The glossary is designed for readers entering the study without a background in molecular biology or biomedical engineering. Where a term has a more technical meaning in research literature than appears here, the definition gives the working sense used in the study. The companion references page points to primary literature for full treatments.
The molecule that carries the genetic instructions for an organism. A long polymer made of four building blocks (called bases — A, C, G, T) arranged in a sequence. The sequence is the information.
Each human cell contains about 2 metres of DNA tightly packaged into a nucleus a few micrometres across. The full human DNA sequence is the genome.
A relative of DNA that serves several roles. Messenger RNA (mRNA) carries instructions from DNA to the cell's protein-making machinery. Other RNAs catalyse reactions, regulate genes, or serve structural roles.
The COVID-19 mRNA vaccines work by delivering RNA instructions that briefly cause cells to produce a viral protein, which the immune system then learns to recognise.
The complete set of an organism's DNA. The human genome contains about 3 billion base pairs and roughly 20,000 protein-coding genes plus many regulatory regions.
"Sequencing the genome" means reading the order of those 3 billion bases. This now costs under $200 and takes hours.
Chemical modifications to DNA and the proteins it wraps around. These modifications don't change the DNA sequence but determine which genes are active in which cells.
A neuron and a liver cell have identical DNA but very different epigenomes — which is why they behave so differently. The epigenome is at least as important as the genome for determining a cell's behaviour.
A specific difference in DNA sequence at a particular position, compared to a reference genome. Each person has roughly 3 million variants from the human reference.
Most variants have no known effect. A small number cause disease. A larger number affect things like drug metabolism. Variant interpretation — figuring out what a variant does — is the central problem of advance B1.
Long chains of amino acids that fold into specific 3D shapes and perform almost all the functional work in biology — catalysing reactions, transporting molecules, providing structure, transmitting signals.
The human genome encodes about 20,000 different proteins. Each one's function is determined by its 3D shape, which is determined by its amino acid sequence, which is determined by the DNA sequence of its gene.
A category of cell with characteristic appearance, behaviour, and gene expression. The human body has hundreds of distinct cell types — neurons, hepatocytes, fibroblasts, T cells, and so on — all from the same genome.
The Human Cell Atlas is a major international effort to comprehensively catalogue and map all human cell types.
A family of techniques for precisely editing DNA in living cells. Derived from a bacterial immune mechanism; adapted to therapeutic and research use starting around 2012.
The first FDA-approved CRISPR therapy (Casgevy, for sickle cell disease) was approved in December 2023. CRISPR is foundational to several Bucket C advances.
Technology that measures gene activity (which RNAs are being produced) in individual cells one at a time. Allows researchers to see how cells in a tissue differ from each other rather than averaging over millions of cells.
Modern scRNA-seq experiments routinely profile tens of thousands of cells in parallel. The data is destructive — the cells must be killed to read their state — which is why doing this in vivo remains an unsolved problem (see A1).
Methods that profile gene activity in a tissue while preserving the spatial relationships between cells. Combines the resolution of single-cell sequencing with the geographic information of imaging.
Currently requires thin frozen tissue sections, meaning samples must be removed from the patient. Advance A2 is, in effect, the goal of doing spatial transcriptomics in a living intact patient.
Microscopic fatty droplets that encapsulate and deliver drug payloads — particularly nucleic acids like mRNA or DNA — to cells in the body.
The mRNA COVID-19 vaccines work because LNPs successfully deliver mRNA into cells. LNPs are now the leading delivery technology for many gene therapies and underlie several Bucket C approaches.
A family of harmless viruses repurposed as delivery vehicles for gene therapy. Different AAV variants target different tissues. AAV9 in particular can cross the blood-brain barrier.
Several FDA-approved gene therapies use AAV delivery. Limitations include immune responses on repeated dosing and difficulty delivering large genetic payloads.
A large AI model trained on broad data that can be adapted (fine-tuned) for many specific tasks. AlphaFold (protein structure), ESM (protein function), and GPT-class language models are all foundation models.
Medical foundation models — trained on imaging, lab values, clinical notes, and genomic data — are central to advance A4 and several Bucket B capabilities.
A computational model of a specific patient detailed enough to simulate that patient's response to interventions. Used in cardiology for surgical planning, with broader applications in development.
Whole-patient digital twins are far off; the realistic near-term target is organ-system-specific twins integrated to cover increasingly more of physiology. See advance B4.
A technique for training machine learning models across multiple data sources without centralising the data. The model travels between data holders rather than the data travelling to a central training site.
Critical for medical AI where patient data cannot leave its host institution due to privacy regulations. See advance B5.
A medical emergency where the body's response to infection causes damage to its own tissues and organs. Kills more than 11 million people annually worldwide.
Sepsis is referenced repeatedly as a likely early AIHS target because it requires fast, integrated, multi-system response — exactly the capability AIHS is designed to provide.
A category of diseases characterised by cells that grow and divide outside normal regulatory controls. Specific cancers differ enormously in causes, behaviour, and treatment.
Many AIHS-relevant diagnostics and therapies (liquid biopsy, multi-target coordinated therapy, in vivo gene editing) are developing fastest in cancer because of clinical urgency and research investment.
A state where cells stop dividing but don't die; instead they secrete inflammatory signals that affect nearby tissues. Accumulates with age and contributes to age-related diseases.
Senolytic drugs — which selectively kill senescent cells — are an active research area and are mentioned in Bucket C priorities.
The system of tight cellular junctions and active transport mechanisms that keeps most molecules in the bloodstream out of the brain. Protective against toxins and pathogens, but also a barrier to most drugs.
Crossing the blood-brain barrier safely is one of the leading challenges in neurological drug development and a major focus of advance C2.
A misfolded protein that can induce normal copies of the same protein to also misfold. Causes uniformly fatal neurodegenerative diseases including Creutzfeldt-Jakob disease and BSE.
Prions are biology's closest real analog to the kind of shape-propagating agent imagined in protomolecule fiction. The Institute's learning module M-Bio-01 covers them in depth.
The Institute's term for closed-loop diagnostic-therapeutic platforms meeting four criteria: continuous real-time diagnosis, patient-specific interpretation, active therapeutic intervention through the same platform, and closed-loop operation.
This is the central object of study S-001. A system meeting some-but-not-all criteria is a partial AIHS.
This study organises AIHS capabilities into three buckets: A (Diagnostics), B (Interpretation), and C (Therapeutic delivery).
Each is independently necessary; none alone is sufficient. The buckets contain specific required advances numbered A1-A5, B1-B5, and C1-C7.
A discrete research target whose completion would meaningfully advance one of the buckets toward AIHS capability. The study identifies 17 such advances (A1-A5, B1-B5, C1-C7), each with its own detail page.
An advance is not a milestone — it can take decades and may not yield to current approaches. Each advance has a risk classification and time horizon.
One of four advances identified as gating constraints on full AIHS maturity: B3 (causal disease modelling), C4 (neural reconnection), A2 (whole-body molecular mapping), and C3 (architectural tissue reconstruction).
These are the advances most likely to remain partial indefinitely. Their progress is what the Phase II and Phase IV decision gates are designed to assess.
An explicit moment in the AIHS roadmap when the programme should be reassessed against pre-registered criteria. Three gates are defined: at year ~5, ~15, and ~25.
Gates are not pass-fail. The decisions are recalibrations — adjusting ambition to evidence — rather than continue-or-stop verdicts. See the decision-gates page.
The Institute's labelling scheme for claims, used throughout published materials: Established for well-validated science, Frontier for active research, Speculative for plausible extrapolations, and Handwave for content that crosses into the science-fictional.
Used to make epistemic status legible to readers without requiring them to know which claims are settled and which aren't.
A system meeting some but not all four AIHS criteria. The realistic near-term target — partial AIHS for narrow indications is achievable on a 15-year horizon while full universal AIHS may require breakthroughs decades away.
The continuous-glucose-monitor-plus-insulin-pump system used by many diabetic patients is, in a meaningful sense, the simplest existing partial AIHS.
The Encyclopedia of DNA Elements — a comprehensive functional annotation of the human genome. Identifies which parts of DNA do what (coding regions, regulatory elements, etc.).
NIH-funded effort to build 3D atlases of human tissues at single-cell resolution. Foundational reference data for AIHS-relevant cellular interpretation.
Large-scale US neuroscience programme covering brain cell census, neural circuit mapping, and tools for in vivo neuroscience. Foundational for any AIHS neural capability.
DeepMind's foundation model for predicting protein 3D structure from amino acid sequence at near-experimental accuracy. The proof of concept that foundation-model approaches can solve longstanding biology problems.