The review searched PubMed, Scopus, Web of Science, bioRxiv, arXiv and Semantic Scholar for English-language work published from January 2020 to July 2026. Two reviewers independently screened titles and abstracts, with disagreements resolved by a third reviewer, and full texts were assessed for relevance. Studies were grouped as BMP4-specific, directed at BMP receptors or shared pathway components, or based on artificial-intelligence methods developed for other targets and presented as potentially transferable. Because this was a narrative review rather than a quantitative meta-analysis, its findings summarize the direction and quality of the literature rather than estimate a pooled effect. Relevant preprints were included and identified as non-peer-reviewed.

BMP4 belongs to the transforming growth factor-beta superfamily and acts as a secreted morphogen. Its mature dimer binds a receptor assembly containing two type I and two type II serine/threonine kinase receptors, initiating SMAD-dependent signaling as well as noncanonical mitogen-activated protein kinase pathways. The effects of this signaling are highly context dependent: BMP4 participates in mesoderm specification, neurogenesis and limb development, but it can also promote epithelial-mesenchymal transition and tumor progression, restrain pulmonary fibrosis, or drive catastrophic heterotopic ossification when the type I receptor ACVR1 carries a gain-of-function mutation. The review presents this functional duality as a central challenge for drug development, because a treatment that blocks BMP4 activity in one tissue could interfere with protective signaling in another.

Structure prediction has produced some of the most visible advances. An AlphaFold2 model reproduced the canonical fold of human BMP4, including its central cysteine-knot core and the two extended fingers that present receptor-binding surfaces. AlphaFold3 was used in 2025 to model BMP4 prodomain phosphorylation and predict how phosphorylation at particular serine residues might change the prodomain’s conformation and accessibility to furin cleavage. That work was paired with experimental mutagenesis, but the review stresses that most AlphaFold3-derived BMP4 and receptor-complex models remain predictions of likely static conformations rather than experimentally established mechanisms. Per-residue and pairwise confidence scores do not guarantee accuracy in flexible loops, disordered regions or post-translational states, and current neural-network predictors do not reproduce the full conformational ensembles underlying receptor binding, prodomain displacement and antagonist engagement.

Experimental structures remain essential because they supply the templates against which computational models can be tested. A 2024 study used X-ray crystallography to determine how the secreted antagonist Twisted Gastrulation 1 blocks the GDF5–BMPR1A interface, preventing type I receptor recruitment while allowing type II receptor engagement. The review emphasizes that this was an experimentally determined structure, not an artificial-intelligence prediction. Molecular dynamics simulations can complement static models by testing whether predicted ligand poses remain stable, as in a 100-nanosecond simulation used to refine retinoic acid and torularhodin docking poses for BMP4. More ambitious tools—including RoseTTAFold-All-Atom, ESMFold, Chai-1 and Boltz-1—could support analysis of BMP4 complexes, engineered variants and small-molecule binding, but the review reports that they have not yet been systematically applied to specific BMP4 questions.

Drug discovery for BMP4 remains less mature than the headlines generated by artificial intelligence might suggest. A 2023 screen combined gene-expression profiling, structure-based docking and molecular dynamics to nominate retinoic acid and torularhodin as possible BMP4 activators, but it used conventional computational-chemistry methods rather than machine learning. A separate FOP campaign screened about 140,000 compounds and synthesized 1,266 candidates for SMAD1 phosphorylation assays, identifying RK783 as a selective BMP type I receptor kinase inhibitor. That effort likewise combined computational prescreening with biochemical testing rather than deep learning. The review says these studies show how computation can narrow chemical space, yet it also draws a firm boundary around present claims: no BMP4-specific small molecule has been identified by AI-driven screening, and no such candidate has entered clinical trials.

More explicitly AI-based methods have produced transferable results elsewhere. A transformer-based platform screened 40 billion molecules against LRRK2, while a ULK1 study found that a Naive Bayes classifier using ECFP6 fingerprints outperformed deep learning when the training set contained only about 200 active compounds. Neuro-symbolic systems, three-dimensional diffusion models and GPT-4o-assisted lead generation have also produced promising results on other targets. The small number of known BMP4 modulators makes the second finding particularly relevant: random forests, support vector machines and Bayesian classifiers may be more suitable than data-hungry deep networks when only limited actives are available. None of these advanced methods, however, has yet been applied successfully to BMP4 itself, so their performance on other molecular targets represents an opportunity rather than a demonstrated BMP4 result.

Single-cell and multi-omic analyses provide the most substantial BMP4-specific evidence described in the review. In pancreatic adenocarcinoma, integrated single-cell and bulk RNA sequencing found BMP4 enriched in cancer cells and cancer stem cells, where it was associated mainly with metabolic rather than canonical BMP–SMAD regulation. In 17 patient-derived glioblastoma cultures, RNA-velocity analysis found that BMP4-sensitive cultures lost pro-inflammatory and neuronal subpopulations and gained a quiescent state with increased BMP/Hippo pathway activity; OLIG1/2 expression was the strongest predictor of sensitivity, while ribosomal translation genes served as early response markers. A separate state-transition model fitted lineage-tracing data and predicted how BMP4 shifts equilibria between stem-like and invasive-like glioblastoma states. A 2025 pipeline combined single-nucleus RNA sequencing, bulk-data deconvolution, phosphoproteomics and Boolean-network simulation to predict cell-state transitions under perturbation.

Retrospective clinical models provide promising but unproven prognostic signals. In pancreatic adenocarcinoma, high BMP4 expression was associated with poor overall, disease-free and progression-free survival, and a nomogram combining BMP4 with sex, age and TNM stage achieved C-index values of 0.738 in training and 0.724 in validation. Separate work reported a random-forest diagnostic AUC of 0.889 and used DeepSurv for survival prediction, but it did not isolate the contribution of BMP4 and generated treatment recommendations that still require prospective assessment. In gastric cancer, a five-gene hypoxia–immune panel containing BMP4 was associated with reduced natural-killer-cell and CD8-positive T-cell infiltration, greater M2 macrophage polarization and higher scores for predicted resistance to immune-checkpoint blockade. These retrospective associations do not establish that a BMP4-guided test improves outcomes, and the independent contribution of BMP4 within many multi-gene panels remains uncertain.

The review proposes a five-year research agenda centered on better data, closed experimental loops and multimodal modeling. Priorities include de novo BMP4 protein design with RFdiffusion and ProteinMPNN, ultra-large virtual screening, active learning for mutation and compound selection, and unified models that combine sequence, structure, cell type, imaging and clinical data. Proposed protein designs include stabilized BMP4 variants, receptor-specific antagonists and conditionally activated prodrugs, while multi-ligand network models could examine how BMP4 and other ligands contribute to FOP. These are research objectives, not existing therapies. The authors also caution that training data dominated by European ancestry, opaque model architectures, off-target biology and unreported validation can make clinical artificial intelligence unreliable. The central conclusion is therefore measured: tools now exist to formulate sharper hypotheses about BMP4, but computational predictions must be checked through structural experiments, biochemical assays, mechanistic models and prospective clinical studies before they can guide treatment.

Subject of Research: Artificial intelligence applications in BMP4 drug discovery, structural modeling and clinical prognostication

Article Title: Bone morphogenetic protein 4 meets artificial intelligence: a comprehensive review of AI-driven advances in drug discovery, structural modeling, and clinical translation

Article References: Sun, W., Huo, T., Ye, Z., Wei, W., Sun, Y., & Cai, H. (2026). Bone morphogenetic protein 4 meets artificial intelligence: a comprehensive review of AI-driven advances in drug discovery, structural modeling, and clinical translation. Frontiers in Molecular Biosciences, 13, Article 1930582. https://doi.org/10.3389/fmolb.2026.1930582

Image Credits: AI Generated

DOI: 10.3389/fmolb.2026.1930582

Keywords: BMP4, artificial intelligence, machine learning, protein structure prediction, drug discovery, single-cell transcriptomics, multi-omics, cancer prognostics, fibrodysplasia ossificans progressiva, fibrosis, Bone, morphogenetic