The Power of Evo 2: AI Meets DNA

Evo 2 is the largest AI model for biology ever created, trained on an astonishing 9.3 trillion DNA "letters" (nucleotides) carefully selected from genomes spanning the entire tree of life. This massive dataset ensures that Evo 2 can recognize patterns and relationships in genetic sequences at an unparalleled scale.

For the first time, scientists can design DNA with AI, moving beyond simple sequence analysis to active DNA generation. Evo 2 enables researchers to predict, modify, and even create entire genetic sequences, opening new possibilities in medicine, agriculture, and synthetic biology.

Decoding the Dark Genome

One of the biggest challenges in genetics is understanding the non-coding regions of DNA—vast stretches of the genome that do not code for proteins but play crucial roles in regulating gene expression. These regions control when and how genes are activated, influencing everything from development to disease.

Evo 2 is designed to decode these non-coding elements, helping researchers uncover their functions and use this knowledge to develop gene-based therapies, synthetic life forms, and precision agriculture solutions.

From Reading DNA to Writing It

To put Evo 2’s impact into perspective:

• Previous AI models could "read" DNA like a book, analyzing genetic sequences and identifying patterns.
• Evo 2 can "write" entirely new DNA, designing functional genes, chromosomes, and even full genomes from scratch.

This means scientists can now engineer biological systems with AI, designing new proteins, metabolic pathways, and genetic circuits to address real-world challenges.

A Step Toward Generative Biology

The Arc Institute describes Evo 2 as a major step toward "generative biology"—a revolutionary approach where AI is used to create novel biological structures rather than just analyzing existing ones. This could lead to breakthroughs such as:

• New medicines: AI-generated enzymes and proteins tailored for targeted therapies.
• Disease-resistant crops: Genetically optimized plants for higher yield and climate resilience.
• Synthetic organisms: Custom-designed microbes for bioremediation, biofuel production, and industrial applications.

An Open-Source Revolution

Unlike many proprietary AI models, Evo 2 is open source, making its capabilities accessible to researchers worldwide. This democratization of AI-driven biology means that scientists from different disciplines can collaborate, experiment, and innovate, accelerating discoveries in genetic engineering and synthetic biology.

With Evo 2, the boundaries of what’s possible in DNA design, genetic engineering, and biological innovation are being redrawn. The future of life sciences is no longer just about understanding life’s code—it’s about writing it.

That is the challenge researchers face when studying how bacteria adapt to different environments. A bacterial gene may appear to be linked to a particular habitat, when the real reason is simply that closely related bacteria happen to live there.

In their 2025 Genome Biology paper, Bujdoš, Walter, and O’Toole introduce aurora, a machine-learning tool designed to tackle this problem.

Aurora identifies potentially mislabeled or unusual bacterial strains before performing genome-wide association studies (GWAS). By cleaning up the dataset and accounting for bacterial evolutionary relationships, it can help researchers find genetic features that are more genuinely connected to habitat adaptation.

The researchers tested aurora using simulated and real bacterial datasets and found that it could recover important genetic associations even when datasets contained misleading labels.

The bigger lesson is simple: better biological discoveries often begin with better data.

Aurora gives researchers a new way to separate real genetic clues from misleading ones—and could help us better understand how bacteria adapt, survive, and evolve in the environments they call home.

Based on Bujdoš et al., “aurora: a machine learning GWAS tool for analyzing microbial habitat adaptation,” Genome Biology (2025).[Read the original paper](https://link.springer.com/article/10.1186/s13059-025-03524-7?)

Research Directions

Stony corals: evolution of novelty and photosymbiosis

Symbiotic acoels: cell type evolution and photosymbiosis

Animal genomes: structural evolution and gene regulation

https://sgel.biodiv.tw/home

A recent study published in Genome Biology (https://link.springer.com/article/10.1186/s13059-026-04280-y) addressed this challenge by introducing dicast (https://github.com/burgshrimps/dicast), a machine-learning approach for detecting structural variants in short-read sequencing data. Structural variants, such as large deletions, insertions, duplications, and inversions, can have significant biological and clinical implications, yet they are challenging to identify with short-read technologies. Because different detection methods frequently yield divergent results, researchers commonly employ consensus calling, considering a variant valid only if multiple tools detect it. While this approach reduces false positives, it relies on the potentially flawed assumption that the majority is always correct.

The researchers explored the impact of evaluating the supporting evidence for each variant, rather than simply tallying the number of algorithms that identified it. To establish a ground truth, they analyzed nine genomes using multiple sequencing technologies and 15 detection methods, initially identifying approximately 35 million potential variants. Through extensive filtering, evidence integration, and manual review of over 11,500 variants, they developed a robust benchmark comprising more than 236,000 structural variants. The findings underscored the complexity of the problem: short-read methods detected fewer than half of deletions and less than 10 percent of insertions, with performance declining markedly in repetitive genomic regions. In contrast, long-read technologies demonstrated superior detection capabilities. However, replacing the substantial volume of existing short-read data in clinical and research settings is not immediately feasible. Consequently, the researchers questioned whether short-read data might harbor more information than conventional analytical pipelines currently extract.

This line of inquiry led to the development of dicast. Rather than merely confirming agreement among multiple tools, dicast identifies patterns in sequencing data, including split and clipped reads, discordant read pairs, alignment characteristics, and the surrounding genomic context. An XGBoost machine-learning model evaluates which combinations of these signals are indicative of genuine structural variants. Thus, the approach shifts from tallying algorithmic consensus to interpreting the underlying evidence.

The researchers subsequently conducted a targeted evaluation by examining structural variants detected by only a single short-read tool, which are typically missed by consensus-based approaches. dicast successfully recovered approximately 81% of these single-caller deletions, insertions, and duplications. The signals for these variants were present in the data, but conventional filtering methods failed to integrate them effectively.

The utility of dicast was further demonstrated in cohorts with rare diseases, including congenital limb malformations, atrial fibrillation, and neuromuscular disorders. In one instance, dicast achieved a deletion recall rate of approximately 0.96, compared to 0.74 using consensus calling. The median number of variants requiring manual review was 29 per sample. Among 31 experimentally validated variants that standard filters would have missed, dicast identified 12, whereas consensus calling detected only one.

Overall, dicast identified approximately 20 percent more potential disease-causing deletions than consensus-based methods. While a 20 percent increase may appear modest, in clinical genomics such improvements can have significant practical implications. Missing a deletion may leave a case unresolved, whereas detecting a structural variant can provide critical diagnostic insights.

The study does not claim that machine learning has rendered short-read sequencing superior to long-read approaches. Instead, the results underscore the effectiveness of long-read sequencing for structural variant detection. However, dicast highlights a more nuanced perspective: substantial biological information may still be recoverable from the extensive short-read datasets already available.

The principal lesson extends beyond the detection of structural variants. For many years, bioinformatics pipelines have relied on threshold-based criteria, such as minimum coverage, quality scores, or support from multiple tools. While these rules are useful, biological phenomena do not always conform to rigid checklists; multiple weak signals, when considered collectively, can provide compelling evidence.

This perspective prompts consideration of the solitary variant: one algorithm identifies it, while several others do not. Traditional consensus techniques might have dismissed it, yet machine learning approaches evaluate the available evidence to determine whether the variant is plausible.

Occasionally, the most significant variant within a genome is the one that is almost universally overlooked.

Read more at https://link.springer.com/article/10.1186/s13059-026-04280-y