Life and medicine

Protein folding and dynamics

A protein as tens of thousands of atoms, shaking, folding and changing shape in water.

Question it answers: How does this chain of amino acids fold into a functional three-dimensional structure?

Leaves out: Even long runs cover microseconds at most, far short of many real folding times.

Runs on: about 10¹⁶ FLOPS rough a GPU cluster running microseconds of all-atom dynamicsSince: 1977 McCammon, Gelin and Karplus's first protein dynamics simulationIn the simulations tree: Protein dynamics

Molecular docking and drug binding

Candidate drug molecules tried in the pocket of a target protein.

Question it answers: Which of a million molecules will stick to this target?

Leaves out: Docking scores are approximate; most predicted binders never work in the lab.

Runs on: about 10¹⁵ FLOPS rough a university cluster docking millions of candidate moleculesSince: 1982 Kuntz and colleagues' DOCK, the first automated molecular docking program

Membranes and ion channels

The fatty skin of a cell and the protein gates that let ions through it.

Question it answers: How does a nerve cell open and close its channels?

Leaves out: Real channels gate over milliseconds, far longer than most simulations reach.

Runs on: about 10¹⁶ FLOPS rough a GPU cluster simulating a lipid bilayer and its channelsSince: 1982 van der Ploeg and Berendsen's dynamics of a lipid bilayer

Whole virus particles

Every atom of a virus shell, sometimes a hundred million of them, simulated at once.

Question it answers: How does the virus hold together, and how does it come apart inside a cell?

Leaves out: Runs cover nanoseconds to microseconds, far short of a real infection cycle.

Runs on: about 10¹⁷ FLOPS rough a national supercomputer simulating a hundred-million-atom shellSince: 2006 Freddolino and Schulten's simulation of a satellite tobacco mosaic virus

RNA folding and structure

An RNA strand as a chain of bases pairing with itself, folding into a specific three-dimensional shape rather than the straight double helix of DNA.

Question it answers: How does a strand of RNA settle into the shape that enables it to act as a catalyst or a regulatory switch?

Leaves out: Predicted structures often need experimental correction, and function does not follow automatically from shape.

Runs on: about 10¹² FLOPS rough a desktop workstation running free-energy folding softwareSince: 1981 Zuker and Stiegler's minimum-free-energy RNA folding algorithm

Enzyme catalysis

The active site of an enzyme, its substrate and the surrounding protein, modelled with quantum mechanics where bonds break and classical mechanics everywhere else.

Question it answers: How does this enzyme speed up a reaction by a factor of a billion?

Leaves out: Only a small part of the protein gets the expensive quantum treatment, and where to draw that line is a modelling choice.

Runs on: about 10¹⁵ FLOPS rough a university cluster running combined quantum and classical calculationsSince: 1976 Warshel and Levitt's combined quantum and classical model of an enzyme reaction

Coarse-grained DNA and RNA models

Each nucleotide is a rigid body with a few interaction sites, allowing simulation of folding, hybridisation and nanostructures.

Question it answers: How do strands pair, fold and assemble into designed structures?

Leaves out: Atomic detail and ions are replaced by effective interactions.

Runs on: about 10¹⁴ FLOPS rough Millions of nucleotides on GPUsSince: 2011 oxDNA, a coarse-grained DNA model (2011) source

Learned protein structure prediction

A neural network trained on known structures and evolutionary data predicts the folded shape of a protein from its sequence.

Question it answers: What three-dimensional shape does this amino-acid chain fold into?

Leaves out: It gives a static structure, not the folding pathway or dynamics.

Runs on: about 10¹⁸ FLOPS rough Training at large scale; minutes per predictionSince: 2018 Improved protein structure prediction using potentials from deep learning (2020, CASP13 entry 2018) source

Generative protein ensembles

A generative model trained on structures and simulations produces many conformations of a protein rather than a single shape.

Question it answers: Which alternative shapes can a protein adopt?

Leaves out: Populations of states are only as good as the training simulations.

Runs on: about 10¹⁶ FLOPS rough Training plus cheap samplingSince: 2024 AlphaFold meets flow matching for generating protein ensembles (2024) source

Volunteer-computed molecular dynamics

Thousands of independent short trajectories of one molecule run on donated home computers and are combined to see folding events.

Question it answers: Can many short runs together reveal a slow molecular process?

Leaves out: Trajectories are short and independent, so statistics must be stitched together.

Runs on: about 10¹⁷ FLOPS rough Aggregate of volunteer machinesSince: 2000 Screen savers of the world unite! (2000), the distributed folding project source

Stochastic biochemistry

Reactions inside a cell fired one at a time at random, because there are too few molecules to average.

Question it answers: Why do genetically identical cells behave differently?

Leaves out: Usually assumes the cell's interior is well mixed, which often does not hold.

Runs on: about 10¹² FLOPS rough a desktop workstation running a stochastic reaction simulationSince: 1977 Gillespie's exact stochastic simulation algorithm for chemical kinetics

Gene regulatory networks

Genes switching each other on and off, as a network of interactions.

Question it answers: How does a cell decide what kind of cell to become?

Leaves out: Most real regulatory networks are only partly known, so links are often inferred.

Runs on: about 10¹² FLOPS rough a desktop workstation simulating a gene network's dynamicsSince: 1969 Kauffman's random Boolean networks of gene regulation

Metabolic networks

Every chemical reaction a cell uses to turn food into energy and parts, balanced as a network of fluxes.

Question it answers: Which gene should we switch off to make this microbe produce more of a useful compound?

Leaves out: Assumes a steady state and an optimal objective a real cell may not pursue.

Runs on: about 10¹² FLOPS rough a desktop workstation solving a genome-scale flux balance modelSince: 1986 early flux balance analysis of a cell's metabolic network

Whole-cell models

Every gene and every class of molecule in a single bacterium, followed through one full division cycle.

Question it answers: Can we predict a living cell from its parts list?

Leaves out: Only one simple bacterium has ever been modelled this completely.

Runs on: about 10¹⁴ FLOPS rough a university cluster running one division cycle of a bacteriumSince: 2012 Karr and colleagues' whole-cell model of Mycoplasma genitalium

Pattern formation

Chemicals diffusing and reacting across a growing tissue, the mechanism Alan Turing proposed for stripes and spots.

Question it answers: How do reaction-diffusion mechanisms account for animal coat patterns and for digit formation in a limb?

Leaves out: Real tissues use mechanical forces and more signals than a two-chemical model.

Runs on: about 10¹² FLOPS rough a desktop workstation solving reaction-diffusion equations on a gridSince: 1952 Turing's reaction-diffusion computations for The Chemical Basis of MorphogenesisIn the simulations tree: Turing’s reaction-diffusion

Run the interactive model

Tissue and tumour growth

Individual cells dividing, moving, competing for oxygen and dying.

Question it answers: How does a tumour grow, and how might a treatment schedule stop it?

Leaves out: Division and death rates inside a real tumour are hard to measure directly.

Runs on: about 10¹³ FLOPS rough a university cluster tracking thousands of individual dividing cellsSince: 1993 early cellular automaton models of solid tumour growth

Cell perturbation-response models

A model trained on single-cell data predicts how a cell's gene expression changes after a drug or gene edit it was not measured with.

Question it answers: What will a cell do after an intervention nobody has yet tried?

Leaves out: Correlational; it extrapolates poorly to cell types or drugs far from training data.

Runs on: about 10¹⁶ FLOPS rough Large single-cell training setsSince: 2019 scGen predicts single-cell perturbation responses (2019) source

Tumour evolution agent-based models

Individual cancer cells on a grid divide, mutate and compete in a changing microenvironment of oxygen and acid.

Question it answers: How do tumour shape and aggressive traits emerge from selection pressure?

Leaves out: Rules are simplified, and blood supply and immune cells are often absent.

Runs on: about 10¹¹ FLOPS rough Millions of cells on one nodeSince: 2006 Tumor morphology and phenotypic evolution driven by selective pressure from the microenvironment (2006) source

Microbial community metabolic models

Genome-scale metabolic networks of several species are joined so that waste from one becomes food for another.

Question it answers: How do gut or soil microbes share nutrients and stabilise together?

Leaves out: Assumes steady state and optimal growth, ignoring regulation and space.

Runs on: about 10¹¹ FLOPS rough Linear programming across many speciesSince: 2007 Metabolic modeling of a mutualistic microbial community (2007) source

Tumour growth and therapy response

A solid tumour treated as a mass fed by diffusing oxygen and nutrient, with a growing proliferating rim and a starved necrotic core, then exposed to radiation or drugs.

Question it answers: How fast does a tumour grow, and how much does a given dose shrink it?

Leaves out: Real tumours vary between patients and within themselves. The model uses averaged parameters and a simplified geometry.

Runs on: about 10¹² FLOPS rough a three-dimensional reaction-diffusion model with treatment schedules over weeks of simulated time.Since: 1972 Greenspan, models for the growth of a solid tumour by diffusion. source

Tumour microenvironment evolution

Individual cancer cells on a grid compete for oxygen and space while gradients of acidity select for new phenotypes, so the tumour evolves as it grows.

Question it answers: Does the surrounding tissue push a tumour towards more invasive cells?

Leaves out: Cell rules are simplified and parameters are hard to measure in living tumours.

Runs on: about 10¹² FLOPS rough hybrid agent-continuum runs over simulated months on a workstation or small cluster.Since: 2006 Anderson and colleagues, hybrid discrete-continuum model of microenvironment-driven phenotypic evolution. source

Tumour-induced blood vessel growth

Endothelial cells sprout and migrate up a chemical gradient released by a tumour, forming a branching vessel network that feeds it.

Question it answers: How does a tumour recruit a blood supply, and what happens if the signal is blocked?

Leaves out: Vessel flow, remodelling and the real chemistry of signalling are heavily simplified.

Runs on: about 10¹¹ FLOPS rough two- or three-dimensional sprouting runs on a workstation.Since: 1998 Anderson and Chaplain, continuous and discrete models of tumour-induced angiogenesis. source

Cancer drug response and resistance dynamics

Populations of normal, sensitive and resistant cancer cells change under drug therapy, so a patient first improves and then may relapse.

Question it answers: How long will a drug control the disease before resistance appears?

Leaves out: Only the compartments that can be measured are modelled, and stem-cell behaviour is inferred.

Runs on: about 10⁸ FLOPS rough ordinary differential equations fitted to patient blood measurements.Since: 2005 Michor and colleagues, dynamics of chronic myeloid leukaemia. source

Blood cell production dynamics

Stem cells, progenitors and mature blood cells are linked by feedback with delays, producing stable counts or oscillations when the control fails.

Question it answers: Why do some blood disorders cycle, and what restores normal production?

Leaves out: Real marrow has spatial structure and many cell types that are lumped.

Runs on: about 10⁶ FLOPS rough delay differential equations on a laptop.Since: 1978 Mackey, a unified hypothesis for aplastic anaemia and periodic haematopoiesis. source

Bacterial colony growth and biofilms

Each bacterial cell is simulated as an individual with its own size, growth and position, feeding on diffusing nutrient inside a colony or biofilm.

Question it answers: How do colonies take their shape and how do cells on the inside survive?

Leaves out: Cell behaviour is simplified and the chemistry of real biofilm matrix is not captured.

Runs on: about 10¹¹ FLOPS rough individual-based growth of millions of cells on a cluster.Since: 1998 Kreft, Booth and Wimpenny, BacSim, a simulator for individual-based modelling of bacterial colony growth. source

The heart

Electrical waves across heart muscle and the squeeze and blood flow they drive.

Question it answers: Will this drug or this device make an irregular heartbeat worse?

Leaves out: Individual variation between real hearts is still hard to capture in one model.

Runs on: about 10¹⁶ FLOPS rough a university cluster running one beat of whole-heart electromechanicsSince: 1960 Noble’s computer model of the action potential of a single Purkinje fibre cell; models of the whole heart came later sourceIn the simulations tree: Cardiac cell model

Blood flow

Blood moving through arteries, veins and devices like stents and valves.

Question it answers: Where is this artery likely to weaken or clog?

Leaves out: Vessel walls are often treated as rigid, which misses real elasticity.

Runs on: about 10¹⁵ FLOPS rough a university cluster running a patient-specific artery flow simulationSince: 1989 Perktold and colleagues' finite-element model of carotid artery flow

Breathing and airways

Air moving through the branching airways and the particles carried in it.

Question it answers: Where do inhaled medicines, or pollutants, end up?

Leaves out: Most models use one idealised airway tree rather than a specific person's.

Runs on: about 10¹⁴ FLOPS rough a university cluster simulating airflow through a branching airway treeSince: 1993 early fluid dynamics models of airflow in the bronchial tree

Musculoskeletal biomechanics

Bones, joints and muscles as a linked mechanical system, walking, lifting and falling.

Question it answers: Which muscles take the strain, and how should this implant be shaped?

Leaves out: Forces inside a real limb are estimated indirectly, not measured.

Runs on: about 10¹² FLOPS rough a desktop workstation solving a musculoskeletal model of one movementSince: 1990 Delp and colleagues' computer models of musculoskeletal motion

Immune response

Immune cells and pathogens meeting, multiplying and signalling inside the body.

Question it answers: Why do some infections clear and others become chronic?

Leaves out: Real immune signalling is far more complex than any model tracks in full.

Runs on: about 10¹³ FLOPS rough a university cluster running an agent-based immune simulationSince: 1992 Celada and Seiden's cellular automaton model of the immune response

Pharmacokinetics

A drug absorbed, spread through the body’s compartments, broken down and cleared.

Question it answers: What dose, how often, for someone of this age and weight?

Leaves out: A population-average model can miss how one particular patient handles a drug.

Runs on: about 10¹⁰ FLOPS rough a desktop workstation solving a compartmental drug modelSince: 1937 Teorell's two-compartment model of drug distribution

In silico clinical trials

Populations of virtual patients with realistic variation, given a treatment before or alongside a real trial.

Question it answers: Who is this treatment likely to help, and who could it harm?

Leaves out: Regulators still require real trials alongside. Virtual patients can't fully replace them.

Runs on: about 10¹⁴ FLOPS rough a university cluster running thousands of virtual patientsSince: 2011 by 2011, early in silico trials submitted alongside evidence for a device

Cardiac digital twins

Heart models fitted to one patient's scans and signals and used to predict how that heart responds to a therapy.

Question it answers: How will this patient's heart respond to a particular treatment?

Leaves out: Fit quality depends on imaging data, and tissue properties are inferred.

Runs on: about 10¹⁵ FLOPS rough Personalised electromechanical runs plus fittingSince: 2020 The 'Digital Twin' to enable the vision of precision cardiology (2020) source

Neural predictive models of sensory brain areas

A neural network trained on recorded responses becomes a stand-in for a brain area, letting researchers test stimuli in silico that excite neurons most.

Question it answers: Which stimulus drives a given neuron hardest?

Leaves out: Predictions need verification on the real tissue and cover only the tested animals.

Runs on: about 10¹⁶ FLOPS rough Training on recordings, then unlimited virtual trialsSince: 2019 Inception loops discover what excites neurons most using deep predictive models (2019) source

Brain-stimulation field models

Finite-element head models trace how current from electrodes or coils spreads through scalp, skull and brain tissue.

Question it answers: Where in the brain does the electric field from a stimulator land?

Leaves out: Tissue conductivities are uncertain, and neural response is not simulated.

Runs on: about 10¹² FLOPS rough Finite-element solves on individual head meshesSince: 2009 Gyri-precise head model of transcranial direct current stimulation (2009) source

Virtual-population drug exposure models

Whole-body drug transport models run across a synthetic population that varies in age, genes and organ function.

Question it answers: How much will drug levels differ across children, the elderly and patients with organ disease?

Leaves out: Depends on measured enzyme and transporter data.

Runs on: about 10¹¹ FLOPS rough Thousands of virtual subjects on a workstationSince: 2009 The Simcyp population-based ADME simulator (2009) source

Gut microbiome community models

Dozens of bacterial species in the intestine interact through competition and cross-feeding, with their abundances tracked over time under diet or antibiotic disturbance.

Question it answers: Which species stay stable in a gut, and which disturbance can tip the community?

Leaves out: Interaction strengths are inferred from noisy data and may not hold outside the sampled conditions.

Runs on: about 10¹⁰ FLOPS rough generalised Lotka-Volterra fitting and metabolic community simulation on a workstation.Since: 2013 Stein and colleagues, ecological modelling from time-series inference of intestinal microbiota. source

Surgical planning with patient-specific models

A scan is turned into a personal anatomical model, a surgical option is virtually applied, and flow or stress is computed to compare outcomes before operating.

Question it answers: Which surgical option gives this patient the best result?

Leaves out: Tissue properties are assumed from averages and healing after surgery is not modelled.

Runs on: about 10¹⁴ FLOPS rough one flow or stress solve per candidate plan, hours on a cluster.Since: 1999 Taylor and colleagues, predictive medicine using computational techniques for therapeutic decision-making. source

Heart valve fluid-structure interaction

A flexible valve leaflet and the blood flowing around it are computed together, as the leaflet is pushed open and slammed shut by the flow it controls.

Question it answers: How does a valve open and close, and how does a replacement valve change the flow?

Leaves out: Leaflet material properties and the surrounding muscle motion are simplified.

Runs on: about 10¹⁵ FLOPS rough moving-boundary fluid solves on a cluster.Since: 1972 Peskin, flow patterns around heart valves, introducing the immersed boundary method. source

Bone adaptation and implant remodelling

Finite-element models of bone in which tissue is added where stress is high and removed where it is low, over months and years.

Question it answers: How will bone change around an implant over time?

Leaves out: The remodelling rule is a phenomenological law and varies between patients.

Runs on: about 10¹¹ FLOPS rough iterative finite-element solves on a workstation.Since: 1987 Huiskes and colleagues, adaptive bone-remodelling theory applied to prosthetic design. source

Musculoskeletal models for surgery

Bones, joints and muscle-tendon actuators of the leg or arm are modelled so that a surgical change, such as a tendon transfer, can be tested for its effect on movement.

Question it answers: Will transferring this tendon improve how the patient walks?

Leaves out: Muscle activation is estimated, not measured, and nerve control is not modelled.

Runs on: about 10⁹ FLOPS rough inverse dynamics and muscle optimisation on a laptop.Since: 1990 Delp and colleagues, an interactive graphics-based model of the lower extremity to study orthopaedic surgery. source

Radiotherapy dose calculation

Photons or protons are followed through a patient scan by Monte Carlo transport to compute the radiation dose delivered to a tumour and to healthy tissue.

Question it answers: Does the planned beam deliver the dose to the tumour while sparing nearby organs?

Leaves out: The scan is a snapshot, so patient motion and anatomy change during treatment are approximations.

Runs on: about 10¹³ FLOPS rough millions to billions of histories per plan on a cluster or GPU.Since: 1988 Mackie and colleagues, photon energy deposition kernels generated with the EGS Monte Carlo code. source

Virtual anatomical phantoms for imaging

Digital human bodies with detailed organs and beating or breathing motion, used as the subject for simulated scans so image methods can be tested against known truth.

Question it answers: Can this scanner and algorithm recover a known abnormality?

Leaves out: The phantom is a shape model fitted to a few subjects and does not capture the full range of anatomy.

Runs on: about 10¹¹ FLOPS rough ray-tracing or transport simulation per scan on a workstation.Since: 2010 Segars and colleagues, the 4D XCAT phantom for multimodality imaging research. source

Magnetic resonance image simulation

The physics of nuclear spins in a magnetic field is simulated through a pulse sequence to generate synthetic scans of a virtual body.

Question it answers: What image would this sequence produce for this tissue?

Leaves out: Tissue relaxation values are averages and motion and hardware imperfections are modelled only partly.

Runs on: about 10¹² FLOPS rough Bloch-equation simulation of a full volume on a GPU.Since: 2005 Benoit-Cattin and colleagues, the SIMRI interactive MRI simulator. source

Fracture healing simulation

Mechanical strain in the healing gap of a broken bone decides whether tissue forms as cartilage, new bone or fibrous scar.

Question it answers: Will this fracture fixation allow the bone to heal?

Leaves out: The tissue-differentiation rule is empirical and blood supply is simplified.

Runs on: about 10¹⁰ FLOPS rough finite-element solves over simulated weeks on a workstation.Since: 1999 Claes and Heigele, local stress and strain along bony surfaces predicting the course and type of fracture healing. source

Within-host virus dynamics

Free virus, infected cells and target cells inside one patient are tracked, and their change after starting treatment reveals how fast virus is made and cleared.

Question it answers: How quickly does the virus replicate inside a person, and what does treatment change?

Leaves out: Tissue structure and sites of infection other than blood are not represented.

Runs on: about 10⁷ FLOPS rough a few coupled differential equations fitted to a patient time series.Since: 1996 Perelson and colleagues, HIV-1 dynamics in vivo. source

Coronary flow reserve from scans

Blood flow through the coronary arteries is computed from a heart scan to estimate whether a narrowing is restricting flow enough to need a stent.

Question it answers: Is this narrowing enough to starve the heart muscle of blood?

Leaves out: Boundary conditions at the small vessels are estimated from the patient rather than measured.

Runs on: about 10¹⁴ FLOPS rough one patient-specific flow solve per scan on a cluster.Since: 2013 Taylor and colleagues, CFD applied to cardiac CT for noninvasive quantification of fractional flow reserve. source

Cochlear mechanics

Sound waves in the fluid of the inner ear are coupled to the vibrating membrane inside the cochlea, generating the travelling wave that sorts pitches.

Question it answers: How does the ear turn sound into a place-coded signal?

Leaves out: The active amplification by hair cells is represented by approximated feedback.

Runs on: about 10¹⁰ FLOPS rough finite-difference solves of the cochlear fluid on a workstation.Since: 1981 Neely, a finite-difference solution of a two-dimensional model of the cochlea. source

Whole-body integrative physiology

Blood pressure, kidney, fluid balance and hormones are linked into one model of the circulation and its regulation.

Question it answers: How does the body regulate blood pressure, and what happens when one organ fails?

Leaves out: Each organ is a lumped compartment and cellular detail is absent.

Runs on: about 10⁶ FLOPS rough hundreds of differential equations on a laptop.Since: 1972 Guyton, Coleman and Granger, circulation: overall regulation. source

Bioheat transfer and hyperthermia

Heat conduction in tissue with blood flow carrying heat away, used to predict temperature during fever, heat stress or heat-based cancer treatment.

Question it answers: How hot does the tissue get and how deep does the heat go?

Leaves out: Blood perfusion is treated as a uniform heat sink.

Runs on: about 10⁹ FLOPS rough finite-element heat solves on a workstation.Since: 1948 Pennes, analysis of tissue and arterial blood temperatures in the resting human forearm. source

Kidney nephron transport models

Water, salt and urea moving through the loops and tubules of the kidney, solved as coupled flows to show how urine becomes concentrated.

Question it answers: How does the countercurrent arrangement of tubules let the kidney concentrate urine?

Leaves out: Tubule properties and three-dimensional layout are measured in few species, so parameters are extrapolated from animals.

Runs on: about 10⁹ FLOPS rough minutes to hours on a workstation.Since: 1966 Stephenson's analysis of concentration in renal counterflow systems. source

Vocal fold and voice production models

The vocal folds as masses on springs driven by air from the lungs, coupled to the shape of the throat and mouth, to produce voiced sound.

Question it answers: What physical settings of the folds make a clear, breathy or rough voice?

Leaves out: The folds are reduced to a few masses, so tissue layers and full three-dimensional airflow are not captured.

Runs on: about 10⁸ FLOPS rough seconds to minutes per sentence of voiced sound.Since: 1972 Ishizaka and Flanagan's two-mass model of the vocal cords for synthesising voiced sounds. source

Circadian clock models

Coupled oscillators standing for the body clock respond to light and the timing of rest, predicting how sleep and alertness shift with changes in schedule or travel.

Question it answers: How long does it take the body clock to adjust after a flight or a change of shift?

Leaves out: A few oscillators stand in for a whole network of neurons, and people differ widely in clock period and light sensitivity.

Runs on: about 10⁵ FLOPS rough seconds on a laptop.Since: 1982 Kronauer and colleagues' mathematical model of the human circadian system with two interacting oscillators. source

Compartment epidemic models

A population divided into susceptible, infected and recovered, with rates flowing between them.

Question it answers: How big will this outbreak get, and when will it peak?

Leaves out: Assumes homogeneous mixing, with every person meeting every other at the same rate, which real contact patterns do not satisfy.

Runs on: about 10⁹ FLOPS rough a laptop solving a handful of coupled epidemic equationsSince: 1927 Kermack and McKendrick's equations for a closed epidemic

Run the interactive model

Agent-based epidemics

A synthetic population of a whole city or country, each person with a household, school or workplace, and contacts along which disease spreads.

Question it answers: Which measures slow the spread most, at the least cost?

Leaves out: Requires detailed household, school and workplace data that many countries lack.

Runs on: about 10¹⁶ FLOPS rough a university cluster simulating a country-scale synthetic populationSince: 2005 Longini, Ferguson and colleagues' agent-based influenza pandemic models

Run the interactive model

Pathogen evolution

Mutations appearing and spreading in a virus or bacterium as it passes between hosts.

Question it answers: How fast could resistance to this drug evolve?

Leaves out: Real mutation and fitness landscapes are far more complex than a handful of loci.

Runs on: about 10¹² FLOPS rough a desktop workstation simulating drug-resistance evolutionSince: 1990 early stochastic models of drug-resistance evolution in HIV

Disease burden and health economic modelling

A disease and its treatments followed as probabilities of illness, disability, cost and death, used to compare which health interventions do the most good per pound spent.

Question it answers: Which treatment or programme gives the most healthy years of life for the money?

Leaves out: Assigning a number to a year of life or suffering is a value judgement presented as arithmetic.

Runs on: about 10¹¹ FLOPS rough a desktop workstation running a cohort cost-effectiveness modelSince: 1993 the World Bank's global disability-adjusted life year burden-of-disease study

Outbreak genomic epidemiology

Pathogen genomes sampled during an outbreak, used to build a family tree of infections and infer who most likely infected whom and how fast it is growing.

Question it answers: How is this outbreak actually spreading, behind what contact tracing alone can see?

Leaves out: Only cases that get sequenced show up in the tree, so a thin or biased sample can mislead.

Runs on: about 10¹³ FLOPS rough a university cluster building trees from thousands of genomesSince: 2013 real-time genomic tracking of an Ebola outbreak in West Africa

Environmental and chemical exposure risk assessment

How a population's exposure to a pollutant, food contaminant or radiation dose translates into a probability of harm, run across the range of uncertainty in the data.

Question it answers: How much of this chemical is safe for a whole population to be exposed to?

Leaves out: Extrapolating from high test doses in animals to low doses in people is a persistent judgement call.

Runs on: about 10¹¹ FLOPS rough a desktop workstation running a Monte Carlo dose-response modelSince: 1983 the US National Research Council's framework for chemical risk assessment

Mobility-network epidemic models

Disease spread simulated across a global network of cities connected by commuting and air-travel data.

Question it answers: How fast will an outbreak reach other regions and countries?

Leaves out: Mixing inside each city is simplified, and behaviour change is not modelled.

Runs on: about 10¹² FLOPS rough Stochastic runs across thousands of subpopulationsSince: 2009 Multiscale mobility networks and the spatial spreading of infectious diseases (2009) source

Individual-based malaria models

Simulated people, mosquitoes and parasites follow within-host infection and transmission to test control strategies.

Question it answers: Which mix of nets, drugs and vaccines would cut malaria transmission?

Leaves out: Mosquito ecology and local health systems are modelled coarsely.

Runs on: about 10¹³ FLOPS rough Many stochastic runs of a simulated villageSince: 2008 Towards a comprehensive simulation model of malaria epidemiology and control (2008) source

Scenario modelling hubs

Several modelling teams run the same what-if assumptions about vaccines or variants, and the projections are combined into one ensemble.

Question it answers: What range of futures follows from each policy or variant assumption?

Leaves out: Scenarios are conditional, not predictions of what will happen.

Runs on: about 10¹² FLOPS rough Multi-team ensemble of epidemic modelsSince: 2020 US COVID-19 Scenario Modeling Hub, from December 2020 source

Digital contact-tracing models

Transmission models with an app-based tracing step estimate how fast tracing and isolation must be to stop spread.

Question it answers: How quickly must contacts be traced to bring an epidemic under control?

Leaves out: Uptake and compliance are assumed values.

Runs on: about 10⁹ FLOPS rough Branching-process or agent runsSince: 2020 Quantifying SARS-CoV-2 transmission suggests epidemic control with digital contact tracing (2020) source

Antimicrobial resistance spread

Susceptible and resistant bacteria compete inside patients and spread between them under different antibiotic policies such as cycling, mixing or combination treatment.

Question it answers: Which treatment policy slows the rise of resistant infections?

Leaves out: Resistance costs and transmission rates are poorly measured, and the results depend on them.

Runs on: about 10⁹ FLOPS rough compartment and individual-based hospital models on a laptop.Since: 1997 Bonhoeffer, Lipsitch and Levin, evaluating treatment protocols to prevent antibiotic resistance. source

Malaria transmission and intervention models

Mosquito biting, parasite development, human immunity and interventions such as bed nets or vaccines are simulated together in a village or country.

Question it answers: Which combination of controls can reduce or eliminate malaria?

Leaves out: Mosquito ecology and local human behaviour are poorly measured.

Runs on: about 10¹¹ FLOPS rough individual-based simulation of a country over decades on a cluster.Since: 2008 Smith and colleagues, towards a comprehensive simulation model of malaria epidemiology and control. source

Indoor airborne transmission simulation

Breathing, coughing and talking release droplets and aerosols into a room, and fluid dynamics tracks where they go and how much a second person inhales.

Question it answers: How far does exhaled virus travel, and how much do masks or ventilation reduce exposure?

Leaves out: Infectious dose and real room conditions vary more widely than any simulation setup assumes.

Runs on: about 10¹⁴ FLOPS rough large-eddy simulation of a room with particle tracking on a cluster.Since: 2020 Mittal, Ni and Seo, the flow physics of COVID-19, reviewing computational studies of respiratory flows. source

Tobacco-control policy simulation

Smoking starts, quits and deaths in a country are projected under taxes, bans and campaigns, producing the change in prevalence and lives saved.

Question it answers: How many lives would a tax or smoking ban save?

Leaves out: Policy effects are taken from past evidence and may not transfer to new settings.

Runs on: about 10⁶ FLOPS rough population projection model on a laptop.Since: 2000 Levy, Cummings and Hyland, a simulation of the effects of youth initiation policies on overall cigarette use. source

Vaccine allocation optimisation

An epidemic model is combined with an optimiser that chooses which age or risk groups receive scarce vaccine first.

Question it answers: Who should be vaccinated first to save the most lives?

Leaves out: Contact patterns and vaccine effectiveness are uncertain, and the best group depends on the aim.

Runs on: about 10⁸ FLOPS rough optimisation over many epidemic runs on a workstation.Since: 2009 Medlock and Galvani, optimising influenza vaccine distribution. source

Virtual imaging trials

Simulated patients, scanners and readers are used to compare imaging technologies without enrolling real people.

Question it answers: Does a new imaging device find cancers as well as the current one?

Leaves out: Virtual lesions and readers are models that may miss real clinical variety.

Runs on: about 10¹³ FLOPS rough thousands of simulated images on a cluster.Since: 2018 Badano and colleagues, in silico imaging trial of digital breast tomosynthesis versus full-field mammography. source

Epidemics on contact networks

Disease spreads between nodes of a network whose degree distribution may be highly uneven, and the outcome depends on the hubs.

Question it answers: At what level of transmissibility does an epidemic take off on a network with hubs?

Leaves out: Real contact networks change over time and are only partly observed.

Runs on: about 10⁷ FLOPS rough Monte Carlo spread on million-node networks on a workstation.Since: 2001 Pastor-Satorras and Vespignani, epidemic spreading in scale-free networks. source

Mosquito population dynamics

Eggs, larvae, pupae and adults of a disease-carrying mosquito develop in containers under local temperature and rainfall, giving the number that bite.

Question it answers: How many mosquitoes will there be next season, and what does removing breeding sites do?

Leaves out: Breeding sites are hard to count and vary street by street.

Runs on: about 10⁷ FLOPS rough life-table simulation on a laptop.Since: 1993 Focks and colleagues, a dynamic life-table model for Aedes aegypti. source

Cancer screening microsimulation

Individual life histories with the natural course of cancer are simulated with and without screening to estimate cases found and deaths prevented.

Question it answers: How many deaths does a screening programme avert, and at what cost in overdiagnosis?

Leaves out: The unobserved early course of cancer must be assumed.

Runs on: about 10⁹ FLOPS rough millions of simulated lives on a workstation.Since: 1985 Habbema and colleagues, the MISCAN simulation program for the evaluation of screening. source

Obesity and diet agent-based models

Agents with weight, income and neighbours eat and move, and the spread of body weight across the population shifts under food price or policy changes.

Question it answers: Which interventions bend population weight over decades?

Leaves out: Behaviour is reduced to a few rules and long-term evidence is thin.

Runs on: about 10⁸ FLOPS rough population agent-based simulation on a workstation.Since: 2007 Hammond and Epstein, exploring price-independent mechanisms in the obesity epidemic. source

Organ allocation policy simulation

Patients on a transplant waiting list, arriving donor organs and allocation rules are simulated to compare waiting time and survival.

Question it answers: Which allocation rule yields the most life-years and the fairest wait?

Leaves out: Organ quality and the decisions of patients and centres are simplified.

Runs on: about 10⁸ FLOPS rough event simulation of years of waiting lists on a laptop.Since: 2000 Zenios, Chertow and Wein, dynamic allocation of kidneys to candidates on the transplant waiting list. source

Ambulance and fire station deployment models

Emergency calls arrive at random, units are dispatched, and a queueing model gives response times for each location and district layout.

Question it answers: Where should stations and units be placed to cut response time?

Leaves out: Travel times and the true spread of demand are assumed.

Runs on: about 10⁷ FLOPS rough queueing and Monte Carlo on a laptop.Since: 1974 Larson, a hypercube queueing model for facility location and redistricting in urban emergency services. source

Synthetic patient record generation

Artificial people are given diseases, treatments and encounters from published clinical rules, to produce realistic health records that contain no real person.

Question it answers: Can software and analytics be tested on realistic records without exposing anyone's data?

Leaves out: The records carry only the patterns put into the generator, so rare diseases and real care pathways are thinly represented.

Runs on: about 10¹⁰ FLOPS rough a laptop-scale run for a million patients.Since: 2010 Buczak and colleagues' data-driven approach for creating synthetic electronic medical records. source

Predictive toxicology

A candidate drug's structure compared to databases of known toxic molecules, or run through models trained on lab and animal data, before it is ever synthesised.

Question it answers: Is this molecule likely to damage the liver, heart or genes before we make it?

Leaves out: Predictions are only as reliable as the past toxic molecules on which the model was trained, and biological exceptions continue to emerge.

Runs on: about 10¹³ FLOPS rough a university cluster running a trained toxicity prediction modelSince: 2000 early structure-activity relationship models used to screen drugs for toxicity

Pharmacodynamic modelling

How the concentration of a drug at its target translates into an effect on the body, from receptor binding up to a measured symptom or biomarker.

Question it answers: How much effect will a given dose actually produce, and for how long?

Leaves out: Assumes a clean concentration-effect relationship that tolerance and disease progression can disrupt.

Runs on: about 10¹⁰ FLOPS rough a desktop workstation solving a handful of coupled effect equationsSince: 1968 early receptor-occupancy models linking drug concentration to effect

Clinical trial design simulation

Thousands of hypothetical trials run against assumed patient variation and drug effects, to choose a trial's size, duration and rules for stopping early.

Question it answers: How many patients does this trial actually need, and when can we stop it safely?

Leaves out: The simulated trial is only as realistic as the assumptions about patient response that are supplied to it.

Runs on: about 10¹² FLOPS rough a small cluster running thousands of simulated trial replicatesSince: 1994 early simulation studies used to plan adaptive clinical trial designs

Drug repurposing network models

Drugs, genes, proteins and diseases linked in one large network, searched for existing drugs whose known connections suggest an unexpected new use.

Question it answers: Could a drug already approved for one disease also treat another?

Leaves out: A network link is a hint, not proof; most suggested repurposings still fail when tested.

Runs on: about 10¹² FLOPS rough a desktop workstation searching a large drug-gene-disease networkSince: 2011 early network-based screens for drug repurposing candidates

Drug formulation and delivery modelling

A tablet dissolving, a patch releasing or a nanoparticle carrying a drug through the body, followed as diffusion and reaction over minutes to days.

Question it answers: How fast and where does this drug actually reach its target once swallowed or injected?

Leaves out: Lab dissolution rigs and real digestion do not always agree.

Runs on: about 10¹¹ FLOPS rough a desktop workstation solving diffusion and release equationsSince: 1961 Higuchi's mathematical model of drug release from a solid matrix

Vaccine antigen design

The surface of a pathogen protein simulated at atomic detail to find or engineer the shapes the immune system will best recognise and remember.

Question it answers: Which piece of this pathogen makes the best target for a vaccine?

Leaves out: A shape that looks like a good target on screen still has to work in a real immune system.

Runs on: about 10¹⁵ FLOPS rough a university cluster running atomic simulations of candidate antigensSince: 2013 structure-based design of a stabilised respiratory virus antigen

Closed-loop diabetes control in silico

A population of virtual people with type 1 diabetes, with meals and glucose-insulin dynamics, used to test delivery control algorithms before human trials.

Question it answers: Would this control algorithm keep glucose safe across many different people?

Leaves out: Real behaviour such as stress, illness and sensor faults is only partly captured.

Runs on: about 10⁹ FLOPS rough long virtual-day simulations of hundreds of subjects on a laptop.Since: 2009 Kovatchev and colleagues, in silico preclinical trials of closed-loop control of type 1 diabetes. source