Appendix D — Glossary

Published

July 7, 2026

Definitions for terminology used throughout the handbook. The goal is practical orientation: what the term means, why it matters, and what evidence or chapter context should shape how it is read.

A

A-Lab

Autonomous materials-science laboratory described by Szymanski and colleagues (Szymanski et al., 2023). Useful as a demonstration of closed-loop experimental automation, with later debate about novelty and validation. In this handbook, A-Lab is a cautionary case for reading autonomous-discovery claims through measured output, baseline comparison, and independent review.

ACMG/AMP guidelines

The American College of Medical Genetics and Genomics and Association for Molecular Pathology standards for clinical interpretation of sequence variants (Richards et al., 2015). Defines five tiers (benign, likely benign, uncertain significance, likely pathogenic, pathogenic) with evidence-strength weights. The framework AI variant predictors must integrate into; see Variant Effect Prediction chapter.

AAV

Adeno-associated virus. Common viral-vector platform for gene therapy because it can deliver genetic cargo to cells in vivo. AI may help engineer capsids or prioritize variants, but clinical usefulness depends on biodistribution, dose, payload, immunogenicity, durability, manufacturing, and human safety evidence.

ADMET

Absorption, distribution, metabolism, excretion, and toxicity. The pharmacokinetic and safety properties that shape whether a compound can become a drug. Predicted by tools like Chemprop, ADMET-AI, and ADMETlab (Yang et al., 2019; Swanson et al., 2024; Fu et al., 2024); clinical PK/PD remains the validation standard.

Agentic workflow

A research workflow driven by software agents that plan, retrieve information, call tools, write code, and coordinate tasks. ChemCrow, Coscientist, and Virtual Lab are the canonical examples (M. Bran et al., 2024; Boiko et al., 2023; Swanson et al., 2025). See Agentic Science Workflows chapter.

AI-ready data

Data with sufficient metadata, provenance, structure, quality control, and governance for reliable machine learning use. NIH Bridge2AI was created around this gap (NIH Common Fund Bridge2AI, 2026). See Biological Data Infrastructure chapter.

AlphaFold

The DeepMind protein structure prediction system. AlphaFold 2 (Jumper et al., 2021) effectively solved single-chain protein structure prediction at CASP14. AlphaFold 3 (Abramson et al., 2024) extended to biomolecular interactions including nucleic acids, ions, and ligands.

AlphaFold DB

The AlphaFold Protein Structure Database hosted by EMBL-EBI, containing over 214 million predicted protein structures (Varadi et al., 2024).

AlphaGenome

A DeepMind model for regulatory variant effect prediction at megabase context (Avsec et al., 2026). Extends the Enformer lineage.

AlphaMissense

A DeepMind classifier for missense variant pathogenicity using AlphaFold-derived structural features and protein language model evidence (Cheng et al., 2023). Classifies about 89% of human missense substitutions; positioned as a research tool, not a clinical classifier.

AlphaProteo

A DeepMind generative model for de novo protein binder design (Zambaldi et al., 2024, preprint). Reports strong binder generation but remains an unpublished arXiv preprint with restricted code.

Analytical validity

Whether a test, assay, or measurement system reliably measures what it claims to measure. In diagnostics and biomarker translation, analytical validity is separate from clinical validity and clinical utility.

Antisense oligonucleotide

Short synthetic nucleic-acid therapeutic designed to bind RNA and alter splicing, degradation, or translation. AI can support target selection and sequence design, but delivery, tissue exposure, off-target effects, and clinical endpoints remain decisive.

B

Backbone generation

The protein design task of producing a three-dimensional protein backbone without final amino-acid identities. RFdiffusion is the canonical method (Watson et al., 2023). Sequence design (ProteinMPNN, LigandMPNN) follows.

Base editing

Genome-editing approach that changes one base class to another without relying on a double-strand break. Base-editing claims require edit-efficiency, bystander-editing, DNA and RNA off-target, delivery, durability, and safety evidence.

Benchmark leakage

Evaluation failure where training data or related examples contaminate a test set. Includes homolog leakage (proteins), scaffold leakage (molecules), donor leakage (cells), and time leakage (sequences). Biology-aware splits address this.

BEST resource

FDA-NIH Biomarkers, EndpointS, and other Tools resource. Provides standard biomarker and endpoint terminology, including diagnostic, prognostic, predictive, pharmacodynamic, safety, monitoring, and risk biomarkers.

BICAN

Brain Initiative Cell Atlas Network. NIH BRAIN Initiative effort to build reference cell atlases across human, non-human primate, and mouse brain resources. In neuroscience AI, BICAN matters because brain models need anatomical and cell-type reference layers, not only recordings.

Biomarker

Defined characteristic measured as an indicator of biological process, disease process, exposure, pharmacologic response, safety, prognosis, prediction, or surrogate endpoint, using the BEST glossary framing (FDA, 2026). Context of use determines evidentiary requirements.

Biomanufacturing

Manufacturing of biological products or biologically produced materials, including biologics, cell therapies, enzymes, vaccines, fermentation products, and synthetic-biology outputs.

Biomedical knowledge graph

A graph that represents biomedical entities, such as genes, diseases, drugs, pathways, assays, publications, and phenotypes, as nodes connected by typed edges. Useful only when relation type, source provenance, and evidence level stay visible.

BioBERT

BERT adapted to biomedical text for biomedical named-entity recognition, relation extraction, and question answering (Lee et al., 2020).

Boltz

Open biomolecular interaction model in the AlphaFold-3-class family. Boltz models are important because they make AF3-style workflows more accessible, but preprint status, benchmark scope, and downstream validation still determine how much weight to give a prediction.

Bridge2AI

The NIH Common Fund program for AI-ready biomedical data and workforce development. Funds four grand-challenge data generation projects: CHORUS, CM4AI, VOICE, AI-READI (NIH Common Fund Bridge2AI, 2026).

BrainLM

Foundation-model approach for fMRI recordings using self-supervised masked prediction on large brain-activity datasets (Ortega Caro et al., 2024). Demonstrates representation learning for brain activity, not general brain reasoning.

C

CAMEO

Continuous Automated Model EvaluatiOn. Ongoing protein structure prediction evaluation resource. Unlike CASP’s biennial blinded rounds, CAMEO supports continuous benchmarking against newly released experimental structures.

CASP

The Critical Assessment of protein Structure Prediction, a biennial blinded benchmark for protein structure prediction since 1994. CASP15 covers structures, complexes, RNA, and ligand binding (Kryshtafovych et al., 2024).

CAR-T

Chimeric antigen receptor T-cell therapy. A patient’s or donor’s T cells are engineered to recognize a target antigen. AI may help with target selection, receptor design, product characterization, toxicity modeling, and manufacturing analytics, but clinical readiness depends on potency, persistence, safety, antigen biology, and patient evidence.

Cell Painting

A high-content imaging assay using six multiplexed fluorescent dyes to stain eight cellular components. Original protocol by Bray et al., 2016; updated protocol by Cimini et al., 2023.

CellProfiler

Open-source image-analysis software for biology (Carpenter et al., 2006; CellProfiler 3.0 by McQuin et al., 2018).

Cellpose

Generalist cell-segmentation algorithm for microscopy images (Stringer et al., 2021).

Cellpose3

Restoration-aware Cellpose release that pairs image restoration with cellular segmentation (Stringer et al., 2025). Useful for degraded microscopy images when the downstream biological measurement is validated.

CEBRA

Contrastive embedding method for learning latent representations from neural and behavioral data (Schneider et al., 2023). In neuroscience AI, CEBRA is evidence for useful neural-behavioral representation learning, not proof of causal mechanism.

ChEMBL

Curated database of bioactivity data for drug-like molecules, hosted by EMBL-EBI (Zdrazil et al., 2024).

Chai-1

Open biomolecular interaction model in the AlphaFold-3-class family. Chai-1 should be read as a useful structure-prediction tool whose output still depends on confidence metrics, target class, ligand or nucleic-acid context, and experimental validation.

ChemCrow

An LLM agent wrapped around chemistry-specific tools (M. Bran et al., 2024). Canonical example of agentic chemistry orchestration.

CHANGE-seq

Genome-wide assay for discovering CRISPR-Cas9 off-target cleavage sites (Lazzarotto et al., 2020). In gene-editing AI, computational off-target prediction should feed empirical assays such as CHANGE-seq rather than replace them.

CIRCLE-seq

In vitro method for sensitive genome-wide nomination of CRISPR off-target cleavage sites (Tsai et al., 2017). It helps define a risk-discovery method for genome-editing programs.

Clinical utility

Evidence that a diagnostic, biomarker, or model improves a decision or outcome in the intended context of use. Clinical utility is separate from analytical validity and association with disease.

Closed-loop experimentation

Workflow where measurements update a model that selects future experiments. The defining feature of a self-driving laboratory.

Cloud lab

Remote laboratory platform where users submit protocols or experimental requests through software and the work is executed on automated or semi-automated infrastructure. Cloud labs are useful only when protocol calibration, sample handling, acceptance criteria, and failure logs are explicit.

Combination therapy

Therapeutic strategy using multiple drugs or modalities together. Computational combination design ranks candidates by predicted additive, antagonistic, or synergistic effects, but dose, schedule, toxicity, and tissue context determine usefulness.

COMPSS

Protein design workflow that combines computed protein sequence statistics with structure generation and experimental validation (Johnson et al., 2024). Useful as evidence that design models still need laboratory build, assay, and fitness measurement.

Companion diagnostic

Diagnostic test used to identify patients for whom a therapeutic product is expected to be beneficial or unsafe. FDA describes a companion diagnostic as a device that provides information essential for the safe and effective use of a corresponding therapeutic product (FDA, 2026). Companion-diagnostic claims require a specific regulatory and clinical context.

CONCH

Pathology vision-language foundation model for computational pathology (Lu et al., 2024). Supports slide and text representation tasks; diagnostic deployment still requires intended-use validation.

Connectomics

Mapping neural connectivity at cellular, synaptic, circuit, or whole-brain scale. Connectomics data can support brain-model training and circuit analysis, but wiring alone does not explain behavior without physiology, task, and organism context.

CONSORT-AI

Extension of CONSORT reporting guidance for clinical trials that evaluate AI interventions (Liu et al., 2020). It improves reporting transparency; it does not prove the AI intervention works.

Connectivity Map

Perturbational gene-expression reference, including the L1000 platform, used for mechanism and drug-repurposing hypotheses (Subramanian et al., 2017).

Coscientist

A GPT-4-driven autonomous chemistry system (Boiko et al., 2023). Demonstrated LLM-planned Pd-catalysed cross-coupling chemistry.

CPA

Compositional Perturbation Autoencoder. Single-cell perturbation model designed to represent perturbation identity, dose, and cellular context (Lotfollahi et al., 2023).

Counterfactual response

Predicted biological state under an intervention that was not directly observed for that cell, donor, condition, or perturbation. In virtual-cell work, counterfactual response claims require explicit held-out design and measured perturbation anchors.

Critical process parameter

Process variable whose variability may affect a critical quality attribute and should be monitored or controlled during manufacturing.

Critical quality attribute

Physical, chemical, biological, or microbiological property that should stay within an appropriate range to ensure product quality.

CRISPR

Genome-editing system adapted from bacterial immune mechanisms. In this handbook, CRISPR appears as an experimental perturbation tool, a therapeutic editing platform, and a source of safety questions around off-target editing, on-target genotoxicity, delivery, and long-term follow-up.

CZ CELLxGENE

The Chan Zuckerberg Initiative single-cell data platform with approximately 100 million curated single-cell observations.

CryoDRGN

Neural-network method for reconstructing heterogeneous cryo-EM structures (Zhong et al., 2021).

Cryo-EM

Cryogenic electron microscopy. Experimental structure-determination modality that images vitrified biological samples and supports molecular structure, complex, and conformational-state analysis.

cryoSPARC

Cryo-EM structure-determination workflow and algorithm suite (Punjani et al., 2017).

crYOLO

Automated particle picker for cryo-EM micrographs (Wagner et al., 2019).

D

DANDI Archive

BRAIN Initiative archive for neurophysiology data, including electrophysiology, optophysiology, behavioral time series, and related metadata. In neuroscience AI, DANDI matters because model reuse depends on raw signals, stimulus timing, behavioral context, and preprocessing details.

Deep-STORM

Deep-learning method for single-molecule super-resolution microscopy (Nehme et al., 2018).

DeepSynergy

Deep-learning method for predicting anti-cancer drug synergy from drug and cell-line features (Preuer et al., 2018).

DiffDock

A diffusion-based blind docking model (Corso et al., 2023, preprint). PoseBusters showed many DiffDock poses fail physical validity checks despite low RMSD.

Diagnostic biomarker

Biomarker used to detect or confirm presence of a disease, condition, or biological state. Diagnostic use requires context of use, analytical validity, clinical validity, thresholding, comparator, and intended-use population.

Digital twin

Computational representation of a physical process, product, or system connected to data streams, assumptions, and decision logic. In biomanufacturing, twin validity depends on process boundary, calibration, and update discipline.

Drug repurposing

Finding a new indication, mechanism, patient subset, or therapeutic use for an existing compound. AI methods support prioritisation through signatures, graphs, networks, and prior evidence; efficacy still requires validation in the new context.

DOME framework

Data, Optimization, Model, Evaluation: a community reporting checklist for supervised ML in biology (Walsh et al., 2021). Professional minimum for biology ML reporting.

DURC

Dual-use research of concern. US framework for identifying and reviewing life sciences research that could be directly misused. The 2024 USG DURC/PEPP policy expanded the oversight context for life sciences research, including AI-enabled work (NIH, 2024).

E

eDNA

Environmental DNA. Genetic material recovered from water, soil, air, sediment, or other environmental samples. eDNA can support biodiversity detection, but abundance, local presence, and living population claims require sampling design, controls, reference databases, and field validation.

Enformer

A transformer model that predicts gene expression and chromatin states from long-range DNA sequence context (Avsec et al., 2021).

ESM

The Evolutionary Scale Modeling protein language model lineage from Meta FAIR (now EvolutionaryScale). ESM-1b (Rives et al., 2021), ESM-2 (Lin et al., 2023), ESM-3 (Hayes et al., 2025).

ESMFold

Protein structure prediction from sequence alone via the ESM-2 protein language model (Lin et al., 2023).

EVE

Evolutionary model of Variant Effect: unsupervised missense pathogenicity prediction from protein family multiple sequence alignments (Frazer et al., 2021).

EVOLVEpro

Few-shot active-learning method that uses protein language embeddings to guide protein activity optimization from small experimental datasets (Jiang et al., 2025).

Evo

A genome-scale foundation model from Arc Institute with collaborators (Nguyen et al., 2024). 7B parameters, reasons across DNA, RNA, and protein. Evo 2 (Brixi et al., 2026) extends to 40B parameters across all domains of life.

F

FAIR principles

Data stewardship principles: findable, accessible, interoperable, and reusable (Wilkinson et al., 2016). In life sciences AI, FAIR is necessary but not sufficient; model-ready data also need assay context, provenance, negative results, and measured-versus-computed label discipline.

Flux balance analysis

Constraint-based method for modeling metabolic network behavior under mass-balance and objective assumptions. Flux balance analysis can support metabolic engineering and systems biology, but predictions depend on the network reconstruction, constraints, objective function, and measured context.

Foundation model

A neural network pretrained on broad data with a self-supervised objective, then adapted to many downstream tasks via fine-tuning or zero-shot transfer. ESM, scGPT, Geneformer, Evo, and AlphaFold are biology examples (Lin et al., 2023; Cui et al., 2024; Theodoris et al., 2023; Nguyen et al., 2024; Jumper et al., 2021).

G

GEARS

Graph-enhanced gene activity transformer for transcriptional perturbation prediction (Roohani et al., 2024). Canonical method for single-cell perturbation prediction; see Ahlmann-Eltze 2025 critique alongside.

Geneformer

Single-cell foundation model from MGH/Broad (Theodoris et al., 2023). Pretrained on approximately 30 million single cells using rank-order gene encoding.

Gene regulatory network

Model of regulatory relationships among genes, transcription factors, enhancers, and cell states. A gene regulatory network edge is a hypothesis about regulation, not proof of direct causal control unless supported by perturbation or orthogonal evidence.

Genomic selection

Breeding method that uses genome-wide marker information to estimate genetic value and guide selection decisions (Meuwissen et al., 2001). In crop AI, genomic selection is a breeding-system tool, not just a prediction model.

GUIDE-seq

Genome-wide method for identifying CRISPR double-strand break sites in living cells (Tsai et al., 2015). Like CIRCLE-seq and CHANGE-seq, it provides empirical off-target evidence that computational guide scores cannot replace.

GPN-MSA

Multispecies-alignment DNA language model for coding and noncoding variant effect prediction (Benegas et al., 2025). Useful for evolutionary-context scoring, not a clinical classifier by itself.

H

Hetionet

Biomedical knowledge graph integrating drugs, diseases, genes, pathways, anatomy, and side effects for drug-repurposing prioritisation (Himmelstein et al., 2017).

Histopathology AI

Use of digitised tissue slides for representation learning, segmentation, tissue phenotyping, and biomarker research. Clinical diagnostic use requires separate validation and governance.

HoVer-Net

Histology method for simultaneous segmentation and classification of nuclei in multi-tissue images (Graham et al., 2019).

Human Cell Atlas

International effort to map human cell types, states, and tissue contexts. In this handbook, the Human Cell Atlas is part of the data infrastructure behind single-cell and tissue foundation models.

I

IGoR

Intelligent Generator of Research: ARPA-H program for AI-powered biomedical research infrastructure focused on Alzheimer’s, Parkinson’s, and autoimmune disease (ARPA-H IGoR, 2026).

Indel

A short added or deleted segment of DNA. A class of genomic variants that AI variant predictors handle differently from substitutions.

Information hazard

Risk created by sharing true information that could enable misuse. In life sciences AI, information-hazard review should preserve verification where possible while avoiding unnecessary operational detail.

Intended use

The specific purpose, population, setting, and decision for which a tool, assay, model, or diagnostic is used. Evidence that supports one intended use does not automatically support another.

Inverse folding

The protein design task of proposing an amino-acid sequence for a target backbone structure. ProteinMPNN and LigandMPNN are the canonical inverse-folding tools.

ipTM

Interface predicted TM-score: confidence metric for protein-protein complex predictions introduced in the AlphaFold-Multimer lineage (Evans et al., 2021, preprint). ipTM > 0.8 suggests reliable interface; ipTM < 0.5 should be treated as hypothesis-grade.

J

JUMP morphology maps

Public perturbation morphology datasets from the Joint Undertaking in Morphological Profiling. JUMP-era data connect Cell Painting profiles to chemical and genetic perturbations and give image-based phenotyping methods a shared benchmark layer.

L

Literature AI

Software for biomedical search, triage, entity extraction, relation extraction, evidence retrieval, and citation-grounded answer drafting over scientific corpora. High-stakes use requires source review and claim-to-evidence mapping.

LigandMPNN

Ligand- and modification-aware sequence design model from the Baker lab (Dauparas et al., 2025). Extends ProteinMPNN to atomic ligand context.

LNP

Lipid nanoparticle. Non-viral delivery platform used for RNA and some genome-editing payloads. AI may help formulation triage, but tissue distribution, endosomal escape, dose, repeat dosing, tolerability, and manufacturing reproducibility determine utility.

M

MERFISH

Multiplexed fluorescence in situ hybridization method with error-correcting barcodes. Spatial transcriptomics method that measures many RNA species in tissue or cells while preserving spatial coordinates.

Metabarcoding

High-throughput sequencing approach that identifies taxa from environmental or mixed samples using marker regions. In ecological AI, metabarcoding is useful for biodiversity detection but depends on sampling design, primer choice, contamination control, and reference databases.

Metabolomics

Measurement of small molecules and metabolites in cells, tissues, biofluids, organisms, or environments. AI can support peak annotation, pathway inference, and multi-omic integration, but metabolite identity and assay provenance are frequent bottlenecks.

Microbiome

Community of microorganisms and their genes, products, and ecological interactions in a host or environment. Microbiome AI must account for compositional data, sequencing protocol, batch effects, diet, geography, medication, and host context.

MoleculeNet

Benchmark suite for molecular machine learning that standardised scaffold-split benchmarks (Wu et al., 2018).

mRNA therapeutic

Therapeutic or vaccine modality that delivers messenger RNA so cells produce an encoded protein. AI can support sequence design, codon choice, stability prediction, antigen selection, and delivery optimization; immune response, LNP behavior, dose, and manufacturing remain decisive.

Multi-omics

Integration of multiple molecular data layers, such as genomics, transcriptomics, epigenomics, proteomics, metabolomics, microbiome, imaging, and clinical data. Integration is useful when samples, timing, missingness, provenance, and endpoint are handled explicitly.

N

Neural optimal transport

Method class that models how one distribution of single-cell states moves to another under a perturbation. Used for counterfactual perturbation-response prediction in single-cell data (Bunne et al., 2023).

Nicheformer

Spatially aware single-cell and spatial-omics foundation model trained on dissociated and spatial transcriptomics data (Tejada-Lapuerta et al., 2025). It supports spatial-context representation, not full tissue-response simulation.

Noise2Void

Self-supervised denoising method that learns from single noisy images (Krull et al., 2019).

Novae

Graph-based foundation model for spatial transcriptomics data (Blampey et al., 2025). Supports spatial-domain representation, not tissue-response simulation.

O

OpenProblems

Community benchmark suite for single-cell analysis covering integration, perturbation, label transfer, and other tasks (Luecken et al., 2025).

Organoid

Three-dimensional biological model grown from stem cells or tissue-derived cells to recapitulate selected features of an organ or tissue. Organoids can strengthen validation for spatial, perturbation, and therapeutic claims, but they are not full organisms.

Orthrus

Mature RNA foundation model trained with evolutionary and isoform-based contrastive objectives (Fradkin et al., 2026).

P

PAE

Predicted Aligned Error: per-residue-pair confidence metric from AlphaFold and successors (Jumper et al., 2021; Abramson et al., 2024). Reveals inter-domain orientation confidence separately from intra-domain confidence.

PDB

Protein Data Bank: the global archive for experimentally determined macromolecular structures, since 1971 (wwPDB, 2026).

Perturbation prediction

Forecasting how a biological system changes after a genetic, chemical, or environmental intervention. GEARS, scGen, CPA, scPRAM are canonical methods (Roohani et al., 2024; Lotfollahi et al., 2019; Lotfollahi et al., 2023). Independent evaluation shows linear baselines often match deep methods (Ahlmann-Eltze et al., 2025).

Perturb-CITE-seq

Perturbation-screening method that combines pooled perturbations with paired single-cell RNA and protein readouts. Useful when a virtual-cell claim depends on protein-state changes as well as transcriptomic response (Frangieh et al., 2021).

Perturb-seq

Pooled perturbation method that links genetic perturbations to single-cell RNA readouts (Dixit et al., 2016). Perturb-seq is a central data layer for virtual-cell, systems-biology, and perturbation-prediction claims.

Phenomics

High-throughput measurement of observable traits. In plant, cell, and organismal AI, phenomics connects genotype, environment, image data, physiology, and measured performance.

pLDDT

Predicted Local Distance Difference Test: per-residue confidence score from the AlphaFold lineage (Jumper et al., 2021). 0-100 scale: above 90 is high confidence; 70-90 confident backbone; 50-70 low confidence; below 50 unreliable (often disordered).

PoseBusters

A docking-pose validity benchmark (Buttenschoen et al., 2024) showing many low-RMSD deep-learning docking poses fail basic physical and chemical validity checks.

Prime editing

Genome-editing approach that writes specified small sequence changes using a Cas nickase fused to reverse transcriptase and a prime-editing guide RNA (Anzalone et al., 2019). Therapeutic claims require delivery, efficiency, bystander, off-target, durability, and safety evidence.

PrimeKG

Precision-medicine knowledge graph connecting diseases, drugs, genes, phenotypes, and biological context for biomedical evidence mapping (Chandak et al., 2023).

Predictive biomarker

Biomarker associated with differential treatment effect. Predictive claims require evidence that treatment benefit or harm differs across biomarker-defined groups.

Prognostic biomarker

Biomarker associated with outcome regardless of treatment. Prognostic value does not by itself establish treatment selection value.

Process analytical technology

FDA manufacturing framework for designing, analyzing, and controlling processes through timely measurement of critical process and quality attributes (FDA, 2004).

Proteomics

Large-scale measurement of proteins, proteoforms, modifications, abundance, localization, or interactions. Proteomics often supplies the functional layer missing from transcript-only models.

ProteinGenerator

Sequence-structure diffusion model for protein generation (Lisanza et al., 2024). Generates protein candidates; experimental validation determines function and utility.

Prov-GigaPath

Whole-slide foundation model for digital pathology trained from real-world pathology data (Xu et al., 2024).

ProteinMPNN

Inverse-folding sequence design model (Dauparas et al., 2022). The de facto sequence-design standard before LigandMPNN.

PubMedBERT

Biomedical language model pretrained from scratch on PubMed text rather than adapted from general-domain text (Gu et al., 2022).

R

Real-world data

Data relating to patient health status or health care delivery routinely collected from sources such as EHRs, claims, registries, devices, imaging, genomics, or patient-reported outcomes (FDA, 2026).

Real-world evidence

Clinical evidence about medical product use and potential benefits or risks derived from analysis of real-world data (FDA, 2026). Regulatory usefulness depends on data quality, design, analysis, and context of use.

Retrieval-augmented generation

Answer drafting pattern that retrieves source passages before generating text. In biomedical settings, the retrieved sources, date, corpus boundary, and evidence type matter as much as the generated answer.

REMARK

Reporting recommendations for tumor-marker prognostic studies. In biomarker AI, REMARK is a reminder that association claims need transparent cohorts, assays, endpoints, statistics, and validation.

RFantibody

Extension of RFdiffusion to de novo antibody design (Bennett et al., 2026). Strongest current peer-reviewed example for AI antibody design.

RFdiffusion

Diffusion-based protein backbone design model from the Baker lab (Watson et al., 2023). Fine-tuned from RoseTTAFold weights.

RoseTTAFold

Three-track protein structure prediction architecture from the Baker lab (Baek et al., 2021). Released open-source as an AlphaFold-2-class alternative.

S

Scaffold split

Split of a molecular dataset by Bemis-Murcko scaffold (central ring system), assigning whole scaffolds to either train or test. Prevents leakage of analogues. MoleculeNet standardised this discipline (Wu et al., 2018).

scBERT

Transformer-based single-cell model that preceded the larger scFM wave (Yang et al., 2022). Useful as an early baseline for cell-state representation learning.

scFM

Single-cell foundation model. scGPT, Geneformer, scFoundation, scBERT are canonical examples.

scFoundation

Single-cell foundation model trained at larger scale than earlier models with a different pretraining recipe (Hao et al., 2024). Like other scFMs, its strongest use is representation and selected transfer, not unrestricted perturbation prediction.

scGen

Variational autoencoder method for modeling single-cell perturbation transfer (Lotfollahi et al., 2019). Important historical baseline for virtual-cell and perturbation-prediction work.

scGPT

Generative pretrained transformer for single-cell multi-omics (Cui et al., 2024). Trained on over 33 million single cells.

scIB

Single-cell integration benchmark (Luecken et al., 2022). Combines batch-correction quality and biological-signal preservation.

SciBERT

BERT-style language model pretrained on scientific text, including biomedical papers (Beltagy et al., 2019).

Self-driving laboratory

A laboratory that closes the experimental loop: a model proposes an experiment, automated hardware runs it, the result returns to the model, the model proposes the next experiment. Coscientist, Virtual Lab, A-Lab, Ada are examples. See Self-Driving Laboratories chapter.

Segment Anything for Microscopy

Microscopy segmentation foundation model adapted from the Segment Anything pattern (Archit et al., 2025).

SemMedDB

PubMed-scale repository of subject-predicate-object semantic predications extracted from biomedical text (Kilicoglu et al., 2012).

siRNA

Small interfering RNA. RNA modality that reduces expression of target transcripts through RNA interference. AI can support sequence choice and off-target review, but delivery, tissue exposure, durability, and safety determine therapeutic usefulness.

Spatial omics

Methods that measure molecular features while preserving spatial location in tissue. Spatial omics includes spatial transcriptomics, spatial proteomics, imaging-based platforms, and integrated tissue atlases.

SpatialData

Open data framework for spatial omics preserving images, shapes, labels, molecular tables, and coordinate systems (Marconato et al., 2024).

SPIRIT-AI

Extension of SPIRIT protocol reporting guidance for clinical trials with AI components (Cruz Rivera et al., 2020). It supports transparent trial planning; it is not an efficacy standard.

SpliceAI

CNN-based splice-altering variant predictor (Jaganathan et al., 2019). Widely used in clinical genetics for splice-region variants.

STARD

Standards for Reporting Diagnostic Accuracy Studies. In diagnostic AI and biomarker translation, STARD helps readers evaluate whether the population, test, reference standard, and accuracy estimates are interpretable.

SynergyFinder

Visual analytics platform for multi-drug combination synergies (Ianevski et al., 2020).

Surrogate endpoint

Endpoint used as a substitute for a direct clinical endpoint in a defined context. Validation depends on the relationship between the surrogate and the clinical outcome.

Systems biology

Study and modeling of interacting biological components across genes, pathways, cells, tissues, organisms, and environments. In this handbook, systems-biology claims are separated by evidence type: networks, perturbation data, mechanistic models, multiscale simulation, and prospective validation.

T

Target trial emulation

Design strategy that specifies the randomized trial one would like to run, then emulates it using observational or real-world data when randomization is unavailable (Hernán et al., 2022).

TCR therapy

T-cell receptor therapy. Immune-cell therapy in which T cells are engineered with receptors that recognize peptide-HLA complexes. AI may help with target, receptor, and cross-reactivity analysis, but antigen processing, HLA restriction, toxicity, and clinical evidence govern translation.

Therapeutics Data Commons

Multi-task benchmark for therapeutic discovery covering small molecules, antibodies, vaccines, gene editing, and clinical trials (Huang et al., 2022).

Three-tier evidence framework

The handbook discipline of placing every capability claim in one of three tiers: Demonstrated (peer-reviewed or reproducible benchmark evidence), Theoretical (plausible but not yet established), Beyond current capabilities (not supported by current evidence).

Topaz

Positive-unlabeled convolutional neural network method for cryo-EM particle picking (Bepler et al., 2019).

TxGNN

Graph foundation model for clinician-centered zero-shot drug repurposing (Huang et al., 2024). Supports prioritization, not efficacy.

U

UCE

Universal Cell Embedding. Single-cell foundation-model approach for representing cell states across datasets and organisms. As with other scFMs, utility depends on pretraining coverage, split design, baselines, and validation task.

UNI

Computational pathology foundation model trained for general slide representation tasks (Chen et al., 2024). Useful for pathology research and biomarker discovery when external validation matches the intended use.

V

Virchow

Whole-slide pathology foundation model used for computational pathology representation and rare-cancer work (Vorontsov et al., 2024). Its evidence should be read by task, site, stain, scanner, and intended use.

Virtual cell

A computational model intended to predict cell behaviour, usually for defined outputs rather than complete cellular simulation. See Perturbation Prediction and Virtual Cells chapter.

Virtual Lab

Stanford multi-agent biology research system (Swanson et al., 2025). Demonstrated multi-agent design of SARS-CoV-2 nanobodies with experimental validation.

Virtual Cell Challenge

Benchmark and field-framing effort for virtual-cell models focused on perturbation-response prediction (Roohani et al., 2025). It matters because virtual-cell claims need shared tasks, held-out designs, and measured biological endpoints.

Virtual organism

Computational model intended to predict organism-level phenotypes, development, physiology, behavior, or response under named conditions. A virtual organism is not a larger virtual cell; it must define organism, genotype, environment, life stage, intervention, and measurable phenotype.

Visium

10x Genomics spatial transcriptomics platform that measures transcript abundance across tissue sections while preserving spatial coordinates at spot-level resolution. Evidence claims depend on tissue handling, spot size, capture efficiency, and downstream segmentation or deconvolution assumptions.

W

Whole-cell model

Mechanistic or hybrid model that attempts to represent many cellular subsystems together. Whole-cell models are important for systems-biology thinking, but broad prediction remains limited by measurement coverage, parameter uncertainty, context, and perturbation validation.

X

Xenium

In situ spatial transcriptomics platform from 10x Genomics. Xenium measures selected RNA targets with subcellular spatial information; interpretation depends on panel design, tissue quality, segmentation, and validation against the biological question.