Building Resilient Biodefense
1AI is giving biodefense a new toolset; now we need to make it operational
Emerging vectors of attack have historically created the need for new systems of defense. Networked computing created continuous cyber defense. Cheap offensive drones drove modern air defense systems. AI is now expanding biological capability, but we have not yet built its defensive equivalent.
Today's biodefense apparatus remains fragmented, and many of the components are vulnerable to novel biological threats, whether natural or engineered. To stay ahead, we will need systems for continuous situational awareness, threat characterization, and adaptive response that improve as underlying models advance.
The tools to build that system are arriving quickly. Specialized models are increasingly capable of making inferences about biology at the DNA and protein levels. Generative models can create de novo designed genes and functional biomolecules without relying on an existing template. Reasoning models can combine these specialized models with bioinformatics tools, experimental data, and scientific literature to execute increasingly complex biological workflows.
So far, these capabilities have mostly been demonstrated as proofs of concept in therapeutic design and scientific discovery.
Valthos applies the same advances to biological defense. We build the layer that brings frontier AI into the biodefense chain: turning new models into usable capabilities for detecting, understanding, and responding to biological threats.
2Grove turns biological data into operational threat assessments
Grove, a platform developed by Valthos, combines sequencing data across laboratory programs, field workflows, and public repositories into interpretable threat assessments. Grove is model-agnostic, and continually updates the frontier AI models, biological models, and bioinformatics tools needed to assess potential threats.
The system runs on both cloud infrastructure and deployable edge compute, to accommodate on-target characterization under connectivity constraints. On DoW-provided data, Grove moved from sample input to readout in approximately 20 minutes in the cloud and 53 minutes on NVIDIA Jetson edge hardware.
Time from FASTQ submission to composition readout
SOURCE: DoW-provided raw sequencing reads, from a field-collection program training event.
Grove analyzed a DoW-provided Oxford Nanopore dataset comprising 133,845 reads (mean read length of 850 bases, 113.8 million total bases). The Grove classification pipeline combined quality control, taxonomic classification, and alignment-based confirmation to identify candidate organisms and assess the sequencing evidence supporting each detection. Reported elapsed time from FASTQ input to this initial profile was approximately 20 minutes on cloud infrastructure and 53 minutes on NVIDIA Jetson AGX Thor edge hardware.
Grove outputs a decision-ready threat assessment: what was detected, the evidence supporting that conclusion, what the threat may do, which countermeasures may work, and where uncertainty remains.
Percent of known Bundibugyo ebolavirus read pairs detected
SOURCE: Los Alamos National Laboratory blinded dataset using a representative wastewater sequencing background and computationally incorporated target reads.
Grove analyzed a LANL-provided blinded benchmark dataset containing approximately 20 million individual reads (10 million paired-end fragments, 151 bases per read; 3.02 billion total bases), with Bundibugyo ebolavirus reference-derived target sequences computationally incorporated into a representative wastewater background. The Grove classification pipeline combined quality control, taxonomic classification, and alignment-based confirmation to identify the pathogen and evaluate supporting sequence evidence. After analysis, comparison against the source labels showed that Grove detected and confirmed 99 of 100 target read pairs (99%).
3Arbor tests and validates new AI capabilities before deployment
Biological AI will continue to improve. Rebuilding the defensive system around every new model is too slow and disruptive.
Arbor is the evaluation layer to keep Grove up to date. New AI models and capabilities are tested on operational data and real mission constraints before they are incorporated into the deployed system. We measure how models perform under real-world conditions, so that we can help operators understand where there is confidence and uncertainty when the models are deployed. As the frontier improves, the system can improve with it.
Percentage of target reads detected across samples
SOURCE: Valthos benchmark datasets, which incorporate reference-derived target pathogen sequencing reads with wastewater, clinical, and air sequencing background datasets.
Grove was evaluated across 330 benchmark datasets in Arbor, created by computationally incorporating reads from 110 reference genomes spanning 42 viral targets into wastewater, clinical, and air sequencing backgrounds. Target reads were derived from natural reference genomes and introduced at defined coverage levels of 1× in wastewater, 5× in air, and 100× in clinical samples. Grove's classification pipeline combined quality control, taxonomic classification, and alignment-based confirmation. Using the known target reads as ground truth, Grove detected and confirmed 87.2% of target reads in air, 95.5% of target reads in clinical samples, and 97.1% of target reads in wastewater. Detection rates are calculated against all target reads incorporated in the test datasets prior to quality filtering.
4Valthos turns AI into a standing system for biological response
Together, Grove and Arbor ensure that advances in AI lead to stronger biological defense. We organize around four core capabilities:
- Detect a threat. Establish whether a biological signal represents a credible threat.
- Characterize its impact. Assess what kind of risk it poses, to inform the urgency of response: for example, how effectively will it transmit? Will prior immunity protect humans?
- Enact a mitigation strategy. Evaluate which intervention is most effective to mitigate the threat, from medical countermeasures to critical logistics decisions. Accelerate the design of novel countermeasures against the landscape of evolving targets.
- Adapt. Evaluate new capabilities and update the system to keep pace with emerging threats.
Resilient biodefense depends on the entire chain. Improvement on a single modeling benchmark is not a guarantee that defense as a whole is stronger. We will need machinery to continually translate model predictions into better decision support at each step.
Consider three sequences: a trace-level match to smallpox, a dengue-like virus with unusual mutations, and a sequence that does not resemble anything we've seen before. Each could be noise, or the first sign of a real threat. Telling the difference depends on the whole body of evidence, not a single model's output.
Grove and Arbor do this continuously: Grove combines the outputs of multiple AI models to turn incoming biological data into threat assessments, and validation from Arbor shows operators where there is enough confidence to act (see the case study on the hantavirus outbreak below).
Each cycle of testing and updating expands the system: more models, more data, more failure modes tested, and increasing capability to see and respond to biological threats.
Hantavirus characterization in Grove
Collection locations for the 2026 Andes virus outbreak sequences, with the corresponding Pathoplexus sequence record for each sample.
The outbreak
After passengers on the MV Hondius were confirmed to be infected with a hantavirus (Andes virus), one of the immediate questions was whether the sequence in this outbreak contained mutations that could change the virus's behavior: for example, could it spread more effectively or has it adapted more readily to humans?
Once genomic data is released, Grove begins characterization and automatically updates the assessment as new sequences arrive.
We have a narrow window in front of us to ensure advances in biological capability are used for human resilience. Our goal is to close the gap between biological risk and our ability to defend against it. If that's your mission too, get in touch.