PLMs trained on large sequence collections develop representations associated with evolutionary relationships, structure, and functional properties even when these attributes are not explicitly provided during pretraining (Rives et al., 2021; Elnaggar et al., 2022). Yet biologically meaningful representations should not be interpreted as complete representations of biological function. In supervised protein function prediction, the immediate learning target is typically an operational label constructed from the observable annotation state, which itself is only a partially observed and epistemically heterogeneous approximation of biological function.

Sequence based predictors show that protein sequence contains substantial information for functional inference. DeepGOPlus combines sequence derived features with sequence similarity to predict Gene Ontology annotations (Kulmanov and Hoehndorf, 2020). Modern predictors, however, increasingly incorporate structure, interaction networks, ontology derived information, and other modalities. DeepFRI combines sequence derived features with structural information through graph convolutional networks and improves prediction relative to sequence based convolutional models (Gligorijević et al., 2021). Contemporary approaches increasingly integrate multiple biological information sources (Boadu et al., 2025).

Multimodal models make visible a distinction between information contained in a biological system and information recoverable by a particular representation and learning algorithm. Improved prediction after incorporating structure or biological context does not necessarily mean that the corresponding information is absent from sequence. It may instead be difficult to recover from sequence alone, insufficiently represented by the learned embedding, or more directly accessible through another modality. The practical question is whether a representation provides sufficient information for the functional inference being attempted, rather than simply whether sequence “contains” protein function.

Predictive sufficiency should also not be confused with mechanistic understanding. A model may learn statistical regularities strongly associated with functional annotations without recovering the biological mechanisms that generate those functions, while additional modalities may improve prediction without resolving uncertainty in the annotations themselves. Successful prediction shows that a representation contains information useful for functional inference, but not that it provides a complete or mechanistic model of biological function.

3.2 Uncertainty and trustworthy protein AI

Predictive accuracy alone is insufficient to determine when a protein function prediction should be trusted. A predictor may perform well on average while remaining overconfident on individual errors. Calibration addresses this by asking whether predictive confidence corresponds to empirical correctness, rather than merely whether predictions are accurate (Guo et al., 2017). The challenge becomes more pronounced under distribution shift, where both predictive performance and uncertainty estimates may deteriorate as test data depart from the training distribution (Ovadia et al., 2019), for instance when models encounter divergent sequences or remote protein families poorly represented during training.

Predictive uncertainty is commonly separated into aleatoric uncertainty, associated with variability in the observed prediction problem, and epistemic uncertainty, associated with limited model knowledge (Kendall and Gal, 2017; Hüllermeier and Waegeman, 2021). Biological knowledge uncertainty arises upstream from incomplete annotations, heterogeneous evidence, and missing context. It is not a separate component estimated by the model, although it shapes the data from which predictive uncertainty is estimated, and has no one-to-one correspondence with either category. For example, epistemic uncertainty may increase for a protein family poorly represented in training, whereas inconsistent labels among otherwise similar proteins may contribute to apparent aleatoric variability in a fixed dataset even when that inconsistency originates from the annotation process. The same unresolved annotation may behave as irreducible label noise relative to the dataset as given yet become reducible biologically through additional evidence.

This distinction also complicates the interpretation of apparent prediction errors. Under the open-world assumption, disagreement between a prediction and the current annotation corpus does not necessarily imply biological incorrectness. Jiang et al. (2014) showed that incomplete experimental annotations can affect performance estimates, with the magnitude and direction depending on the prediction method, evaluation metric, and ontology structure, though large-scale evaluations can remain informative despite incomplete knowledge. The epistemic limitation is therefore not that current benchmarks are invalid, but that their outcomes quantify agreement with the currently observable functional state, not with a complete biological ground truth.

Methods act at different levels: positive–unlabeled or evidence-aware supervision can make assumptions about the annotation process explicit or reduce their influence on training, whereas calibration, out-of-distribution detection, and selective prediction characterize model behavior and applicability. Boger et al. (2025), for example, applied conformal prediction and conformal risk control to protein search and functional annotation, transforming raw model scores into calibrated probabilities and retrieval sets with user-specified statistical risk guarantees. These guarantees depend on assumptions linking calibration data to future observations. Their value lies in extending the question from what function a model predicts to the conditions and statistical guarantees under which that prediction can be trusted.

A trustworthy predictor should communicate both its functional prediction and the limits of the inference supporting it. Uncertainty quantification is useful not because it eliminates epistemic uncertainty, but because it can make the boundaries of predictive knowledge explicit. Beyond those boundaries, abstention is preferable to silent failure.

3.3 Toward a universal computational blueprint for protein function inference

The preceding limitations raise a broader question: what would universality mean in protein function prediction? Here, universality does not denote a single architecture that assigns every possible function to every protein. We use the term to describe a universal computational blueprint for inference: one that identifies which molecular and contextual variables are sufficient for a specified functional target, generalizes beyond familiar evolutionary and statistical relationships, and recognizes when available information is insufficient for prediction.

The required information may differ across functional scopes. A molecular function assignment may sometimes be supported predominantly by sequence and structure, whereas biological process or cellular component assignments may require interaction, localization, or condition-specific information. Individual proteins can also perform different functions across contexts (Singh and Bhalla, 2020), and localization can alter which interactions and activities are available (Sigaeva et al., 2026). Universality therefore should not assume that the same modalities are sufficient for every functional question.

Under this definition, progress toward universality should be evaluated by the scope and limits of justified inference rather than coverage alone. A more general predictor should degrade gracefully, or abstain, across increasingly distinct evolutionary and statistical regimes, and use additional context when it contributes information not recoverable from the available molecular representation. Progress could be assessed by coverage at a fixed risk level under defined shifts, together with target-specific gains from adding relevant contextual information. Abstention when information or support is absent is therefore part of this notion of universality, not evidence against it.

Whether a finite, learnable representation can support this form of inference remains open. If relevant molecular and contextual variables can be observed, increasingly general models may approximate the mapping to functional state. If essential context remains unobserved, a universal blueprint would instead need to recognize non-identifiability.

These relationships among representation, context, applicability domains, and a universal computational blueprint are summarized in Figure 2.

FIGURE 2

4 Discussion

Advances in deep learning and protein foundation models have substantially expanded the capacity to infer protein function from biological data. Yet increasing predictive performance does not eliminate the epistemic limitations of the knowledge from which these models learn. Protein function annotation remains incomplete, negative instances are intrinsically difficult to establish, functional assertions differ in evidential support, and evolutionary relationships shape both the construction of biological knowledge and its computational inference. Together, these properties form the epistemic substrate on which protein function predictors are trained and evaluated.

Several practical decisions follow from this view. Benchmark designers should report how operational negatives were constructed, how evolutionary overlap between training and evaluation data was controlled, and, where evidence metadata permit, results by evidence type. Model developers should report calibration against the operational target used, evaluate performance under relevant distribution shifts, and define an applicability domain outside which the model abstains. Users should interpret benchmark results as evidence about the currently annotated state, not as validation against complete biological truth. Predictive success does not establish mechanistic understanding, and low sequence similarity does not establish evolutionary independence.

A near-term agenda has three priorities. First, calibration should be reported against a stated operational target and, where evidence permits, stratified by evidence provenance. Second, applicability domains should be tested under defined evolutionary and statistical shifts, including selective prediction when support is insufficient. Third, context sufficiency should be tested for each functional target by comparing molecular-only and context-augmented models.

A longer-term question is whether these principles can converge into a universal computational blueprint for protein function inference. Progress would then depend less on identifying a single architecture than on determining which molecular and contextual variables are sufficient for a specified functional target, how models generalize beyond supported regimes, and when available information is insufficient for prediction. Whether such a blueprint exists remains unresolved. Its pursuit, however, reframes progress in protein function prediction: the objective is not only to predict more functions, but to establish what can be inferred from available biological knowledge, why that inference is justified, and where its limits lie.

Statements

Data availability statement

The original contributions presented in the study are included in the article/Supplementary Material, further inquiries can be directed to the corresponding author.

Author contributions

LL-D: Conceptualization, Writing – original draft, Writing – review and editing, Project administration. EC-D: Formal Analysis, Writing – review and editing. EP-R: Writing – review and editing. RC-E: Writing – review and editing, Funding acquisition.

Funding

The author(s) declared that financial support was received for this work and/or its publication. This work was supported by the Instituto Tecnológico y de Estudios Superiores de Monterrey through the Publications Support Fund (FAP, by its initials in Spanish) to RCE, and funded by the Dirección General de Asuntos del Personal Académico, Universidad Nacional Autónoma de México (IN206326) to EPR.

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. Generative artificial intelligence tools, including ChatGPT (GPT-5.6 Sol, OpenAI) and Claude Sonnet 5 (Anthropic), were used during manuscript preparation to assist with language editing, synthesis, organization, and reference checking. All AI-assisted content, scientific claims, interpretations, citations, and references were critically reviewed by the authors, who take full responsibility for the accuracy, integrity, and final content of the manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fbinf.2026.2010708/full#supplementary-material

References

Summary

Keywords

calibration, homology bias, open-world learning, protein function prediction, protein language models, trustworthy AI, uncertainty

Citation

Ledesma-Dominguez L, Carbajal-Degante E, Perez-Rueda E and Cortes-Espinosa R (2026) Epistemic blind spots in protein function prediction: uncertainty, homology bias and failure modes in deep learning models. Front. Bioinform. 6:2010708. doi: 10.3389/fbinf.2026.2010708

Received

26 September 2026

Revised

20 October 2026

Accepted

28 October 2026

Published

20 November 2026

Volume

6 - 2026

Edited by

Rita Casadio, University of Bologna, Italy

Reviewed by

Yiquan Wang, Shenzhen Bay Laboratory, China

Updates

Copyright

*Correspondence: Leonardo Ledesma-Dominguez,

Disclaimer

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.

Outline

Figures

Cite article

Copy to clipboard


Export citation file


Share article

Article metrics