Evo 2 emerges from collaborative efforts by the Arc Institute, NVIDIA, and leading academic institutions such as Stanford University, UC Berkeley, and UC San Francisco. The model’s architecture enables it to scan and interpret massive data sets, revealing intricate patterns in DNA sequences that eluded researchers for decades. Unlike traditional methods that may require years or even decades to unravel comparable genetic insights, Evo 2 can perform these tasks in a fraction of the time with extraordinary accuracy.

A deep dive into the technical specifications reveals that Evo 2 was trained on over 9.3 trillion nucleotides sourced from more than 128,000 whole genomes as well as metagenomic data, encompassing bacteria, archaea, phages, plants, humans, and a wide range of other eukaryotic organisms. This expansive training set allows Evo 2 to develop a profound understanding across life’s domains, solidifying its standing as the largest, most sophisticated AI model in biology to date. Its predecessors focused on more limited genetic windows, particularly single-cell genomes, while Evo 2 simultaneously processes genetic sequences up to one million nucleotides long. This capability is critical for analyzing long-range relationships in genomes, which play pivotal roles in gene expression and organismal complexity.

The foundation of Evo 2’s advanced performance is the novel AI architecture known as StripedHyena 2, a significant evolution from prior models. By leveraging the computational power of over 2,000 NVIDIA H100 GPUs on the DGX Cloud AI platform through AWS, the Evo 2 training required reimagining AI data ingestion and inference strategies for genomics. The model’s ability to handle 30 times more data and reason over eight times as many nucleotides compared to its predecessor empowers it to detect subtle, context-dependent variations within DNA sequences, a task that has historically challenged both machine and human experts.

Beyond raw computational prowess, Evo 2 shines in its practical applications. For example, it has demonstrated exceptional proficiency in pinpointing disease-causing mutations in critical human genes such as BRCA1, achieving over 90% accuracy in distinguishing benign from potentially pathogenic variants. This degree of precision can vastly accelerate genetic diagnostics and personalized medicine, reducing both time and financial costs linked to experimental validation in wet-lab environments.

The researchers behind Evo 2 envision the model as a versatile platform, capable of fostering innovations across scientific domains. The model’s publicly accessible code and integration into NVIDIA’s BioNeMo framework encourage the global research community to expand and fine-tune its capabilities. Arc Institute also partnered with the AI research lab Goodfire to develop sophisticated mechanistic interpretability tools that unlock the biological features and genomic motifs Evo 2 learns. This transparency bridges the gap between black-box AI and biological insight, enhancing trust and fostering deeper explorations of genetic code.

One of the potentially transformative applications of Evo 2 lies in synthetic biology and genome engineering. By interpreting genomic languages and evolutionary imprints, the model can guide the design of novel genomes that are as extensive as those found in relatively simple organisms like bacteria. Arc Institute’s work has already demonstrated functional synthetic bacteriophages designed using Evo 2, which could revolutionize approaches to combating antibiotic-resistant infections, a pressing global health crisis.

Evo 2’s prowess extends to biological specificity. A promising example cited by the developers includes engineering genes to express uniquely in specific cell types, such as neurons or liver cells. This precision genome editing could mitigate side effects in gene therapy by limiting gene expression to targeted tissues, an advancement with profound implications for developing safer, more effective medical interventions.

The AI model captures an essential truth about biology: just as evolution has left its fingerprints on nucleotides through millennia, these genetic patterns hold keys to molecular interactions and functional dynamics. By assimilating this evolutionary information, Evo 2 has acquired a form of generalized genomic literacy. Its capacity to “think” in the language of nucleotides represents an extraordinary step for generative biology, empowering machines to actively read, write, and even innovate within genetic sequences.

Importantly, the research team has exercised ethical caution in developing Evo 2. The training data explicitly omits pathogens harmful to humans and other complex organisms. Mechanisms have been put in place to prevent the model from generating productive outputs related to these excluded pathogens. Stanford’s Professor Tina Hernandez-Boussard and her team have been instrumental in guiding the responsible development and deployment of Evo 2 to ensure safe, ethical use of this transformative technology.

The scale and ambition of Evo 2 have not only broadened scientific horizons but also set new standards in the realm of biological AI. By creating a foundation model trained on one of the most extensive and diverse biological datasets yet assembled, the Arc Institute and NVIDIA have provided researchers worldwide a powerful tool akin to a universal operating system for genetic research. As co-author Dave Burke suggests, this model represents a “kernel” upon which countless future scientific applications can be built, from genome annotation to therapeutic design.

The widespread enthusiasm surrounding Evo 2 is underscored by its rapid adoption and application across diverse biological challenges—from predicting Alzheimer’s genetic risk to assessing variant effects in domestic animals. Arc Institute’s open approach, which shares training data, model weights, and code, invites the global research community to innovate, accelerating discovery while maintaining transparency—a critical factor in AI’s intersection with biology.

As Evo 2 continues to evolve and integrate with other technological advancements, its impact on medicine, synthetic biology, conservation, and fundamental biological research is poised to expand dramatically. With the power to decode life’s language at scale, this AI model not only deepens human understanding of genetic architecture but also opens avenues for designing life forms and therapies hitherto imagined only in science fiction.

Subject of Research: Not applicable

Article Title: Genome modeling and design across all domains of life with Evo 2

News Publication Date: 4-Mar-2026

Web References:

• Arc Institute GitHub: https://github.com/arcinstitute/evo2
• NVIDIA BioNeMo Framework: https://github.com/NVIDIA/bionemo-framework
• Goodfire mechanistic interpretability tool: https://arcinstitute.org/tools/evo/evo-mech-interp

References:
Brixi, G., Durrant, M.G., Ku, J., et al. (2026). Genome modeling and design across all domains of life with Evo 2. Nature. DOI: 10.1038/s41586-026-10176-5

Image Credits: Arc Institute

Keywords: Artificial intelligence, Life sciences, DNA, Genomics, Genome modeling, Synthetic biology, Evolutionary biology, Machine learning

Transcriptional responses are crucial elements in the framework of gene expression, dictating how genetic information is converted into functional products like proteins. Traditional methods of forecasting these responses have primarily revolved around statistical models, which can often fall short due to their reliance on known data patterns and assumptions. Scouter, however, breaks new ground by leveraging the expansive learning capabilities of language models that have been trained on vast datasets—effectively transforming how scientists can approach genetic perturbations and their myriad effects.

The utilization of large language models in this context is particularly striking. These models have shown unprecedented proficiency in understanding and generating human language, and their application in biological sciences is a testament to their versatility. The researchers behind Scouter have tapped into this robustness by fine-tuning the LLMs on sequence data and relevant biological literature, allowing the model to understand biological contexts and generate meaningful predictions based on genetic alterations.

One of the most prominent features of Scouter is its ability to provide contextually rich embeddings for gene sequences. This means that rather than treating genetic data as isolated strings of information, Scouter contextualizes these sequences within the broader biological environment they operate. By capturing the intricate web of interactions and regulatory mechanisms that govern gene expression, Scouter offers a more nuanced approach to predicting how alterations in the genetic code might lead to specific transcriptional outcomes.

The implications of Scouter’s predictions are vast. For one, it could significantly speed up the drug discovery process, allowing researchers to predict how cancer cells, for example, might respond to novel therapies based on their genetic make-up. Moreover, this technology could facilitate personalized medicine by enabling clinicians to tailor treatments that align with the individual genetic profiles of their patients, thus enhancing efficacy while minimizing adverse effects.

In practical terms, Scouter demonstrates its strength in numerous experimental setups, including those involving synthetic and endogenous perturbations. By simulating various genetic scenarios, the researchers were able to validate the model’s predictions against empirical data, showcasing its capacity to accurately forecast transcriptional changes. The robustness of Scouter was particularly emphasized in experiments involving gene knockouts and overexpression models, where it consistently outperformed traditional methods.

The collaborative effort that led to the development of Scouter is also noteworthy. The project brought together a multidisciplinary team comprising computational scientists, biologists, and machine learning experts. This convergence of expertise allowed for a comprehensive approach in creating a tool that is not only mathematically robust but also biologically relevant. The integration of insights from different fields ensured that Scouter operates at the nexus of computational efficiency and biological accuracy.

As the research landscape evolves, tools like Scouter are becoming increasingly important. The volume of genomic data generated in recent years has outpaced analysts’ ability to interpret it meaningfully. Scouter addresses this challenge head-on, acting as a bridge that connects raw genomic data with actionable insights. The model’s architecture allows for continuous updates as new data emerges, ensuring that its predictive capabilities remain current and relevant in a fast-paced research environment.

Moreover, Scouter’s design embraces open science principles, fostering transparency and collaboration within the scientific community. By making the model and its underlying code accessible, the research team encourages other scientists to build upon their work, potentially leading to further innovations in predictive modeling in genetics. This ethos not only accelerates the pace of discovery but also cultivates a culture of shared knowledge and collective advancement in the biotechnological landscape.

Looking towards future applications, one can envision Scouter being implemented in various domains beyond oncology. For instance, in the field of agriculture, predictions derived from Scouter could aid in developing crops that are more resilient to environmental stresses or diseases by understanding how specific genetic traits confer advantages. Similarly, the tool could pave the way for advancements in synthetic biology, enabling the design of microorganisms that can produce valuable compounds or perform complex bioconversions.

Zhu and Li’s Scouter is poised to usher in a new era of genetic research, where predictions can be made with unprecedented accuracy and depth. As the tool gains traction within the scientific community, it opens the door to innovations across numerous fields, underscoring the significant potential of large language models in domains that extend far beyond their initial scope. The research community stands at the brink of a transformative shift that intertwines biology with powerful computational techniques, promising a future where the intricacies of life can be untangled with newfound clarity.

In conclusion, the advent of Scouter represents a significant leap forward in the intersection of artificial intelligence and genetics. By harnessing the predictive power of large language models, Zhu, Li, and their team have not only created a tool capable of transforming how we understand transcriptional responses but also laid the groundwork for future explorations into the genome. As researchers continue to unravel the complexities of life at the molecular level, innovations like Scouter will undoubtedly play a pivotal role in guiding their discoveries and applications.

Subject of Research: Predicting transcriptional responses to genetic perturbations using large language model embeddings.

Article Title: Scouter predicts transcriptional responses to genetic perturbations with large language model embeddings.

Article References:
Zhu, O., Li, J. Scouter predicts transcriptional responses to genetic perturbations with large language model embeddings.
Nat Comput Sci (2025). https://doi.org/10.1038/s43588-025-00912-8

Image Credits: AI Generated

DOI: https://doi.org/10.1038/s43588-025-00912-8

Keywords: Genetic perturbations, large language models, transcriptional responses, computational biology, predictive modeling, gene expression, artificial intelligence, synthetic biology, personalized medicine.

The rise of antimicrobial resistance, particularly in common foodborne pathogens like Salmonella, has been exacerbated by the overuse of antibiotics in agriculture and human medicine. The genetic mutations that accompany such practices have led to strains that are resistant to standard treatment protocols. Traditionally, predicting this resistance has relied on bacterial antimicrobial susceptibility tests (ASTs), which can be resource-intensive and time-consuming. Furthermore, current genomic analyses often fall prey to issues of overfitting, largely due to the high-dimensional nature of whole-genome sequencing (WGS) data.

Given these complexities, the researchers turned to a dual-layered feature-selection process, which is crucial for extracting pivotal genetic information relevant to antimicrobial resistance. The initial step involves the application of a chi-square test to sift through vast datasets and identify the key resistance genes within the Salmonella genome. This technique is further complemented by a conditional mutual information maximization approach that enhances the process of identifying critical features.

Building on the initial findings, the team then developed a sophisticated algorithm, dubbed the Salmonella Antimicrobial Resistance Predictive Large Language Model (SARPLLM). This model is rooted in the Qwen2 LLM, augmented with low-rank adaptation (LoRA) technology, allowing for the conversion of complex genomic data into a sentence-like structure that the model can effectively analyze. By translating genetic variations and resistance markers into this textual format, the SARPLLM can predict antimicrobial resistance with a level of accuracy that far surpasses prevailing methodologies.

However, despite the promise of this innovative approach, the study does not shy away from acknowledging inherent challenges. Notably, the datasets used in genomic analyses often exhibit an imbalance in the representation of antimicrobial-resistant versus sensitive strains of Salmonella. This discrepancy can severely limit the efficacy of predictive models. To tackle this issue, the researchers developed the QSMOTEN algorithm, an adaptation of the existing SMOTEN algorithm that utilizes the principles of quantum computing. This novel algorithm encodes sample features into quantum states, enabling efficient computation of distances between samples. By drastically reducing the complexity of these calculations from a linear to a logarithmic scale, QSMOTEN effectively enhances the processing of complex high-dimensional data associated with WGS.

The remarkable findings presented in this study also encompass the creation of a user-oriented online platform designed for antimicrobial resistance prediction. Utilizing the Django framework for backend operations and Echarts for robust knowledge graph visualization, the platform features multiple modules. These include an intuitive predictive module for antimicrobial resistance, a dedicated space for presenting pan-genomics analysis results, an interactive gene-sample-antimicrobial knowledge-graph module, and a streamlined data download feature. This user-friendly interface enables researchers and practitioners to upload gene feature files for real-time predictions, bolstered by extensive data visualization capabilities.

Impressive experimental results underpin the claims made by the researchers regarding the superior performance of the SARPLLM algorithm in predicting antimicrobial resistance. With elevated F1-scores across various antimicrobial drugs, it is evident that the integration of LLMs into this domain can yield significant advancements. Moreover, the efficiency of the QSMOTEN algorithm in accurately measuring sample similarities has been validated on both virtual and physical quantum computing platforms, demonstrating the transformative potential of this technology in the realm of biological data analysis and augmentation.

Nonetheless, the authors are candid in recognizing the limitations that currently exist within their work. The intricate biological and genetic landscapes surrounding antimicrobial resistance present considerable challenges, particularly concerning the limitations of LLMs in fully grasping complex domain knowledge. The efficacy of these models is tightly intertwined with the quality and scope of the training data employed, further emphasizing the necessity for continuous improvements in this field of study. The nascent stage of quantum computing technology also remains a noteworthy consideration, underscoring the need for ongoing research aimed at refining integration processes and establishing more robust quantum infrastructures.

As the healthcare landscape grapples with the rising tide of antimicrobial resistance, this study represents a monumental shift towards more predictive and proactive healthcare solutions. The implications of leveraging advanced computational power to preemptively identify strains with resistance potential could not only pave the way for more effective treatment strategies but also inform public health policies aimed at reducing the prevalence of resistant pathogens in the environment.

Moving forward, the researchers plan to concentrate their efforts on enhancing the predictive platform by integrating multi-source datasets alongside domain-specific knowledge. This approach may significantly bolster the accuracy of resistance predictions, further mitigating the healthcare crisis posed by antimicrobial-resistant strains. Additionally, advancements in quantum hardware will be pivotal in achieving more streamlined and stable performance for such complex applications.

The innovative strides made in this research signify a promising trajectory towards harnessing cutting-edge technologies like quantum computing and LLMs to tackle some of the most formidable challenges in public health. The potential of these methods extends beyond the current study, setting a precedent for future explorations into the intersection of computational techniques and biological insights, ultimately aiming for better health outcomes globally.

—
Subject of Research: Prediction of Salmonella Antimicrobial Resistance
Article Title: Developing a Predictive Platform for Salmonella Antimicrobial Resistance Based on a Large Language Model and Quantum Computing
News Publication Date: 28-Jan-2025
Web References: https://doi.org/10.1016/j.eng.2025.01.013
References: 10.1016/j.eng.2025.01.013
Image Credits: Yujie You et al.

Keywords: Salmonella, antimicrobial resistance, large language models, quantum computing, predictive algorithms, pan-genomics, QSMOTEN, SARPLLM.