Proteins serve as the fundamental workhorses of the biological system, orchestrating myriad cellular processes including tissue growth, metabolic regulation, and immune defense. However, the functionality of proteins is not solely determined by their amino acid sequence; it is extensively modulated by chemical modifications introduced after the protein has been synthesized. These modifications, collectively known as post-translational modifications, involve the covalent attachment of various chemical groups such as phosphates, sugars, or acetyl groups. These PTMs finely tune protein activity, stability, localization, and interactions, thereby dictating the broader cellular response and health outcomes.

PTMs represent critical regulatory nodes within the proteome, directing signaling pathways and cellular machinery in both normal and pathological states. Dysfunctional PTMs have been directly implicated in the etiology of numerous complex diseases, including malignancies, cardiovascular conditions, and degenerative neurological disorders. A mutation in the DNA sequence can disrupt normal PTM patterns by abolishing a modification site, creating ectopic sites, or perturbing the surrounding amino acid environment, thereby derailing protein function and precipitating disease. Therefore, precisely pinpointing PTM sites and understanding mutation-driven alterations are paramount to elucidating disease mechanisms.

Addressing this challenge, the Baylor research team led by Dr. Bing Zhang developed DeepMVP—a deep learning framework meticulously trained to identify PTM sites across the human proteome and predict how mutations reshape these sites. The model was constructed using a novel dataset named PTMAtlas, which represents a comprehensive and rigorously curated collection of 397,524 verified PTM sites derived from the systematic reanalysis of 241 publicly available proteomic datasets. Focusing on six prevalent PTM types, including phosphorylation and glycosylation, PTMAtlas provides a densely annotated resource that dramatically surpasses existing databases in both breadth and accuracy.

DeepMVP’s architecture leverages modern deep neural networks capable of discerning subtle sequence patterns indicative of PTM sites, integrating contextual biochemical properties to enhance predictive power. This approach enables not only precise site identification but also the assessment of how specific amino acid substitutions may enhance or diminish PTM occurrence. The model’s flexibility extends to non-human proteins, effectively predicting PTM sites in viral proteins such as those from the SARS-CoV-2 virus, highlighting its wide utility across biomedical research domains.

Benchmarking DeepMVP against eight state-of-the-art computational tools revealed a clear superiority in performance. Evaluation on a curated set of 235 experimentally validated mutation-PTM pairs demonstrated an impressive 81% accuracy in pinpointing exact PTM sites. More strikingly, DeepMVP correctly predicted the directional change—increase or decrease—of PTM levels caused by mutations in 97% of the cases. These results underscore DeepMVP’s effectiveness in interpreting the functional repercussions of genetic variation at the post-translational level.

The implications of DeepMVP’s predictive capabilities extend far beyond academic interest. By enabling a high-resolution view of how mutations perturb PTM landscapes, this tool offers a powerful platform for the identification of novel therapeutic targets and the design of precision medicine approaches. For example, in cancer biology, understanding aberrant PTM patterns linked to oncogenic mutations may drive the development of targeted inhibitors that restore normal cellular signaling. Similarly, in neurological and cardiovascular diseases, identifying mutation-induced PTM changes could illuminate pathophysiological processes hitherto obscured in genetic studies.

DeepMVP is freely accessible to the global research community, fostering collaborative efforts to exploit its potential across various health disciplines. This open-access availability ensures that scientists investigating disease genetics, drug discovery, and molecular biology can integrate DeepMVP predictions into their workflows, accelerating the translation of genetic insights into tangible clinical interventions.

Complementing the AI model, PTMAtlas stands as a monumental achievement, synthesizing extensive proteomic data into one unified framework. Its creation involved the harmonization of heterogeneous datasets, rigorous quality control measures, and sophisticated bioinformatic pipelines. This assembly provides an unprecedented foundation for future studies in proteomics, molecular evolution, and systems biology, enabling researchers to navigate the complexity of protein modifications with newfound clarity.

The Baylor team acknowledges significant support from various funding bodies, including the National Cancer Institute (NCI) and the Cancer Prevention and Research Institutes of Texas, underscoring the critical role of sustained investment in biomedical innovation. Additionally, computational resources such as the NVIDIA Titan Xp GPU facilitated the model’s training and optimization, reflecting the increasingly interdisciplinary nature of modern bioscience combining biology, computer science, and engineering.

Looking ahead, the researchers envision expanding DeepMVP’s capabilities to encompass additional PTM types and incorporating structural protein information to further refine predictions. Coupled with advances in high-throughput proteomics and functional genomics, such enhancements could revolutionize our capacity to decode the molecular underpinnings of human diseases.

In summary, the deployment of DeepMVP marks a seminal leap in the application of AI to biomedical research, offering a powerful avenue to decode the molecular grammar that links genetic variation to functional protein changes. This work not only deepens our understanding of cellular regulation at the molecular level but also propels the potential for innovative therapeutic strategies targeting post-translational modifications, thus opening new frontiers in precision medicine.

Subject of Research: People

Article Title: DeepMVP: deep learning models trained on high-quality data accurately predict PTM sites and variant-induced alterations

News Publication Date: 26-Aug-2025

Web References: https://www.nature.com/articles/s41592-025-02797-x

References:
Zhang, B., Wang, C., Wen, B., Li, K., Han, P., Holt, M. V., Savage, S. R., Lei, J. T., Dou, Y., Shi, Z., & Li, Y. DeepMVP: deep learning models trained on high-quality data accurately predict PTM sites and variant-induced alterations. Nature Methods, 26 August 2025. DOI: 10.1038/s41592-025-02797-x

Image Credits: Baylor College of Medicine

Keywords: Applied sciences and engineering, Applied mathematics, Computer science, Health and medicine

The innovative AI tool developed by the CRG researchers is adept at creating DNA regulatory sequences that are not naturally occurring. This tool allows scientists to specify criteria for DNA fragments, leading to precise alterations in gene expression. For instance, researchers can instruct the AI to fabricate DNA sequences targeted specifically for stem cells, guiding them to differentiate into red blood cells while avoiding the formation of platelets. This level of specificity in genetic manipulation was previously unattainable, showcasing the immense potential of this technology.

One of the notable aspects of this study is the methodical approach taken by the researchers. By predicting the requisite combination of DNA nucleotides – adenine (A), thymine (T), cytosine (C), and guanine (G) – the model can generate synthetic fragments that meet the desired gene expression patterns for designated cell types. Following the design process, the researchers chemically synthesized roughly 250-nucleotide long DNA fragments, which were subsequently delivered to cells using viral vectors. This methodology yielded successful outcomes, validating the predictive capabilities of the AI model.

In a proof-of-concept experiment, the researchers tasked the AI with generating synthetic sequences that would activate a gene responsible for producing a fluorescent protein. This was achieved while ensuring the surrounding gene expression patterns remained unchanged. The fragments were introduced into mouse blood cells, resulting in successful integration of the genes into random locations within the genome, all aligning with the predictions made by the AI. Such precision exemplifies the transformative impact that AI can have on genetic research and therapy.

Dr. Robert Frömel, the first author of the study, emphasized the vast ramifications of this advancement, likening the process of designing genetic sequences to writing software for biological systems. This analogy captures the essence of the research, highlighting the potential for inducing specific cellular behaviors and developmental pathways with pinpoint accuracy. As gene therapy continues to evolve, the ability to finely tune gene expression could hold the key to enhancing treatment effectiveness while minimizing side effects, particularly in cells and tissues where adjustment is necessary.

Another significant aspect of this research is its contribution to understanding gene regulation and enhancer elements, small DNA fragments integral to controlling gene activity. Traditionally, geneticists have relied on naturally occurring enhancers, which can limit their options to sequences that evolution has already provided. In contrast, AI-generated enhancers possess the potential to engineer novel switching mechanisms that nature has yet to produce, enabling researchers to tailor gene expression patterns for specific therapeutic outcomes.

However, the successful development of such AI models necessitates access to high-quality data, which has historically been sparse for enhancers. To address this challenge, Dr. Lars Velten, the corresponding author of the study, explained the need for deciphering the “grammar” of enhancer sequences. By systematically investigating the nuances associated with enhancer functionality, researchers can begin to generate entirely new combinations of DNA sequences that could redefine our approach to genetic engineering.

Over the course of five years, the research team compiled an expansive dataset, synthesizing over 64,000 distinct synthetic enhancers. Each enhancer was meticulously designed to explore varying arrangements and strengths of binding sites for 38 different transcription factors, resulting in the largest library of synthetic enhancers created to date within blood cells. This ingenuity not only surpassed previous approaches but also provided a clearer insight into the mechanisms governing blood cell development and immune system functionality.

Upon inserting synthetic enhancers into cells, the researchers meticulously observed their activity across seven distinct stages of blood cell development. Unexpectedly, many enhancers were found to activate gene expression in specific cell types, yet functioned to repress gene activity in others. Such contrasting effects challenge conventional understandings of enhancer behavior and introduce novel concepts such as “negative synergy,” where two factors that typically induce gene activation together might actually suppress the gene when combined.

The experimental data generated from the research played a pivotal role in establishing the guiding principles for the AI-driven design model. As the model absorbed substantial metrics on enhancer-induced gene activity in real cellular contexts, it became proficient at predicting new sequences capable of producing on/off effects, even for sequences previously absent from the natural world. This predictive power of the AI marks a significant leap forward in the field and resonates with the aspirations to expand the horizons of genetic engineering.

The study ultimately serves as a testament to the potential of AI in biological research, illustrating that these technologies can address practical challenges in genetic modification before larger-scale implementation is pursued. The endeavor remains at the precipice of discovery, with human and mouse genomes containing an estimated 1,600 transcription factors that continue to be crucial in regulating gene expression.

As the researchers embark on further exploration, they are well-positioned to unlock new pathways in genetic therapy, offering an era where gene expression can be finely controlled to improve health outcomes. This work will likely catalyze future research endeavors, propelling innovation forward in both the fields of artificial intelligence and genetics, as scientists continue to seek remedies for complex diseases and genetic disorders.

The collective efforts of the research group, including notables like Lars Velten, Robert Frömel, Julia Rühle, Aina Bernal Martínez, Chelsea Szu-Tu, and Felix Pacheco Pastor, demonstrate how interdisciplinary collaboration can yield profound scientific advances. As the CRG team builds upon these findings, the implications of their work will reverberate through the scientific community, inspiring generations to come.

In conclusion, the marriage of AI and genetic engineering as showcased in this study not only represents a monumental shift in ability but also poses exciting possibilities for the future of medicine. As researchers grapple with the implications of their findings, the broader question remains: How can we harness this newfound power to address some of humanity’s most pressing health challenges?

Subject of Research: Cells
Article Title: Design principles of cell-state-specific enhancers in hematopoiesis
News Publication Date: 8-May-2025
Web References:
References:
Image Credits: Aina Bernal Martínez/Centro de Regulación Genómica

Keywords