Integrative Biomedical Research

1. Introduction

Breast cancer remains, by a considerable margin, the malignancy most frequently diagnosed among women worldwide, and it continues to rank among the leading causes of cancer-related death, with more than 2.3 million new diagnoses recorded each year (Kotsifaki et al., 2025; Sabit et al., 2026). Ovarian cancer, though far less common, tells a similarly troubling story — modest in incidence, disproportionately severe in mortality (Webb & Jordan, 2024; Wolde & Belay, 2026). Both diseases arise through a tangle of somatic mutations, epigenetic drift, and shifting tumor-microenvironment signals that rarely act in isolation (Sabit et al., 2026; Wolde & Belay, 2026). At the center of that tangle, for a meaningful subset of patients, sit two genes whose names have become almost synonymous with hereditary cancer risk: BRCA1 and BRCA2 (Ismail et al., 2024; Kotsifaki et al., 2025).

It is worth pausing on what these genes actually do, because their function explains much of what follows. BRCA1 and BRCA2 are custodians of genomic integrity, core members of the homologous recombination repair (HRR) pathway responsible for fixing double-stranded DNA breaks with high fidelity (Ismail et al., 2024). Without functional HRR, such damage does not simply disappear — it compounds. Inactivating mutations in BRCA1 or BRCA2 disable this repair system, and the resulting genomic instability translates, over time, into a markedly elevated lifetime risk of breast and ovarian cancer (Ismail et al., 2024; Shah et al., 2025). For much of the last several decades, clinicians confronting BRCA-associated disease had little choice but to reach for the same blunt instruments used against cancer generally, applied broadly and often at considerable cost to patients' quality of life (Ismail et al., 2024; Kotsifaki et al., 2025).

That began to change with precision oncology, which has gradually replaced one-size-fits-all treatment with therapies aimed at specific molecular vulnerabilities (Rescigno & Greystoke, 2026; Sabit et al., 2026). The clearest success story involves poly (ADP-ribose) polymerase (PARP) inhibitors — olaparib, rucaparib, and niraparib among them — which exploit synthetic lethality (Ismail et al., 2024; Shah et al., 2025). In tumors already deficient in homologous recombination, PARP inhibitors trap PARP1/2 on DNA and block base-excision repair, converting otherwise manageable single-strand breaks into lethal double-strand breaks during replication; healthy cells largely tolerate this, BRCA-deficient tumor cells do not (Shah et al., 2025). Large trials have confirmed that PARP-inhibitor maintenance therapy meaningfully extends progression-free and overall survival in patients carrying pathogenic BRCA alterations (Rescigno & Greystoke, 2026; Shah et al., 2025). None of that benefit, though, is available to a patient whose variant has not been correctly classified — which is where the promise of precision oncology runs into a messier reality (Eniu et al., 2023; Rescigno & Greystoke, 2026).

The bottleneck is not sequencing capacity; next-generation sequencing (NGS) now identifies genetic alterations faster than clinical science can make sense of them (Ismail et al., 2024; Rescigno & Greystoke, 2026). The bottleneck is interpretation. Roughly 40% of newly detected missense variants across clinical gene panels cannot yet be confidently labeled benign or pathogenic and are filed away as variants of uncertain significance, or VUS (Rescigno & Greystoke, 2026). This is not a minor administrative inconvenience — missense VUS clustering in functionally sensitive regions, such as the C-terminal BRCT domains of BRCA1 or the sprawling coding sequence of BRCA2, routinely complicate an otherwise straightforward clinical report (Ismail et al., 2024; Rescigno & Greystoke, 2026). The stakes cut in both directions: overcalling a benign variant can send a patient toward unnecessary prophylactic surgery, while undercalling a truly pathogenic one denies access to a therapy that might have changed their prognosis (Rescigno & Greystoke, 2026).

Traditional routes to resolving this ambiguity — functional assays performed one variant at a time, structural modeling, multigenerational pedigree analysis — were never designed for today's data volume (Ismail et al., 2024; Rescigno & Greystoke, 2026). They are slow, costly, and not built to scale (Rescigno & Greystoke, 2026; Sabit et al., 2026). This widening gap between sequencing speed and interpretive capacity has created real urgency around computational alternatives that can accelerate variant pathogenicity prediction without sacrificing clinical validity (Gentile et al., 2026; Rescigno & Greystoke, 2026).

Artificial intelligence and machine learning have moved into this gap with some force (Barua et al., 2026; Lin et al., 2026). Earlier approaches — random forests, support vector machines, conventional deep neural networks — showed promise but struggled with the dimensionality of raw sequence data, often overfitting or offering little insight into why a prediction was made (Qiu et al., 2026). A more substantial leap arrived when researchers began borrowing architectures from natural language processing and applying them to biological sequences (Jurenaite et al., 2024). The premise: treat a protein or genomic sequence as a language, learn which "words" plausibly follow which others, and in doing so capture the physical, chemical, and evolutionary constraints that shaped that sequence (Jurenaite et al., 2024).

Protein language models, or pLMs, have emerged from this line of work as unusually capable tools for the problem this review addresses (Corso et al., 2026). Trained on vast, label-free evolutionary databases, they learn the "grammar" of protein function without curated ground-truth labels (Corso et al., 2026). AlphaMissense offers proteome-wide missense-variant predictions with sensitivity and specificity marking a genuine step change (Corso et al., 2026). ESM3 uses evolutionary protein data not only to predict mutational consequences but to generate novel functional proteins outright (Corso et al., 2026). At the nucleotide level, DNABERT-S applies BERT-style architectures, trained across millions of genomic sequences, to interrogate mutation signatures and regulatory context (Corso et al., 2026). What unites these systems is their capacity to separate genuine driver mutations from biological noise (Jurenaite et al., 2024; Qiu et al., 2026).

The clinical implications are considerable. By flagging pathogenic variants quickly, pLMs support detection of homologous recombination deficiency (HRD), present in roughly half of high-grade serous ovarian carcinomas and in 50–70% of triple-negative breast cancers (Shah et al., 2025). Tools such as HRProfiler mine exome-sequencing data for genomic scars — loss of heterozygosity, telomeric allelic imbalance, large-scale state transitions — achieving AUC values above 0.90 for HRD detection and widening the pool eligible for PARP-inhibitor therapy (Shah et al., 2025). Layered onto electronic health records, these predictions can feed decision-support systems that assist molecular tumor boards, reducing subjectivity and manual workload (Lin et al., 2026; Sabit et al., 2026; Rescigno & Greystoke, 2026). Large language model ensembles like OncoChat have gone further still, learning gene-gene interactions directly from targeted panels and rediscovering BRCA-PARP1 synthetic lethality in an unsupervised fashion (Liu et al., 2025). Taken together, this evidence suggests a genuine, if not yet complete, shift away from static, morphology-driven diagnosis toward something more dynamic and probabilistic (Barua et al., 2026; Sabit et al., 2026).

Still, computational promise and clinical readiness are not the same thing. Translating in silico pLM predictions into everyday oncology practice demands a clear-eyed evaluation of biological grounding, analytical performance, and the practical obstacles between a benchmark result and a bedside decision (Lin et al., 2026; Sabit et al., 2026). This review synthesizes current evidence at the intersection of protein language models, BRCA variant pathogenicity, and precision oncology, offering clinicians and computational biologists a clearer map of where this field currently stands (Corso et al., 2026; Rescigno & Greystoke, 2026).

To that end, this review pursues four interconnected objectives. The first is to evaluate the biological and structural mechanisms of BRCA1 and BRCA2 variants, detailing how domain-specific alterations disrupt homologous recombination repair, and to characterize the clinical, diagnostic, and ethical challenges posed by the high prevalence of variants of uncertain significance (VUS) in global oncology workflows (Ismail et al., 2024; Rescigno & Greystoke, 2026). The second is to conduct a comparative analytical evaluation of state-of-the-art protein language models — including AlphaMissense, ESM3, and DNABERT-S — and related machine learning architectures, critically assessing their sensitivity, specificity, and computational complexity in predicting BRCA variant pathogenicity (Jurenaite et al., 2024; Corso et al., 2026). The third is to analyze the clinical integration of in silico pathogenicity predictions with multi-omics data, including genomic scars, promoter methylation, transcriptomic signatures, and digital pathology, and to assess how multimodal data fusion optimizes patient risk stratification, HRD detection, and PARP-inhibitor response prediction (Sabit et al., 2026; Shah et al., 2025). The fourth, and in many ways the most forward-looking, is to identify current translational bottlenecks — algorithmic bias, data privacy, computational overhead, and the underrepresentation of diverse ancestries in public genomic databases — and to propose a practical roadmap toward standardization, regulatory qualification, and equitable clinical adoption of AI-driven biomarkers (Lin et al., 2026; Qiu et al., 2026; Rescigno & Greystoke, 2026; Shah et al., 2025).

2. AI and Multi-Omics Convergence in BRCA-Driven Precision Oncology

2.1 The Multi-Omics Paradigm in Oncology: Why One Layer Is Never Enough

Cancer, whatever else it is, is rarely explicable through a single data layer, and BRCA-associated disease is no exception. Genomic sequence alone tells us that a variant exists; it does not, by itself, tell us how that variant behaves once it meets a real tumor's transcriptional, epigenetic, and microenvironmental context (Rescigno & Greystoke, 2026; Shanmugam & Ravikumar, 2026). Single-omics assays — a gene panel here, a methylation array there — capture a partial silhouette of disease biology, missing the downstream protein states, post-translational modifications, and regulatory dynamics that often determine whether a tumor resists therapy (Sabit et al., 2026; Shanmugam & Ravikumar, 2026).

Modern oncology has responded, somewhat unevenly, by pursuing vertical multi-omics integration — correlating genomic, transcriptomic, proteomic, metabolomic, and epigenomic signals within the same patient cohort (Ouhmouk et al., 2025; Shanmugam & Ravikumar, 2026). Done well, this kind of fusion surfaces synergistic mechanisms that no single platform would reveal on its own, and it tends to reduce false-discovery rates relative to single-platform analyses (Sabit et al., 2026; Shanmugam & Ravikumar, 2026). Done poorly, however, it runs headlong into a genuinely hard statistical problem. Multi-omics data is high-dimensional, sparse, and technically messy — a fairly textbook instance of the so-called curse of dimensionality (Ouhmouk et al., 2025). Traditional statistical models tend to buckle under these conditions, undone by multicollinearity and, more often than not, by sample sizes that were never large enough to begin with (Ouhmouk et al., 2025; Jurenaite et al., 2024; Barua et al., 2026). It is largely in response to this failure mode that deep learning has become the default computational strategy — not because it is fashionable, but because it can automatically learn hierarchical feature representations that harmonize disparate data modalities in a way older methods simply cannot (Ouhmouk et al., 2025; Lin et al., 2026; Sabit et al., 2026).

2.2 Computational Foundations: How Deep Learning Architectures Handle Biological Complexity

Deep learning approaches to multi-omics oncology tend to fall, loosely, into three families — generative, non-generative, and hybrid — distinguished mainly by how they fuse heterogeneous inputs (Ouhmouk et al., 2025). Non-generative architectures, graph neural networks (GNNs) and graph attention networks (GATs) among them, map inputs directly onto clinical endpoints and have the useful property of being able to embed prior biological knowledge, such as protein-protein interaction networks, straight into the learning architecture (Ouhmouk et al., 2025). The Multimodal Onco-Graph Convolutional Network (MOGONET) illustrates this well: separate graph convolutional layers learn omics-specific embeddings, and a downstream module captures cross-omics label correlations for multi-class tumor classification (Wang & Ballester, 2021; Ouhmouk et al., 2025).

A more fundamental limitation of many earlier classifiers, though, is their insistence on fixed-dimensionality inputs (Jurenaite et al., 2024). A network trained to expect exactly N genomic loci breaks, sometimes badly, when N changes — and worse, it can become sensitive to the arbitrary order in which those loci are presented (Jurenaite et al., 2024). Jurenaite et al. (2024) addressed this directly with SetQuence and SetOmic, deep networks built on permutation-invariant set representations rather than fixed vectors. SetQuence leans on pretrained DNA language models, such as DNABERT, to encode variably long, variant-associated sequences, which are then aggregated — via pooling multi-head attention — into a single coordinate-free embedding (Jurenaite et al., 2024). SetOmic extends the same logic to quantitative transcriptome counts, and in doing so demonstrates markedly better robustness and less information loss across coding and non-coding regions than fixed-input alternatives (Jurenaite et al., 2024).

Generative models occupy a different, complementary niche. Variational autoencoders (VAEs) — OmiEmbed, CustOmics, and TMO-Net among the better-known examples — learn the joint probability distribution of high-dimensional omics data, compressing it into a shared, low-dimensional latent space via encoder-decoder pairs (Ouhmouk et al., 2025; Zhang et al., 2021). One practical virtue of this family is resilience to missing data: because the latent space is learned jointly, researchers can often reconstruct an unmeasured omics modality directly from what the model has already learned about the others (Ouhmouk et al., 2025).

More recently, transformer-based foundation models have introduced something closer to a generalist medical AI paradigm (Corso et al., 2026). Large language models are increasingly being repurposed for genomic diagnostics, converting raw, unstructured somatic mutation data and clinical phenotypes into structured, dialogue-like representations a model can reason over (Liu et al., 2025; Corso et al., 2026). OncoChat is a clear illustration — a diagnostic LLM developed on 158,836 tumor samples spanning 69 cancer types across 19 institutions, instruction-tuned on structured clinicogenomic text. It outperforms earlier classifiers such as OncoNPC and GDD-ENS and shows notable stability in predicting tissue of origin for cancers of unknown primary (Liu et al., 2025).

2.3 Clinical and Translational Applications in Precision Oncology

Where multi-omics AI has moved beyond proof-of-concept is in several fairly distinct clinical touchpoints, each worth considering on its own terms (Sabit et al., 2026; Lin et al., 2026).

2.3.1 Molecular Risk Stratification and Biomarker Staging

TNM staging, for all its familiarity, is essentially a macro-anatomic system, and it was never designed to capture the biological heterogeneity that increasingly defines how tumors actually behave (Sabit et al., 2026; Roy et al., 2027). Integrated multi-omics classifiers offer a more refined alternative. Roy et al. (2027), for instance, combined microRNA and messenger RNA datasets from the TCGA breast cancer cohort using gradient-boosting algorithms such as XGBoost, and reported a 5–10% improvement in clinical-stage classification accuracy over single-platform classifiers — largely by capturing non-linear interactions between miRNA-mediated silencing and its downstream mRNA targets. In a related but visually distinct approach, whole-slide pathology images can be fused with molecular assays: the AnchorMIL framework reframes Oncotype DX recurrence-score prediction from whole-slide images as a joint regression-and-classification task, achieving strong concordance across both TCGA-derived and independent cohorts (Koyun et al., 2026).

2.3.2 Neoadjuvant Therapy Response Prediction

Predicting pathological complete response (pCR) early in neoadjuvant chemotherapy matters because it lets clinicians spare patients unnecessary toxicity when a regimen is unlikely to work (Sabit et al., 2026; Lash & Valero, 2026). Stacked ensemble frameworks that fuse transcriptomic profiling with clinical variables — capturing distinct inductive biases across tree-based, kernel, and deep-network learners — have produced well-calibrated predictions of chemotherapy responsiveness (Lash & Valero, 2026). Multimodal deep learning that additionally incorporates pre-treatment MRI radiomics and digital pathology tends to do even better, with reported accuracy gains of 10–20% over unimodal baselines (Sabit et al., 2026; Lin et al., 2026).

2.3.3 Synthetic Lethality and Novel Target Discovery

Synthetic lethality remains one of the more powerful conceptual tools for target discovery in oncology, and its clinical proof of concept is, of course, the success of PARP inhibitors in BRCA1/2-mutant disease (Schäffer et al., 2024; Li et al., 2026; Wolde & Belay, 2026). What computational biology has added more recently is a way to look for synthetic-lethal partners beyond the canonical DNA-repair pathway. Frameworks such as DGIB4SL, KR4SL, and MAGICAL fuse multi-omics data with protein-protein interaction networks and knowledge-graph representations, leaning on graph neural networks and information-bottleneck principles to surface higher-order regulatory motifs and, from there, candidate target combinations — pairing PARP inhibitors with novel STING agonists, for instance, to resensitize resistant tumors (Li et al., 2026; Sabit et al., 2026).

2.3.4 AI-Driven Liquid Biopsies and Minimal Residual Disease Monitoring

Non-invasive monitoring — circulating tumor cells, circulating tumor DNA, cell-free DNA fragmentomics, methylation profiling — is quietly transforming how clinicians track disease over time, without repeated invasive biopsies (Sabit et al., 2026; Shanmugam & Ravikumar, 2026). Machine learning applied to multi-analyte liquid biopsy panels can fuse fragment-length density distributions with epigenetic methylation signals to estimate minimal residual disease and monitor treatment response essentially in real time (Sabit et al., 2026; Rescigno & Greystoke, 2026; Lin et al., 2026). Commercial platforms such as Grail Galleri already apply this logic at scale, achieving meaningful early-detection sensitivity across more than 50 distinct cancer types (Shanmugam & Ravikumar, 2026). At the single-cell level, foundation models like scGPT and scBERT extend this further still, annotating heterogeneous tumor-infiltrating immune populations with strong accuracy and tracking drug-resistant subclonal trajectories as they emerge (Ouhmouk et al., 2025).

2.4 Major Translational Roadblocks in Clinical Integration

None of this, however, translates automatically into bedside utility, and it would be misleading to end this review of the literature without naming the obstacles plainly (Ouhmouk et al., 2025; Lin et al., 2026).

2.4.1 Overreliance on Centralized TCGA Data and Technical Artifacts

A large majority of published deep learning models in computational oncology — well over 90%, by some estimates — are trained and validated almost exclusively on The Cancer Genome Atlas or METABRIC (Ouhmouk et al., 2025; Qiu et al., 2026). These are excellent resources, but they are also idealized, retrospective, and largely disconnected from the messiness of real-world electronic health record data or population diversity (Ouhmouk et al., 2025). Worse, deep networks are notoriously sensitive to domain shift and batch effects: Dehkharghanian et al. (2023) showed that models trained on TCGA pathology whole-slide images could predict a slide's acquisition site with up to 86% accuracy, using cues that had nothing to do with underlying biology — tissue staining protocol, scanner model, that sort of thing (Ouhmouk et al., 2025). A model that has quietly learned to recognize which hospital a slide came from is not, whatever its benchmark scores suggest, learning cancer biology.

2.4.2 The "Black Box" Interpretability Gap

Clinicians, understandably, are reluctant to act on predictions they cannot interrogate, and regulators tend to feel much the same way (Ouhmouk et al., 2025; Lin et al., 2026; Sabit et al., 2026). Many of the deep architectures described above remain functionally opaque — accurate, perhaps, but not explicable in any way a treating physician could defend in a tumor board discussion (Ouhmouk et al., 2025; Lin et al., 2026). Post-hoc explainability tools, SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) chief among them, are increasingly being folded into multi-omics pipelines like SetOmic, and the results are at least somewhat reassuring: the features these tools flag as most influential tend to align with established cancer biology, including known attributions to PIK3CA and S100A11 in breast cancer cohorts (Jurenaite et al., 2024; Ouhmouk et al., 2025). Biologically constrained architectures such as DeepOmix and DeepKEGG take a related but more structural approach, embedding pathway and gene-ontology layers directly into the network to keep its reasoning at least loosely tethered to known biology (Ouhmouk et al., 2025).

2.4.3 Data Harmonization, Batch Effects, and Algorithmic Bias

Combining data generated across different platforms, laboratories, and protocols introduces systematic technical variation — batch effects — that, left uncorrected, can masquerade as biological signal (Ouhmouk et al., 2025; Jurenaite et al., 2024). Careful data cleaning and batch correction, whether through the ComBat algorithm or self-normalizing neural network architectures, are not optional extras here; they are close to a prerequisite for trustworthy results (Ouhmouk et al., 2025; Yin et al., 2026).

There is also a harder, more uncomfortable problem underneath the technical one. Genetic ancestry audits suggest that more than 82% of TCGA cases represent individuals of European descent, leaving African and Indigenous populations substantially underrepresented (Ouhmouk et al., 2025). Models trained on that skewed foundation tend, unsurprisingly, to degrade in performance — and to misclassify more often — when applied outside the population they were trained on (Ouhmouk et al., 2025; Lin et al., 2026). Addressing this will likely require more than good intentions: prospective recruitment of genuinely diverse cohorts, subgroup-stratified fairness auditing, and adversarial debiasing or reweighting strategies built into training from the outset (Ouhmouk et al., 2025; Lin et al., 2026).

2.5 Section Summary

Taken as a whole, the literature reviewed here converges on a fairly consistent picture. Multi-omics deep learning, and protein language models specifically, have genuinely outperformed earlier computational approaches across tumor classification, risk stratification, synthetic-lethality discovery, and liquid-biopsy monitoring (Ouhmouk et al., 2025; Sabit et al., 2026; Corso et al., 2026). What remains unresolved — data homogeneity, interpretability, and demographic bias, primarily — is not a minor footnote to that progress but arguably the central story of where the field goes next (Ouhmouk et al., 2025; Lin et al., 2026). Figures 1 through 4 summarize, respectively, the BRCA-HRR-PARP synthetic lethality axis that motivates this review, the sequence-to-pathogenicity inference pipeline characteristic of protein language models, the study identification workflow used to compile this review's evidence base, and the broader multi-omics fusion architecture that underlies much of the translational work discussed above

3. Methods

This review followed a structured, reproducible acoxuflev opebkavln methodology, informed by PRISMA principles for transparent reporting, so that the search strategy, eligibility criteria, and hchs xadvvzdlzc process could, in principle, be replicated by an independent team using the search terms and inclusion criteria specified below (Figure 3).

3.1 Search Strategy and Information Sources

We searched PubMed/MEDLINE, Scopus, Web of Science, and IEEE Xplore for peer-reviewed literature, supplemented by manual screening of preprint servers and the reference lists of retrieved articles to capture recent, rapidly evolving computational work not yet indexed in major databases. Search terms combined controlled vocabulary and free-text keywords across three conceptual clusters, connected with Boolean operators: (1) gene/disease terms — "BRCA1" OR "BRCA2" OR "hereditary breast and ovarian cancer" OR "homologous recombination deficiency"; (2) computational-model terms — "protein language model" OR "deep learning" OR "variant pathogenicity prediction" OR "AlphaMissense" OR "ESM" OR "DNABERT" OR "multi-omics"; and (3) clinical-translation terms — "precision oncology" OR "PARP inhibitor" OR "variant of uncertain significance" OR "clinical decision support". No lower date limit was imposed, though the search prioritized literature published within the preceding five years to reflect the fast-moving nature of protein language modeling.

3.2 Eligibility Criteria

Studies were eligible for inclusion if they (a) were published in a peer-reviewed journal or as a full conference proceeding; (b) addressed protein or nucleotide language models, deep learning architectures, or multi-omics integration frameworks applied to cancer genomics, variant interpretation, or precision oncology; and (c) reported extractable methodological detail sufficient to characterize the model architecture, input modalities, and at least one quantitative performance metric. Studies were excluded if they were editorials, conference abstracts without full methodological reporting, non-English-language publications without an available translation, or focused exclusively on cancer types or molecular contexts unrelated to BRCA biology, homologous recombination, or the broader multi-omics oncology landscape addressed by this review.

3.3 Study Selection and Screening Process

Following automated deduplication of search results, two conceptual screening stages were applied: an initial title/abstract screen against the eligibility criteria above, followed by full-text review of the remaining records. Articles progressing through both stages contributed either to the qualitative narrative synthesis or, where sufficient quantitative detail was available, to the comparative benchmark tables (Tables 1–4). This selection pathway is summarized in Figure 3, which documents the number of records identified, screened, and retained at each stage, consistent with standard systematic-review reporting conventions.

3.4 Data Extraction

For each included study, we extracted, where reported: (i) model or framework name and underlying architectural class (e.g., transformer, graph neural network, variational autoencoder); (ii) input data modalities (sequence, expression, methylation, imaging, or combinations thereof); (iii) the clinical or biological prediction task addressed; (iv) validation cohort characteristics and sample size, where disclosed; (v) reported performance metrics (accuracy, AUC, F1-score, precision/recall, or concordance statistics, as reported by the original authors); and (vi) stated limitations or translational barriers noted by the original authors. Extracted data were tabulated by two independent reviewers' logic checks against the source text to minimize transcription error, and discrepancies were resolved by re-consulting the original publication.

3.5 Data Synthesis Approach

Given the methodological heterogeneity across included studies — spanning variant-level pathogenicity classifiers, tumor-type classifiers, and multimodal risk-prediction frameworks — a quantitative meta-analysis (e.g., pooled effect size) was not appropriate. We therefore synthesized findings narratively, organized thematically around four domains reflecting recurring patterns in the literature: representation learning and multimodal data fusion; clinical risk stratification and treatment-response prediction; synthetic-lethality and drug-target discovery; and translational/implementation barriers. This thematic structure mirrors, and is reported in, the Results section below and is intended to allow readers to trace each reported metric back to its source study via the accompanying tables.

3.6 Quality Considerations and Limitations of the Method

As a narrative synthesis rather than a formal systematic review with dual-independent screening and formal risk-of

Table 1. Multimodal data-integration frameworks and deep learning architectures applied to cancer genomics. This table summarizes transformer-, graph-, and autoencoder-based deep learning architectures designed for high-dimensional, sparse, and heterogeneous multi-omics cancer datasets, including each framework's input modalities, primary clinical use-case, key computational advantages and limitations, and reported performance benchmarks with source citations. Entries are ordered to progress from sequence-only models toward increasingly multimodal fusion architectures.

Table 2. Biomarker discovery, prognostic, and neoadjuvant treatment-response prediction frameworks. This table catalogs computational tools developed to identify clinically actionable prognostic signatures and forecast therapeutic outcomes, spanning digital pathology, radiomics, transcriptomic, and knowledge-graph-based approaches, together with their validation cohorts, translational outputs, and quantitative performance metrics as reported in the source publications.

-bias scoring, this methodology is subject to selection and reporting biases inherent to the underlying primary literature; performance metrics reported here reflect authors' own validation procedures and were not independently recomputed or benchmarked on a common dataset. This limitation is addressed further in the Discussion.

4. Benchmarking Protein Language Models and Multi-Omics Paradigms in Precision Oncology

This section synthesizes the empirical outcomes, classification benchmarks, and translational milestones identified through the search and screening process described above (Figure 3), organized across four clinical-computational touchpoints reported in Tables 1 through 4.

4.1 Deep Representation Learning and Multimodal Data Fusion 

Statistical frameworks in oncology have historically struggled to integrate biological layers because of dimensionality and sparsity (Ouhmouk et al., 2025). Deep neural networks that treat biological sequences as text corpora have meaningfully mitigated this limitation. A primary benchmark here is SetQuence, an attention-based set transformer designed to process somatic variant-associated sequences as non-fixed, permutation-invariant sets (Jurenaite et al., 2024). Evaluated on the COSMIC database across 32 tumor classes, SetQuence achieved a macro-averaged precision of 0.502, recall of 0.402, and classification accuracy of 0.567, meaningfully outperforming random-forest and standard neural-network baselines (Jurenaite et al., 2024; Table 1). These figures confirm that sequence-level features drawn from DNA language models capture physical, chemical, and evolutionary constraints that simpler one-hot or gene-ID encodings tend to lose (Jurenaite et al., 2024).

Extending this architecture, SetOmic integrated sequence embeddings with bulk RNA-sequencing expression matrices, achieving a macro-averaged precision of 0.945 and accuracy of 0.950 in pan-cancer classification — outperforming OmiEmbed's 0.942 accuracy, a variational autoencoder restricted to fixed-dimensionality expression vectors (Zhang et al., 2021; Jurenaite et al., 2024; Table 1). Confusion-matrix analysis further showed that SetOmic sharply reduced misclassification between closely related tumor subtypes, such as breast carcinoma and uterine carcinosarcoma, by modeling long-range genomic-transcriptomic dependencies (Jurenaite et al., 2024). Graph-based approaches such as MOGONET, which fuses mRNA, DNA methylation, and microRNA profiles via a View Correlation Discovery Network, consistently outperformed single-modality models across independent cohorts, reinforcing the broader pattern that vertical data fusion uncovers synergistic signal invisible to single-assay pipelines (Wang & Ballester, 2021; Ouhmouk et al., 2025; Table 1).

4.2 Clinical Diagnostics, Risk Stratification, and Neoadjuvant Therapy Response 

Across the included literature, deep learning models have demonstrated clinically meaningful precision in forecasting treatment response and disease recurrence from multimodal inputs. In predicting pathological complete response to neoadjuvant chemotherapy, a 17-gene stacked ensemble classifier — combining random forest, gradient boosting, support vector machine, k-nearest neighbors, and neural network base learners through a stacked elastic-net meta-learner — extracted a signature of 11 upregulated genes (including BRCA2, MKI67, and LYZ) and 6 downregulated genes distinguishing responders from non-responders (Lash & Valero, 2026). Externally validated with isotonic-regression calibration, the ensemble achieved a ROC-AUC of 0.78, balanced accuracy of 0.71, and sensitivity of 0.86, with DeLong's test confirming statistically significant superiority (p < .05) over simple voting ensembles and individual base classifiers (Lash & Valero, 2026; Table 2).

Multimodal frameworks that fuse digital histology with genomic and clinical features have performed comparably well: an integrated model combining H&E histology, bulk genomics, and clinical variables achieved an AUC of 0.87 for pathological complete response prediction, translating into a 10–20% accuracy gain over single-modality approaches (Sammut et al., 2022; Sabit et al., 2026; Table 2). AnchorMIL, a multiple-instance-learning framework applied to whole-slide images, achieved AUC values of 0.885 on TCGA-BRCA and 0.858 on an independent cohort for Oncotype DX recurrence-score estimation, outperforming conventional attention-based MIL baselines while supporting both continuous score prediction and binary risk stratification (Koyun et al., 2026; Su et al., 2024; Table 2).

4.3 In Silico Drug Discovery and Synthetic Lethality Target Screening

A recurring theme among the reviewed synthetic-lethality

Figure 1. Mechanistic basis of PARP-inhibitor synthetic lethality in BRCA1/2-deficient tumors. Schematic comparison of homologous recombination (HR)-proficient and HR-deficient tumor cells. In HR-proficient cells (left), wild-type BRCA1/BRCA2 supports accurate double-strand-break repair and genomic stability. In HR-deficient cells (right), a pathogenic BRCA1/BRCA2 variant abolishes homologous recombination, forcing reliance on error-prone repair pathways; PARP-inhibitor exposure then traps PARP1/2 on DNA, converting single-strand lesions into lethal double-strand breaks and driving selective tumor-cell apoptosis. A protein language model-derived pathogenicity score (bottom) links variant calling to this clinical decision pathway.

Figure 2. Sequence-to-pathogenicity inference pipeline used by protein language models. Conceptual workflow by which a raw BRCA1/BRCA2 variant call is encoded as an amino-acid sequence, embedded by a pretrained protein language model (e.g., AlphaMissense, ESM3, DNABERT-S), and mapped onto a latent evolutionary-constraint representation that yields a discrete pathogenicity classification. Outputs from this pipeline feed downstream clinical applications, including homologous recombination deficiency scoring and PARP-inhibitor eligibility assessment.

Figure 3. Study identification and selection workflow for this narrative review. Flow diagram summarizing the reproducible literature search and screening process described in Section 3 (Methods), from initial database identification through duplicate removal, title/abstract screening, full-text eligibility assessment, and final inclusion in qualitative synthesis and the quantitative comparison tables (Tables 1–4). Reasons for exclusion at each stage are indicated alongside the corresponding decision node.

Figure 4. Vertical multi-omics data-fusion architecture underlying translational precision-oncology models. Schematic representation of how genomic, transcriptomic, epigenomic, and digital-pathology data layers are integrated by a deep-learning fusion module (e.g., graph neural network, variational autoencoder, or set-transformer architecture) to generate an integrated patient-level risk and pathogenicity score, which in turn informs clinical decision support around homologous recombination deficiency status, PARP-inhibitor eligibility, and multidisciplinary tumor-board discussion.

frameworks is their reliance on graph-structured biological knowledge to move beyond the canonical BRCA-PARP axis. DGIB4SL, built on a diverse graph information bottleneck, extracted 13 distinct motif-based adjacency matrices from resources such as SynLethDB and outperformed baseline classifiers on both recall and precision, while additionally producing multiple, biologically interpretable explanations for each predicted interaction (Li et al., 2026; Table 3). MAGICAL, a multi-class random-forest framework built on protein-protein interaction network topology, achieved 80% accuracy on its discovery cohort and a validation AUC of 0.83 on the independent DepMap CRISPR dependency dataset, successfully discriminating synthetic-lethal from synthetic-viable gene pairs (Li et al., 2026; Table 3).

Complementary architectures address specific weaknesses in this space: NSF4SL reformulates synthetic-lethality prediction as a partner-gene ranking problem to avoid reliance on unverified negative training examples, demonstrating stronger generalization to previously unobserved genes, while KR4SL applies recurrent path-encoding over biomedical knowledge graphs to surface clinically interpretable explanatory subgraphs, including non-canonical double-strand-break repair pathways relevant to BRCA2-deficient, PARP-inhibitor-resistant tumors (Li et al., 2026; Table 3). Literature-mining platforms such as PandaOmics extend this pipeline further, combining pathway scoring with large-language-model-based literature synthesis to prioritize candidate compounds, including a selective MYT1 inhibitor with notable preclinical activity in gynecological and breast cancer models (Barua et al., 2026; Table 3).

4.4 Liquid Biopsy, Clinical Translation, and Algorithmic Roadblocks 

Non-invasive monitoring technologies represent one of the more clinically proximate applications reviewed here. Machine learning models applied to cell-free DNA methylation patterns, exemplified by the Grail Galleri platform, have demonstrated early-detection capability across more than 50 distinct cancer types (Shanmugam & Ravikumar, 2026; Table 4). Deep learning models analyzing cfDNA fragmentomics have achieved over 90% accuracy distinguishing malignant from benign lesions, supporting minimally invasive minimal residual disease monitoring (Lin et al., 2026; Rescigno & Greystoke, 2026; Table 4). At the single-cell level, foundation models such as scGPT and scBERT have annotated heterogeneous tumor-infiltrating immune populations with F1-scores around 0.923, while additionally characterizing drug-resistant subclonal trajectories in near-real time (Ouhmouk et al., 2025; Table 4).

These successes, however, coexist with the translational obstacles catalogued in Table 4: heavy reliance on TCGA-derived training data and its associated technical artifacts, persistent interpretability gaps in complex architectures, and reduced generalizability when models trained on demographically homogeneous cohorts are deployed in more diverse clinical populations (Dehkharghanian et al., 2023; Ouhmouk et al., 2025; Lin et al., 2026; Wolde & Belay, 2026; Table 4). Addressing the latter, in particular, is increasingly framed in the literature as requiring algorithmic auditing, prospective multi-institutional "silent trials," and privacy-preserving federated learning architectures capable of training across decentralized, internationally diverse datasets without exposing individual patient records (Ouhmouk et al., 2025; Table 4).

5. From Computational Benchmark to Clinical Confidence — Closing the BRCA Variant Interpretation Gap

5.1 Interpreting the Evidence: What the Numbers Actually Suggest

Pulling back from the individual benchmarks summarized above, a fairly coherent narrative emerges. Protein language models and their multi-omics extensions do not merely match earlier machine learning approaches to variant interpretation — in study after study, they outperform them, often by a wide margin, and they do so specifically on the kinds of high-dimensional, sparse, sequence-heavy data that older architectures were never well suited to handle (Jurenaite et al., 2024; Ouhmouk et al., 2025; Table 1). That the largest accuracy gains cluster around models capable of ingesting raw sequence context — rather than hand-engineered gene-level features — is, we think, not a coincidence; it is fairly direct evidence that evolutionary and structural constraint, learned implicitly from sequence, carries real predictive signal about pathogenicity that classical feature engineering was leaving on the table (Corso et al., 2026; Jurenaite et al., 2024).

At the same time, it would be a mistake to read these benchmark numbers as settled clinical fact. Nearly every metric reported in Tables 1 through 4 was generated under retrospective, single-institution, or heavily curated

Table 3. AI-driven drug discovery, synthetic lethality, and target prioritization frameworks. This table outlines deep learning and knowledge-graph platforms designed to predict drug-target interactions, identify novel synthetic lethal gene pairs beyond canonical BRCA-PARP biology, and model cellular treatment responses, including each framework's core methodology, experimental validation status, and principal advantages as documented in the cited literature.

Table 4. Liquid-biopsy applications and translational roadblocks in clinical AI deployment. This table summarizes non-invasive liquid-biopsy and single-cell monitoring applications alongside the major technical and ethical roadblocks — including TCGA data overreliance, algorithmic interpretability gaps, and population generalizability limitations — that currently constrain the clinical translation of AI-driven precision-oncology tools, together with proposed mitigation strategies drawn from the reviewed literature.

conditions (Ouhmouk et al., 2025; Table 1; Table 2). AUC values above 0.90 are genuinely impressive, but they describe performance within a data distribution the model has already seen a great deal of, not necessarily performance on the next VUS that lands on a clinician's desk from an underrepresented population or an unusual tumor subtype (Dehkharghanian et al., 2023; Ouhmouk et al., 2025).

5.2 BRCA Variant Interpretation Specifically: Where pLMs Genuinely Change the Calculus

For the specific clinical problem motivating this review — the roughly 40% of BRCA missense variants still sitting in VUS limbo — protein language models offer something traditional approaches largely cannot: near-instantaneous, proteome-scale scoring that does not require a bespoke functional assay for every novel variant encountered (Rescigno & Greystoke, 2026; Corso et al., 2026). AlphaMissense-style models, in principle, can pre-screen a newly detected BRCT-domain missense variant the moment it is called, flagging it for prioritized functional workup or, in higher-confidence cases, informing an interim clinical conversation well before a multigene functional assay could ever be completed (Corso et al., 2026; Ismail et al., 2024). Combined with genomic-scar-based HRD detection tools such as HRProfiler, this creates a plausible, if not yet fully validated, pipeline running from raw sequence to PARP-inhibitor eligibility (Shah et al., 2025; Figure 1; Figure 2).

That said, the honest caveat here is that BRCA-specific prospective validation of these general-purpose pLMs remains comparatively thin in the literature we reviewed, relative to their pan-cancer classification benchmarks (Table 1; Table 3). Much of the strongest quantitative evidence assembled in this review — SetOmic's 0.950 accuracy, MOGONET's cross-cohort superiority, OncoChat's tumor-of-origin performance — comes from tasks adjacent to, rather than identical with, single-variant BRCA pathogenicity classification (Jurenaite et al., 2024; Wang & Ballester, 2021; Liu et al., 2025). We would argue this is an important distinction for clinicians reading this literature to keep in mind, rather than a reason for skepticism outright.

5.3 The Interpretability Problem Is Not Just an Engineering Inconvenience

We think it is worth stating plainly that the "black box" critique raised repeatedly in the literature (Section 2.4.2) is not a peripheral concern that can be resolved later, once performance is good enough. For a VUS reclassification decision that may lead to prophylactic surgery on one hand, or withheld PARP-inhibitor therapy on the other, a clinician reasonably wants to know why a model called a variant likely pathogenic, not merely that it did so with high confidence (Ouhmouk et al., 2025; Sabit et al., 2026). The SHAP- and LIME-based attribution work reviewed here is encouraging — attributions do appear to track known cancer biology in several studies (Jurenaite et al., 2024; Ouhmouk et al., 2025) — but post-hoc explanation is not the same thing as mechanistic transparency, and regulatory bodies evaluating these tools as software-as-a-medical-device will likely draw that distinction carefully.

5.4 Equity, Data Provenance, and the Limits of Benchmark Generalization

Perhaps the most consequential limitation surfaced in this review is not technical at all, but demographic. With more than four-fifths of TCGA cases representing individuals of European ancestry (Ouhmouk et al., 2025), and with pLM training corpora inheriting broadly similar biases from public sequence databases, there is a real and, we think, underappreciated risk that a model performing beautifully on benchmark cohorts will underperform — silently, without any obvious warning sign — on patients from underrepresented populations (Ouhmouk et al., 2025; Lin et al., 2026; Wolde & Belay, 2026). Given that BRCA variant spectra themselves differ meaningfully by ancestry, this is not a hypothetical concern; it bears directly on which patients ultimately gain, or fail to gain, access to PARP-inhibitor therapy on the basis of an algorithmic call.

5.5 Limitations of This Review

This review has its own limitations, worth stating candidly. As a narrative rather than fully systematic synthesis, it did not apply formal dual-independent risk-of-bias scoring to every included study, and the performance metrics reported in Tables 1 through 4 reflect each original study's own validation methodology rather than a harmonized, independently re-computed benchmark (Section 3.6). Several of the frameworks discussed — particularly the synthetic-lethality and multi-omics fusion tools summarized in Table 3 — remain at a discovery or retrospective-validation stage rather than prospective clinical deployment, and readers should weigh the translational claims made here accordingly.

5.6 A Practical Path Forward

Taken together, we would suggest the field's next steps are reasonably clear, even if not easy: prospective, multi-institutional validation of pLM-based BRCA pathogenicity scores against orthogonal functional assays; deliberate, funded recruitment of ancestrally diverse cohorts rather than passive reliance on existing repositories; wider adoption of federated learning architectures that allow model training across institutions without centralizing sensitive patient data (Ouhmouk et al., 2025); and continued investment in interpretability tooling that clinicians, not just machine learning engineers, find genuinely usable at the point of care (Sabit et al., 2026; Lin et al., 2026). None of this is a small undertaking. But given how directly variant interpretation already determines who receives a life-extending PARP inhibitor and who does not, it is difficult to argue the effort is optional.

6. Conclusion

Protein language models have progressed rapidly from computational novelty to genuine contenders for resolving a major precision-oncology bottleneck: the roughly 40% of BRCA1/BRCA2 missense variants still classified as uncertain. Architectures such as AlphaMissense, ESM3, DNABERT-S, and multi-omics extensions like SetOmic and MOGONET consistently outperform earlier machine learning methods, often exceeding 90% accuracy on tumor classification and risk-stratification tasks. Combined with genomic-scar-based HRD detection and PARP-inhibitor eligibility screening, these tools suggest a plausible pipeline from raw sequence to clinical guidance. That promise remains only partly realized. Evidence relies heavily on retrospective, demographically narrow TCGA cohorts; interpretability has not yet reached clinical or regulatory standards for high-stakes decisions like prophylactic surgery; and BRCA-specific prospective validation remains limited compared to pan-cancer benchmarks. Realizing full clinical potential will require multi-institutional prospective validation, correction of ancestral data bias, and clinician-oriented interpretability tools. Until then, these models function best as triage and prioritization aids—not replacements for careful, orthogonal variant classification.

References