🔬 Bio-security is an AI Arms Race - Eric Nguyen (CEO, Radical Numerics)

Eric Nguyen explains that a genomic language model (GLM) is a large language model trained on DNA sequences rather than words. Radical Numerics’ goal is not only to have models read DNA and predict its regulatory function, but also to generate new DNA sequences. The motivation is that biology’s “grammar”—how combinations of letters encode traits, regulation, and disease—is still poorly understood, even though the human genome contains roughly three billion letters. The group’s early HyenaDNA model addressed DNA’s unusually long context. Instead of relying on attention, it used a convolutional “Hyena operator” that could process sequences up to one million positions, then used those sequences to predict genomic function and long-range interactions. Nguyen says this was important because short-context models miss relationships between distant motifs, much as natural-language models can suffer from “context rot” when inputs become long. Evo extended the idea from reading DNA to writing it. Nguyen says the team wanted to ask what generative AI could accomplish in biology that humans could not do easily in a lab. Evo could work across DNA, RNA, and protein-related information, and the team demonstrated that it could generate a new CRISPR-Cas system involving both RNA and protein. Later, researchers used Evo to generate a functional bacteriophage genome from scratch—an achievement Nguyen describes as both evidence of the technology’s potential and a warning about what manipulating biological function could enable.

Nguyen says Evo was an important proof of concept but often remained behind specialist DNA models, particularly on human genetics. Some researchers therefore questioned why they should use a large general model instead of a smaller model designed for one task. Radical Numerics’ newer Omni release was intended to show that a genomic foundation model could become useful across many tasks through mid-training and post-training, rather than stopping at raw pretraining. The team structured genomic tasks in a question-and-answer format. A prompt can indicate whether the model should predict disease risk, score a mutation, or perform a design task, with special tokens and examples teaching the desired input-output structure. Nguyen compares this to the alignment work that turns a natural-language base model into something that answers users’ questions in a useful format. The exact recipe remains partly proprietary, but it combines task-specific datasets, structured prompts, and additional training; reinforcement learning may also be included. The goal is a unified model rather than a collection of disconnected task heads. Nguyen says early experiments used separate heads in some cases, but Radical Numerics is moving toward one model that can generalize across modalities and scales. He argues that DNA is a foundational modality from which the model can learn information relevant to RNA, proteins, and other biological systems. Early interest has come from hospitals and nonprofits that have patients with symptoms but cannot identify which variants are causal—especially “variants of unknown significance.”

The hosts ask Nguyen to explain the benchmark results in the blog post. A central task is variant-effect prediction: given a DNA mutation, determine whether it is likely to cause disease. A variant may change one DNA letter, delete a letter, or remove a larger region; most changes do nothing, but some affect critical genomic regions. The benchmark uses collections such as ClinVar and other datasets containing known and uncertain variant classifications. For a zero-shot approach, the model compares the likelihood it assigns to the normal, or wild-type, sequence with the likelihood assigned after the mutation. Nguyen describes this as a “surprise factor”: sequences that are unusual relative to what the model learned from genomes may be more likely to disrupt biological function. Omni also receives additional mid-training that teaches it the structure of the specific disease-prediction task, allowing it to improve on the raw likelihood score. The hosts emphasize that many previous tools perform best in coding regions—the roughly 1.5% of the genome that directly encodes proteins. Nguyen says Omni is especially promising in non-coding and long-range regulatory regions, which control when and how strongly genes are expressed and may contain many disease-causing mutations. He says Omni was competitive or state of the art across most of the reported benchmarks, including against methods such as Borzoi and CAD. Borzoi directly predicts annotated functional-genomics tracks from DNA, while Omni first learns from much larger quantities of unannotated genomic sequence and then applies alignment and task-specific training.

The hosts question whether Omni’s benchmark performance could be inflated by data leakage. Because the model is pretrained on large genomic corpora, benchmark variants from ClinVar and related datasets might appear directly or in near-duplicate form in the training data. Nguyen says the team treated this as a serious risk: bioinformaticians curated and deduplicated the data, aligned sequences, and removed training examples that were too similar to benchmark examples. The discussion distinguishes several levels of performance. Evo 2’s column in the benchmark represents a largely zero-shot use of the pretrained model’s likelihood scores. Another result, from Goodfire’s work, fine-tunes or uses Evo 2 scores for a particular task. Omni adds another layer by incorporating the benchmark’s task structure during mid-training, so the model can receive a prompt in a defined format and return a prediction without requiring researchers to extract embeddings and train a separate regression head. Nguyen says this matters practically: a scientist does not have to repeatedly fine-tune a model for every question. The model can be trained on many task formats and then prompted to produce the relevant score or prediction. The hosts note that CAD had long been a strong baseline because it effectively combined many methods—the “kitchen sink.” Nguyen says Radical Numerics deliberately treated that as the bar rather than comparing Omni with a single weaker model.

The conversation then turns to whether genomic models can use a version of chain-of-thought reasoning. Nguyen describes the natural-language idea as giving a model intermediate steps or “scratch space,” which often improves its answer. Since DNA and RNA do not naturally present their reasoning in sentences, the team tested a simpler biological analogue using RNA aptamers—sequences associated with experimentally measured fitness scores. The experiment showed the model a sequence of aptamers ordered from lower to higher fitness, then asked it to continue the trajectory. A portion of the highest-performing examples was held out. In silico, Omni was able to generate plausible aptamers with scores better than the examples it had been shown; Nguyen says the team was still validating the results in the wet lab. The hosts explain that the underlying dataset came from an iterative experimental process: researchers generated many sequences, measured how effectively they bound or acted on a target, selected better sequences, and repeated the process. The “chain” therefore represents progressive biological refinement rather than verbal reasoning. Nguyen says the same approach could apply whenever sequences have associated functional measurements. Examples include designing bacteriophages with stronger antimicrobial activity or proteins that selectively bind rare-earth minerals. In the latter case, the model could be shown sequences with high affinity for a desired mineral and low affinity for others, then asked to propose better variants. The team’s hypothesis is that sequence-plus-fitness trajectories can become a general prompting and design paradigm.

The hosts ask whether rare-earth extraction is really a protein-design problem, where structure-specific models might have an advantage. Nguyen replies that Radical Numerics is most interested when biological context matters. Rather than designing an isolated protein, Omni can examine the surrounding genomic neighborhood—the organism, regulatory regions, and other sequences—and use that context to find plausible variants that evolution has already explored in related organisms. Nguyen says the model does not yet need to design an entire organism for this use case. The broader genome acts as a record of the physical and biological environment in which a protein evolved. That context may reveal why a protein has a particular function or selectivity. The same reasoning motivates the company’s interest in all domains of life: humans, prokaryotes, viruses, and microbial systems. Viral genomes matter both for therapies and for biodefense; bacteriophages could potentially become targeted antimicrobials against infections, including antibiotic-resistant bacteria. Long context remains a major technical bottleneck. The human genome is far longer than Omni’s roughly two-million-position context, and even engineered bacterial genomes can exceed it. Nguyen says the company therefore operates first as an AI research lab, continuing to develop architectures and GPU-efficient kernels rather than simply adopting existing open-source models. He recounts that the team began with long-context language modeling, noticed that its models handled long inputs unusually well, and then chose DNA because it was an extreme test: three billion base pairs and a field where earlier models commonly used contexts of only one or two thousand positions.

Radical Numerics is also exploring mechanistic interpretability—using the model’s weights, embeddings, and activations to identify the biological patterns it has learned. Nguyen describes the model as compressing large amounts of genomic information into internal representations. The interpretability task is to unpack that compression and determine whether the representations correspond to recognizable biological structure. Early targets include GC content, repeated sequences, transcription-factor motifs, and the regulatory patterns that influence gene expression. Transcription factors bind DNA and alter expression without changing the DNA sequence itself; because transcription-factor genes can regulate one another, their effects create a highly combinatorial system. Nguyen says that understanding these motifs could eventually support transcription-factor design, an area relevant to disease and aging research. The team’s broader ambition is to map a model’s internal “manifold” or landscape of disease. Instead of merely producing a disease score, the model might organize relationships among diseases, variants, regulatory motifs, and other biological signals. Nguyen stresses that DNA is only one sensor of biology. Future systems should combine DNA with RNA, proteins, epigenomics, chromatin accessibility, methylation, metabolomics, and other molecular measurements. The hosts suggest that this multimodal structure could also make biological chain-of-thought more natural, because a model could move between sequence, molecular state, and functional outcome rather than reasoning only through DNA letters.

Nguyen says Radical Numerics has a dual mandate: develop biological design capabilities while also building defenses against their misuse. He argues that the same models that generate useful sequences can often discriminate between sequences or predict whether a sequence is pathogenic. Since generative biology is advancing faster than many safeguards, a design lab is well positioned—and has a responsibility—to build defensive tools too. The company’s biodefense strategy has three main technical pillars. Detection and surveillance would identify pathogenic or dangerous sequences in environmental samples such as airport swabs or sewage. Attribution would estimate whether a threat is natural, associated with a particular country, or engineered in a specific laboratory. Countermeasures would help create responses such as antivirals or antimicrobials. Nguyen also mentions deterrence as a broader government-level function. He says current systems often rely on sequence matching against known pathogen databases, which can miss new or intentionally altered sequences that preserve function while changing their spelling. The hosts discuss DNA-synthesis companies that manufacture customer-designed DNA and ask how a detector could avoid blocking legitimate scientific innovation. Nguyen does not claim perfection. His argument is that defense has been far behind design, so even an imperfect frontier model can improve the baseline and give regulators, manufacturers, and biodefense researchers something to build on. He frames the situation as an arms race analogous to cybersecurity: design systems will improve, defensive systems must test and catch up, and neither side can be treated as finished. Nguyen worries most about lowering the expertise and time required to engineer biological functions. The danger could come from intentional state programs, especially during wartime, or from accidental release by researchers pursuing a legitimate experiment. Nevertheless, he remains optimistic that the benefits for discovery and human health outweigh the risks if the defensive ecosystem develops alongside the design technology.