The whole-cell atlas, and how to read it

A manual for the atlas of one budding-yeast cell: what is on the screen, how to move around it, what every number and every mark means, and where each piece of data comes from.

This page describes the build of 2026-09-29. Its counts, stops, compartments and datasets are written from that build, the same one the atlas shows.

Using it

What the atlas is

The atlas draws one cell of budding yeast, Saccharomyces cerevisiae S288C, from public databases. Every gene is a dot, 6,459 of them: 6,021 protein-coding genes and 438 RNA genes. Each dot sits in its compartment and is coloured by when its family arose. Around the genes are the cell's 4,105 metabolic reactions and 2,748 metabolites (the Yeast9 model), its 634 complexes, 702,499 links of five kinds, and 926 complexes, pathways and processes dated as systems.

A timeline walks the cell back through the tree of life, from what all life shares to today. At each stop the cell shows what existed then. Four layers put the same question, how old is it, to other properties of the cell:

• Energy: where the cell makes and spends its ATP, in a metabolic model (flux balance analysis).
• Connections: how old each protein–protein interaction is, from the same interaction seen again in other species.
• Heat: which layers the cell switches off, condenses, or calls on through Hsf1 and Msn2/4 when it is heated.
• The frontier: the young translated sequences outside the annotated genes, and the genes they were born beside.

A third view draws the cell's networks: what each link says (it activates, represses, marks or touches), how the control network is wired, and how close its signed part comes to monotone, two camps that activate within and repress between. A band beside it holds what the networks act on: the cell's matter, lumped.

It is a research tool and a working draft, built from 40 public datasets. The drawing is a diagram, not a micrograph: each compartment is drawn big enough to hold its genes, not to scale.

Quick start

Each example link opens the atlas in the state it describes.

1. Find a gene. Type a name such as PMA1, or an ORF such as YGL008C, in Find a gene. The atlas selects it and centres it. PMA1
2. Read about it. The panel beside the drawing (below it on a phone) gives its age, place, copies, cost, its response to heat, the young sequences around it, its links and its reactions. Every underlined name in the panel leads on.
3. Walk back in time. Press play: the cell rebuilds from all life to today, one stop every 1.4 seconds, with what did not exist yet faded out. Or click a stop on the timeline. the cell at LECA
4. Size the dots by a quantity. Choose size: heat: leaves solution: the proteins that condense when the cell is heated grow. leaves solution
5. Follow the links back. Select a gene and step the timeline back. Its physical links stay solid while the interaction itself is known that far back, and turn dashed where only the two partners are. RPT1 at LECA
6. See the systems. Press systems (top left of the drawing) for the complexes, pathways and processes, grouped by when each came together. systems
7. See the networks. Press networks for the control network drawn as a bow-tie and split into its two camps, each link's head saying what it does. Select a gene and press circuit for its own. networks HSF1's circuit
8. See what makes the matter. In the networks view, set mass flow to a growth condition: a ribbon runs from each enzyme to the pool of matter it makes. Click a pool for its makers and the cell's mass budget; switch on small molecules for the metabolites that act back on the enzymes. mass flow in the cell
9. Change what is drawn. The colour menu changes the age scale or turns it off, and show switches the proteins or the RNAs off. only the RNAs, in Blues

The screen

1. The header: find a gene, and the menus and switches that set what is drawn (see The controls).
2. The drawing: the cell, every gene in its place. The switch at its top left changes it to the systems view or the networks view.
3. The panel: an introduction with the atlas's findings, or whatever is selected: a gene, a compartment, a complex or a system.
4. The timeline: play, the stops at their dates, and the log and linear time scales.
5. The readout and the key: the chosen stop and what the cell holds there, then the colours, the shapes and how genes are placed.

The controls

The browser remembers the theme, the time scale and the colour scale between visits. The other settings start fresh each time.

Ways to interact

Links into the atlas

A link can open the atlas in a chosen state, which makes a view easy to share. Add any of these to the atlas's address after a ?, joined by &, for example ?gene=SSA4&size=heatUp&stop=2. try it

Feedback

The Feedback button beside theme opens a box for a note: what looks wrong, what is unclear, what you would change. With it goes the atlas's state, unless you untick it: the view, the stop, the time scale, every menu and switch, the selection, the systems and networks filters, the theme, and where each drawing is zoomed to (its centre and scale, so a window of another size shows the same place). With the state also go the page's address, the build's date, your browser's name and your window's size. Nothing else is collected.

• On thinkingembryo.com the record is sent to the site's author, and kept as one file per note.
• Run locally with python -m cellatlas serve, it is saved in outputs/feedback/, and the box lists the latest notes. ?feedback=<id> reopens the atlas as a note left it.
• Anywhere else (a plain file server, or when sending fails) it is saved as a file for you to pass on.

Download JSON saves the same record as a file. Open a saved record… (or dropping a record's file on the box) puts the atlas back as the record left it. That works with any record, including one sent from the site. A record made on this page brings you back here first. Reopening a record never changes the theme, colour scale or time scale your browser remembers.

Reading it

Place, shape and ghosts

Where a dot sits

• Compartments. 20 are drawn. The mother cell is a disc inside its plasma membrane and wall, with the bud on the right, joined by the neck. The nucleus, with its nucleolus, envelope and spindle pole body, sits in its ring of endoplasmic reticulum, and the other organelles float in the cytosol. A ring around the cell stands for its outside, and a tray below it holds the 549 genes with no known place.
• Size of a compartment. Its area follows the number of genes drawn in it, so the dots have one density everywhere. A large compartment has many parts, not a large volume.
• One place per gene. Each gene is drawn once, in its primary compartment. Its panel lists its other places. See Place for how the primary one is chosen.
• Growth rings. Inside each compartment the oldest genes sit at the centre, or on the inner rows of a ring-shaped compartment, and the youngest at the rim. Within one stop, the most abundant sit nearest the centre. How far a dot is from its compartment's centre therefore reflects its age, not where the protein sits in the real organelle.

Shape

a disc is a protein-coding gene; a diamond is an RNA gene.

What exists at a stop

• A gene is drawn from the first stop at which both it and the compartment it is drawn in exist. When the compartment is the younger of the two, the gene's panel says so: "Drawn from … on".
• With ghost what is not there yet on, genes that do not exist yet are drawn at a tenth of their strength; with it off they are not drawn.
• A compartment that does not exist yet has a dashed outline, no fill, and a faded name.
• Genes that nothing dates appear only in today's cell, the last stop.

Colour

Colour is age. Each gene takes the colour of its era, from the darkest, the oldest, to the lightest, the youngest. The timeline keeps all 12 stops, and the colours group them into five eras. Grey is for genes with no call, or RNA genes of undated classes. The same eras colour the timeline's stops, the bars of the systems view, the key and the swatches in the panel.

The colour menu

Every scale below uses the stretch of its colour map whose two ends still stand out against both the light and the dark background (2:1 contrast or more), with lightness rising evenly from the oldest step to the youngest.

Size

With size: equal every dot is the same size. Every other choice sizes the dots by one quantity. The keys below are drawn with the atlas's own scales, enlarged for reading. The dashed circle is a gene without the quantity, drawn as small as the scale allows.

• Small can mean unmeasured. A gene without the quantity is drawn as small as one with none of it. Check its panel before reading a small dot as a small value.
• There is a largest size. The largest dot touches its neighbours, and every value past the one that reaches it looks the same.
• Read the order, not the area. The log scales compress orders of magnitude into a few steps of radius.

equal

by copies copies per million protein molecules

Radius 3 + 1.4 log₁₀(copies + 1), on a log scale. See Copies.

by cost share of the protein-building budget

The same log scale, on the share in parts per million of the budget. See Cost and budget.

by flux, respiring · by flux, fermenting mmol per gram of dry weight per hour

The flux through the gene's reactions in the metabolic model, in one growth condition. The same log scale, with a flux of 0.001 counting as 1. A gene whose reactions carry no flux is drawn a little larger than a gene with no reactions at all. See Energy.

by information flow share of every pair of genes' current

The current through the gene when the contact network carries a unit between every pair of genes. The same log scale, with a hundred-thousandth counting as 1. A gene with one contact passes none on; a gene in no contact is the smallest. See Flows compared.

by energy, respiring · by energy, fermenting mmol ATP per gram of dry weight per hour

The ATP spent building the gene's copies plus the ATP its reactions make or spend. The same log scale, with a hundred-thousandth counting as 1. See Flows compared.

by mass flow, respiring · by mass flow, fermenting mg per gram of dry weight per hour

The grams the gene's reactions take in. The same log scale, with a milligram counting as 1. A gene whose reactions carry nothing is drawn a little larger than a gene with no reactions. See Flows compared.

heat: mRNA up · heat: mRNA down fold change of the mRNA, 10–20 minutes into heat

Linear in the log₂ change: 1.2 units of radius for each doubling (mRNA up) or halving (mRNA down). A gene that moves the other way, or not at all, is as small as one not measured. See Heat.

heat: leaves solution fall in the share in solution, 30 to 46 °C

Linear: 9 units of radius for a fall of 1, the whole protein. A fall of 0.3 is the cut-off for leaving solution. See Heat.

young ORFs around it translated ORFs placed by the gene

The square root of their count. See The frontier.

Selection and links

• Rings. The selected gene has a dark ring and the gene under the pointer a grey one. Selecting a complex or a system rings every member.
• Links run from the selected gene to its partners of the kind chosen in links, in the accent colour ( aqua; deep orange under viridis). Only links whose two genes exist at the chosen stop are drawn.
• Solid or dashed. A solid line is an interaction known at the chosen stop. A faint dashed line joins two partners that both exist at that stop, although the interaction is not known that far back. For physical links, "known" means dated by another species at or before the stop (see Connections back in time). A physical link that no other species dates is known from the stop S. cerevisiae on, or from its younger partner's stop if that is later. Every other kind of link is known whenever both its genes exist, so it is always solid.
• Partner rings. A partner gets a full ring in the accent colour when its link is known at the stop, and a faint thin one otherwise.
• Arrowheads. Regulation and modification links have a direction: the arrowhead sits at the target, pointing away from the regulator or the modifying enzyme.

The timeline

The timeline is the evolutionary axis: 12 stops, each at its date in years before present. A stop's genes are those already present in the ancestor at that split, or, for the last stop, what nothing could date.

• Dates are TimeTree 5's median of published estimates for the split between S. cerevisiae and a species that branched off at that node (Kumar et al. 2022, Mol Biol Evol 39:msac174 (doi:10.1093/molbev/msac174); retrieved 2026-09-18). The 95% interval is the coloured bar under the line.
• Windows. Two stops are windows, drawn as dashed spans on the line. Genes shared only with archaea, or only with bacteria, entered the eukaryotic lineage at some point between the root and LECA, and when is not known. Genes found in no other yeast of the gene-order browser arose after the split from S. mikatae.
• A gene's own date. A gene dated by gene order carries the date of the split from the most distant yeast that shares it, which can be finer than its stop's.
• Scales. Log runs from 4.5 billion to 1 million years ago, then a dotted break before today. Linear runs from 4.5 billion years ago to today, and shows how recent most of the stops are.

What "exists at a stop" means for each kind of element:

The systems view

The systems view lists 926 systems: 615 complexes (Complex Portal), 198 metabolic pathways (YeastPathways) and 113 processes (SGD's GO slim). They are grouped under the stop at which each came together.

• A row gives a system's name, its kind, its number of parts and a bar of its parts in age order, oldest on the left. Each stop is a segment as wide as its share of the parts, in its era's colour. The thin mark halfway along is the part that dates the system.
• A part is a gene. A set of paralogs that can fill the same place in a complex counts as one part, as old as its oldest member.
• Walking the timeline fades the parts that do not exist yet, and dims the groups that have not come together. The line above the list counts the systems that have come together and those with all their parts.
• Filter by name, switch kinds on and off, and click a row for its panel.

The networks view

The networks view draws the cell's networks as networks. The dots are the same genes, with the same colours, sizes and shapes, and they follow the same timeline, selection and switches as the cell. The networks' data load when the view first opens.

Layouts

By topology: control

The control network as its bow-tie (see The bow-tie): a band for each region, from the top upstream, core, downstream regulators, downstream targets, elsewhere, and outside the network. Across the bands run three columns: the stress camp, the genes with no camp, and the growth camp. In each block the genes sit in rows, oldest first and, within a stop, the most abundant first.

Above the bands, a band holds the matter the networks act on: the biomass lumped into pools, each disc's area following its mass, with its name and its grams per gram of dry weight (see Matter). With mass flow on, a ribbon runs up to each pool from each enzyme that makes it.

By topology: contact and coupling

The network's modules (see Modules), each a disc at the cell's density of dots with its oldest genes at the centre, and named for the process most over-represented in it. The genes in no module wait in the tray below.

In the cell

Every gene where the cell view draws it, inside the compartments' outlines: the network laid over the cell's places. The pools of matter sit in their compartments, at the same size as in the band: the wall's glucans and mannan on the wall, chitin at the bud neck, the membrane lipids and acyl chains on the plasma membrane, DNA in the nucleus, the storage lipids in a lipid droplet, and the rest in the cytosol, spread through its free space. The protein is not drawn as a pool: its parts are the dots.

Circuit

The selected gene at the centre and its partners in the network. For control, what acts on it is in an arc above and what it acts on in an arc below; for contact and coupling, its partners are all around it. Partners are ordered by kind of link, then by age, in rows that fill outward. At most 300 are drawn, and the summary line says how many more there are. Every name is shown.

The morph

Changing the network or the layout moves every dot from where it was to where it goes, over 0.7 seconds (at once when the system asks for reduced motion), so a gene can be followed from where it sits to how it is wired.

Links

• Zoomed out, while neighbouring dots are less than 10 pixels apart, links are drawn as bundles: one curve for each pair of groups and kind of link, its width and strength growing with the square root of its number of links. The groups are the bow-tie's regions split by camp, the modules, or the compartments. Each bundle bows to its right, so the two directions between a pair of groups stay apart, and a directed bundle ends in its kind's head.
• Zoomed in, every link with an end in view is drawn, up to 6,000 at once. The selected gene's links are always drawn.
• The head says what a link does. Every link is in the one link colour, so that colour stays with age:

Control

• activates
• represses
• dual
• sign unknown
• adds a mark
• removes a mark
• adds and removes marks
• acts on

Contact and coupling

• contact
• co-complex
• to an mRNA
• aggravating
• alleviating

• Lines. A contact is a plain line; a co-complex pair and a protein bound to an mRNA are fainter. The two kinds of genetic coupling share a dotted line: switch one off to see the other. What each kind means is under Kinds of link.
• Dashed means what it means on the cell: a physical link between two genes that exist at the chosen stop, although the interaction itself is not known that far back.
• Frustrated links, the signed links that disagree with the best split into two camps, are drawn over a wider stroke of ink, so they read as double lines (see Monotone or not).
• Camps. With camps on, a gene of the growth camp has a thin solid ring and one of the stress camp a dotted ring, once neighbouring dots are 7 pixels apart (see The two camps).
• At an earlier stop, the genes that do not exist yet are ghosts, and only links whose two genes exist are drawn.
• Names appear beside the dots once they are 30 pixels apart, and always in the circuit.
• Ribbons are mass flow, in neutral ink: from an enzyme to a pool it makes, as wide as the square root of its grams per hour. Selecting a gene or a pool keeps its own ribbons and fades the rest. See What makes each pool.
• The genes outside a network are counted with their share of the protein mass: in the control network's last band, and beside the modules' tray.
• Small squares in the matter band, right of the pools, are small molecules that act on the enzymes, those acting on the most first; their names show when zoomed in. Their links take the heads of control: an arrow where the molecule activates the enzyme, a bar where it inhibits it, both where BRENDA records both.

The summary line

At the foot of the drawing: the stop, and the links drawn there, by kind. For control it adds the signed links that exist at that stop, how many of them the best split at that stop frustrates, and how many it frustrates on average with the signs shuffled. With mass flow on, it adds the matter laid down per hour. In the circuit it says how many partners are not drawn.

In the panel

Once the networks have loaded, a gene's panel gains In the networks: its role (see Roles), its region of the bow-tie, its camp and how many of its signed links are frustrated, whether it regulates itself, how many feed-forward loops it is in, its links by kind (acting on others, acted on, and with others), its modules, and the pools it helps make, with its share of each. Hover over a small molecule for the enzymes it acts on, and click it for their list; hover over a pool for its name and weight; click it for its own panel: what it holds, how fast the cell lays it down, the reactions that make it, and the cell's mass budget drawn as three columns (what it takes in, the elements, what it gives out and builds), each band as wide as its mass.

Every quantity

A gene's identity

Name, ORF and UniProt entry

The gene's standard name from SGD, or its systematic name when it has none; its systematic ORF name, such as YGL008C; and its UniProt accession.

Shown at the top of its panel and in the tooltip.

Protein, or type

UniProt's name for the protein, or SGD's feature type for an RNA gene (tRNA gene, snoRNA gene, …).

Which genes

SGD's protein-coding genes, Verified and Uncharacterized, and its RNA genes: 6,459. Dubious ORFs, unlikely to encode a protein, are left out.

Roles

The chips under a gene's heat and frontier sections. enzyme: it catalyses a metabolic reaction of Yeast9. transporter: a transport reaction. complex subunit: a member of a Complex Portal complex. transcription factor: a regulator of that type in SGD's regulation data. kinase, phosphatase, ubiquitin ligase: the enzyme named for a phosphorylation, dephosphorylation or ubiquitination site in SGD's modification data.

About

SGD's one-line description, at the end of the panel.

Age

A gene's age is the stop at which its family appears: the deepest split at which a relative is found. It is the age of the gene's lineage, not of its present function.

Stop, class and date

The stop's name in bold; in brackets, the age table's class where it names the stop differently (Euk_Archaea for "from archaea or bacteria", say, or the RNA gene's type); then its date. The line below says what the stop means.

Shown as the colour of the dot, and under Age in the panel.

How a protein is dated

In this order, by the first rule that gives an age:

1. The consensus age (5,000 proteins): Liebeskind, McWhite and Marcotte 2016 pooled 13 orthology algorithms into a probability for each age class, and the most likely class is taken. Entries the 2016 table names by accessions that UniProt has since merged or split are traced to today's entries.
2. Its duplicate (17): the call of its partner from the whole-genome duplication.
3. Its family (100): the most common call among the dated members of its UniProt family, in yeast or, for the classes both tables share, in human.
4. Gene order (548): the Yeast Gene Order Browser finds the same locus in related yeasts. A gene at the same locus in yeasts that split before the whole-genome duplication is at least as old as the budding-yeast family; one found only in species descended from the duplication, only in Saccharomyces, or in no other yeast of the browser dates to the three younger stops.
5. No call (356): nothing dates it. This is the last stop, drawn in grey.

How an RNA gene is dated

By its class (409 genes), since the age tables are built from proteins. rRNA and tRNA: all life. snoRNA: shared with archaea. snRNA and telomerase RNA: LECA. RNase P RNA (RPR1) and the signal recognition particle's RNA (SCR1): all life. RNase MRP RNA (NME1): LECA. The other 29 RNA genes, of mixed classes, are undated and grey.

Flags

Notes under a gene's age, where the call deserves care:

Flag	What it says	Genes
weak call	the consensus is unsure: the most likely age class has a posterior below 0.5	872
few algorithms	fewer than six of the orthology algorithms contributed to the call	582
possible horizontal transfer	the age table flags a possible horizontal gene transfer: a deep call may come from a transfer, not from descent	258
age from its WGD ohnolog	no call of its own: it takes the call of its partner from the whole-genome duplication	17
age from its UniProt family (yeast)	no call of its own: the most common call among the dated yeast members of its UniProt family	69
age from its UniProt family (human)	no call of its own: the most common call among the dated human members of its UniProt family (only the classes both age tables share)	31
dated by gene order (YGOB)	no call from orthology: dated by the most distant yeast that keeps a gene at the same place in its genome	548
no orthologs called	nothing dates it: the no-call stop	356
undated RNA class	an RNA gene of a mixed class with no date of its own: grey	29

The usual error of age calls is too young: a short or fast-evolving gene can lose every detectable relative. Interactions dated older than a partner's own age (see Connections back in time) point to some of these.

Place

Compartment, and where else

The compartment a gene is drawn in, then "also" and its other places. Both come from SGD's Gene Ontology annotations of where each gene product is found, which include the proteome-wide GFP screen of Huh et al. 2003. Each annotated term is mapped up the GO hierarchy (GO release 2026-07-26) to the nearest compartment the atlas draws. For a protein with no GO location, UniProt's location text is used.

The primary place is the one with the best evidence (experimental, then curated, then electronic, then UniProt's text), then the most annotations, then the most specific compartment.

Shown by where the dot sits, in the tooltip, and under Place in the panel.

No known place

549 genes have no location in GO or UniProt. They wait in the tray below the cell, oldest first.

When a compartment appears

Each compartment's date is a judgement from the literature, not computed. The panel of a compartment sets it beside the ages of the genes drawn in it, so a date the genes contradict shows up there.

A compartment's panel gives its make-up by era: a bar for each era, as long as its share of the genes drawn there relative to the largest era, with its percentage.

Copies, cost and budget

Copies per million protein molecules

How many copies of the protein a cell holds, as a share of all its protein molecules: PaxDb 6.1's integrated whole-organism abundance, measured for 5,815 proteins, up to 23,256.

Shown under Place in the panel, and as size: by copies.

Cost to make ATP equivalents for one copy

What one copy costs in high-energy phosphate bonds: making its amino acids plus four bonds for each peptide bond that joins them (two to charge the tRNA, one GTP to deliver it, one to move the ribosome on). Given for respiring and for fermenting cells (6,021 proteins). An amino acid's cost counts the bonds spent making it and those lost by diverting its precursors from energy production, so fermenting cells, whose carbon yields little energy anyway, pay much less.

Amino acids: Wagner 2005, Mol Biol Evol 22:1365 (doi:10.1093/molbev/msi126), as tabulated in Barton et al. 2010, PLoS ONE 5:e11935, Table 1 (doi:10.1371/journal.pone.0011935; CC BY 4.0). Peptide bonds: Cox & Garcia 2013, PLoS ONE 8:e59883 (doi:10.1371/journal.pone.0059883).

Share of the building budget percent

A protein's copies times its cost (respiring), over the same product summed over every protein with both (5,815 proteins): its share of what the cell spends building its proteins. A compartment's panel sums it over the proteins drawn there.

Shown under Cost to make in the panel, and as size: by cost.

Reading it: the budget goes to the abundant proteins far more than to the long ones. It counts building each copy once; how fast proteins are broken down and remade is not measured here.

Energy

The energy layer is a prediction from Yeast9 (yeastGEM_v9.1.1), a genome-scale model of the cell's metabolism with 4,105 reactions and 2,748 metabolites. Flux balance analysis finds a steady state, in which each metabolite is made as fast as it is used, that grows as fast as the medium allows. Of the many such states, the parsimonious one is taken: the least total flux that keeps the best growth (parsimonious flux balance analysis (Lewis et al. 2010, doi:10.1038/msb.2010.47); SciPy's HiGHS solver).

Uptake limits, ethanol and ATP are in mmol per gram of dry weight per hour. The respiring condition is glucose-limited with oxygen free, as in a slow chemostat. The fermenting one is oxygen-limited on ten times the glucose, since plain Yeast9 cannot grow with no oxygen at all.

A reaction's flux mmol per gram of dry weight per hour

Under each reaction in a gene's panel, for each condition. A negative flux runs the reaction backwards; 0 means the solution does not use it. ⇌ marks a reversible reaction, → one that runs one way. Each metabolite carries its compartment in brackets.

Flux through a gene's reactions mmol per gram of dry weight per hour

The sum of the flux sizes of every reaction whose enzyme rule names the gene. Each subunit of an enzyme complex, and each isozyme, is credited with the reaction's whole flux. 1,143 genes have reactions in the model.

Shown above a gene's reactions, and as size: by flux, respiring and by flux, fermenting.

ATP by era

A reaction is dated by its oldest enzyme gene. In the model, enzymes shared with prokaryotes make 100% of the ATP in both conditions, and enzymes younger than LECA carry at most 0.07% of the flux. The introduction quotes both.

Reading it: the fluxes are the model's prediction of one optimal state. Other states can grow as fast, and a flux of 0 does not mean the cell never uses the reaction.

Links

702,499 links of five kinds join the genes. Each pair of genes counts once per kind, and each direction once for the directed kinds.

• In the panel, Links counts a gene's links of each kind. Click a kind to draw it; "drawn" marks the one on the drawing.
• Physical and genetic links are BioGRID's curation, as SGD serves it, including the global genetic interaction maps of Costanzo et al. 2010 and 2016.
• Reading them: a link says that some experiment related the two genes. A physical link may be indirect, through a shared complex. A genetic link is functional, not a contact. High-throughput screens supply most of the links.
• Finer kinds. The networks view splits each kind by what its evidence says: a contact from a co-complex, an activation from a repression (see Kinds of link).

Connections back in time

An interaction seen again in another species, between the counterparts of both partners (an interolog), was most simply already there in the two species' last common ancestor: it is at least as old as their split. The atlas tests each of the cell's 167,630 protein–protein interactions this way. 102,301 could be tested, and 9,285 are seen again beyond chance: 10 as far back as all life, 1,212 as far back as LECA, 7,882 as far back as opisthokonts and 181 as far back as Ascomycota.

The species

Counterparts are PANTHER's least-diverged orthologs, the closest corresponding protein and not its paralogs; the other copy from the whole-genome duplication shares them. The other species' interactions are BioGRID's physical ones, leaving out those whose partner is an RNA.

Species	Dates to
fission yeast Schizosaccharomyces pombe	Ascomycota (stop 5)
human Homo sapiens	opisthokonts (stop 3)
mouse Mus musculus	opisthokonts (stop 3)
rat Rattus norvegicus	opisthokonts (stop 3)
chicken Gallus gallus	opisthokonts (stop 3)
zebrafish Danio rerio	opisthokonts (stop 3)
fly Drosophila melanogaster	opisthokonts (stop 3)
worm Caenorhabditis elegans	opisthokonts (stop 3)
slime mould Dictyostelium discoideum	opisthokonts (stop 3)
thale cress Arabidopsis thaliana	LECA (stop 2)
E. coli Escherichia coli K-12	all life (stop 0)

Beyond chance

A match can happen by chance, more easily between well-connected proteins in a densely mapped network. The chance of a match is the probability that the counterparts would interact in a random network in which each protein keeps its number of partners, about k_xk_y/2M for proteins with k_x and k_y partners among M interacting pairs. Species that split from yeast at the same node are weighed together. At each node the threshold is set so that matches expected by chance are at most 10% of those it dates.

Since a stop

The deepest node that dates the interaction, but never before its younger partner exists.

Interactions back in time, in a gene's panel

How many of the gene's protein–protein interactions could be tested, which means both partners have counterparts in the other species' interaction data (3,445 genes have at least one). Then the dated partners, oldest first, each with the stop it is known since and the species that date it. "Seen as far back as …, but a partner is dated younger" marks an interaction older than a partner's own age: that partner's age call may be too young.

In a complex's panel

How many interactions between its subunits are seen again in other species, and the oldest.

Reading it: not seeing an interaction again dates nothing, since every other species' network is mapped incompletely. A date is a lower bound. By construction about 10% of the dated interactions may be chance matches.

Complexes, pathways and processes

A complex's panel

Its Complex Portal identifier and number of subunits. Oldest subunit: the stop of its oldest member. Came together: the first stop at which half its subunits exist. Complete: its youngest subunit's stop, or today when a subunit has no age call. "Present from … on" when its compartment is younger than its oldest subunit. Then its internal interactions dated by other species, its places, and its subunits from oldest to youngest.

A system's panel

Its kind and identifier, its parts and genes, and three stops: came together (half its parts exist), its oldest part, and all its parts. Then how many parts exist at the chosen stop, show on the cell, a bar for each era with its share of the parts, and the genes from oldest to youngest.

Which systems

Complexes from the Complex Portal, metabolic pathways from YeastPathways (through SGD), and processes from SGD's GO-slim terms for biological processes. The root and the six catch-all terms of more than 500 genes each are left out (biological process (the root; process unknown), biosynthetic process, transport, catabolic process, cellular component assembly, response to stress and small molecule metabolic process). A system needs at least 2 parts.

Reading it: "came together" is robust to one young part. The youngest part alone would make an ancient machine look young; the 25-residue RPL41A, for one, would put the cytosolic large ribosomal subunit in the budding-yeast family. A long gap between coming together and being complete means late additions, or old parts whose older relatives went undetected.

Heat

What happens to each protein when the cell is heated, in four measurements from different experiments. A gene's panel shows those it has, under Under heat.

mRNA, 10–20 min and 60–80 min fold change

Gasch et al. 2000: cells moved from 25 to 37 °C. The early response is the mean log₂ change at 10, 15 and 20 minutes; the late one at 60 and 80 minutes, by when most of the response has subsided. Shown as a fold: ×4 is four times more, ÷2 half, and "unchanged" within a tenth of a fold. Measured for 5,668 genes.

Shown as size: heat: mRNA up and heat: mRNA down, from the early response.

The stress response's group

Gasch et al. 2017's groups of the environmental stress response, for 828 genes: induced by every stress; or repressed by every stress, either the ribosome-biogenesis genes or the ribosomal proteins.

Nascent RNA with and without Hsf1 fold change at 5 minutes

Pincus et al. 2018: new RNA 5 minutes after 30 to 39 °C, with Hsf1 in the nucleus or anchored away from it, each against its own unheated sample (5,343 genes). Its induction needs Hsf1 when the gene is expressed under heat (at least 50 normalized counts), induced at least 4-fold with Hsf1, and without Hsf1 loses at least 1 log₂ unit of that induction and keeps at most half of it. 35 genes pass.

mRNA with and without Msn2/4 fold change

Solís et al. 2016: mRNA under heat in a strain whose PKA can be inhibited (TPK1/2/3-as), and in the same strain without Msn2 and Msn4, the general stress response's regulators (5,246 genes). The same rule gives its induction needs Msn2/4 (46 genes). Genes with several conflicting rows in either table are left out.

In solution percent of the protein

Wallace et al. 2015: the share of each protein left in the supernatant of a 100,000 g spin, at 30 °C (the mean of three replicates) and after 8 minutes at 37, 42 and 46 °C (3,085 proteins). Leaves solution by the first temperature at which the share falls by 0.3 or more.

Shown as size: heat: leaves solution: the fall from 30 to 46 °C.

By era, in the introduction

Among proteins mostly in solution at 30 °C (at least 0.6), the share that leaves by 46 °C:

Era	Leaves solution
shared with prokaryotes	16%
LECA-born	36%
opisthokont	47%
fungal and younger	38%

Reading it: the four measurements come from different laboratories, strains and temperatures, so read each on its own terms. Leaving solution means condensing, often reversibly, not necessarily damage. The "needs" calls are strict: a gene can depend partly on a regulator without passing. Proteins mostly pelleted even at 30 °C, membrane proteins for instance, cannot fall far.

The frontier

Ribosome profiling finds translation far beyond the annotated genes. Wacholder et al. 2023's reference translatome pools hundreds of experiments: 26,344 translated ORFs, of which 5,280 are annotated genes, 228 are Dubious ORFs, and 20,836 lie outside the genes. The atlas places each of these against the annotated protein-coding genes around it, in the first of these contexts that holds:

Transient

An ORF not held by purifying selection across the Saccharomyces species, by Wacholder et al.'s test: 13,848 of those outside the genes. An ORF that overlaps a gene shares that gene's conservation, so its status says little about its own age. The transient ORFs upstream, downstream and between genes are the frontier proper.

Class

Rich et al. 2024's evolutionary classification of the ORFs in their coexpression data: conserved, de novo (arisen from non-genic sequence), or other (5,777 of the ORFs outside the genes). The panel names an ORF de novo when Rich et al. do, and otherwise says whether it is transient.

Moves with growth, stress, both or neither

The side of yeast's main axis of expression an ORF, or a gene, rises and falls with. Rich et al. measured each ORF's coexpression across thousands of RNA-seq samples and tested it for enrichment in SGD's GO slim processes. An ORF is on the growth side when its coexpression is significantly and positively enriched (adjusted p < 0.01) for ribosome biogenesis, rRNA processing, cytoplasmic translation, ribosomal large subunit biogenesis, ribosomal small subunit biogenesis. It is on the stress side for response to stress, carbohydrate metabolic process, response to heat, response to starvation, response to oxidative stress, generation of precursor metabolites and energy, cellular respiration, and it can be on both. 5,777 ORFs outside the genes have a side. A gene's own side comes from its row in the same data.

The frontier around it, in a gene's panel

The gene's own side; how many translated ORFs sit by it and how many of them are transient; then up to ten, in genome order, each with its context, its length in amino acids (not counting the stop codon), its class or status, and its side.

Shown as size: young ORFs around it. In the introduction, 72% of the de novo ORFs move with stress.

Reading it: translation is evidence that a ribosome read the ORF, not that its product does anything. ORFs are placed against the annotated protein-coding genes only, not against the RNA genes.

The networks

Each linked pair of genes has one kind of link in each family, and each kind keeps its own sign: an effect (activates or represses) is not a chemical mark (put on or taken off), which is not an epistasis (two losses worse or better together).

Kinds of link

Physical pairs BioGRID

A pair takes the first kind in this list that one of its experiments shows:

Kind	Experiments
contact	Two-hybrid, PCA, Reconstituted Complex, Co-crystal Structure, FRET, Far Western, Protein-peptide, Cross-Linking-MS (XL-MS), Thermal Shift Assay
co-complex	Affinity Capture-MS, Affinity Capture-Western, Affinity Capture-Luminescence, Co-fractionation, Co-purification, Proximity Label-MS
acts on	Biochemical Activity
to an mRNA	Affinity Capture-RNA, Protein-RNA
co-localization	Co-localization

For acts on and to an mRNA, the bait is the actor: the enzyme, or the protein that pulled the transcript down. A contact or co-complex pair whose bait also acts on its hit (1,037 pairs) counts as acts on in the control network.

Regulation from the regulator to its target

41,267 pairs from SGD's regulation, each over all its records. A pair with records that raise the target activates, one with records that lower it represses, one with both is dual, and one whose records say neither has its sign unknown (37,038 pairs).

• Binding only: 24,656 pairs are seen only by binding assays, such as chromatin immunoprecipitation: the regulator sits at its target, with no sign and perhaps no effect.
• Only under heat: 14,881 pairs are recorded only in heat, almost all from one ChIP-chip study at 37 °C (Venters et al. 2011, doi:10.1016/j.molcel.2011.01.015).
• Targeted: 3,502 pairs have a record from an assay of one gene rather than of the whole genome at once. Their signs are the control for the signs of genome-wide screens, which are often stress experiments.
• Self-regulation: 78 regulators act on their own gene (62 sign unknown, 8 activates, 8 represses). These are kept apart from the links; a gene's panel says when it regulates itself.

Modification from the enzyme to its substrate

SGD's modification sites with a named enzyme. The sign is chemical: the enzyme puts a mark on, or takes one off (dephosphorylation, deacetylation, deubiquitination, demethylation, desumoylation or deglutathionylation). Whether the mark turns the substrate on or off is not recorded.

Genetic undirected

BioGRID's genetic interactions, aggravating (negative) or alleviating (positive). An epistasis says how two losses combine, not what one gene does to the other, so it is never read as activation.

Monotone or not

A signed network is monotone when its genes split into two camps with every activation inside a camp and every repression between the two. No loop then has an odd number of repressions, whichever way its links point, and the network's responses are orderly: pushed, it settles rather than oscillating (Ma'ayan et al. 2008, doi:10.1049/iet-syb:20070036). The frustration index is the fewest links that break every split: the links the network would have to lose to become monotone (Iacono & Altafini 2010, doi:10.1186/1752-0509-4-83).

The signed network

The regulation pairs with one sign, without self-regulation: 4,172 links, 1,580 of them repressing, among 2,332 genes. The 41 dual pairs and the unsigned ones are left out, and the links are read without their direction.

Frustrated links of the signed links

The best split found leaves 366 links frustrated (8.8%). With the signs shuffled among each regulator's own links, so that each keeps its mix of activations and repressions, the best split leaves 599 ± 28 (mean ± sd of 100 shuffles). With the links rewired, each gene keeping its numbers of links in and out, 580 ± 40.

Shown as double lines, in the summary line, and in a gene's panel.

How the split is found

The best split is the lowest-energy state of an Ising spin glass whose couplings are the signs. It is searched for by annealing from 21 starts, one from the leading eigenvector of the signed links and the rest random, each cooled over 400 sweeps and then descended until no single gene's move helps. A search can miss the best split, so the index is an upper bound. The shuffled and rewired networks, the leave-one-out runs and the stops use a quicker search (4 starts of 200 sweeps), which overestimates a little: searched that way, today's network leaves 385 links frustrated, still far below its shuffles.

At each stop

The same measure on the signed links whose two genes exist at each stop, against 10 sign shuffles, with and without the busiest regulator (below). At LECA, the first stop with a hundred signed links, 24 of 856 are frustrated, against 73 with the signs shuffled: the oldest wiring is the nearest to monotone, also against its own shuffles.

Stop	Signed links	Frustrated	Shuffled signs	Without ZNF1	Frustrated	Shuffled signs
all life	1			1
from archaea or bacteria	60	0 (0.0%)	0	60	0 (0.0%)	0
LECA	856	24 (2.8%)	73	856	24 (2.8%)	72
opisthokonts	1,325	63 (4.8%)	140	1,325	60 (4.5%)	144
Dikarya	1,609	80 (5.0%)	178	1,609	86 (5.3%)	180
Ascomycota	1,741	105 (6.0%)	193	1,741	103 (5.9%)	199
Saccharomyceta	3,655	373 (10.2%)	507	2,890	212 (7.3%)	330
budding-yeast family	4,011	405 (10.1%)	577	3,200	248 (7.8%)	360
after the duplication	4,014	398 (9.9%)	566	3,202	230 (7.2%)	361
Saccharomyces	4,056	404 (10.0%)	583	3,242	231 (7.1%)	363
S. cerevisiae	4,094	401 (9.8%)	592	3,277	235 (7.2%)	369
today	4,172	385 (9.2%)	614	3,342	234 (7.0%)	382

Reading it: every loop with an odd number of repressions has at least one frustrated link, but which links are marked depends on the split: another split as good may mark others. A sign is what some experiment saw, not what happens in every condition.

The two camps

The best split puts each gene of the signed network in one of two camps, of 876 and 1,456 genes. The split does not say which camp is which. The one holding more of the genes that move with growth (their side of expression, as under The frontier) is called the growth camp, the other the stress camp. Against the side of expression the odds ratio is 11.2 (p 1 × 10⁻¹⁰⁷). Against the stress response's groups, with the repressed genes in the growth camp and the induced ones in the stress camp, it is 76.4 (p 4 × 10⁻⁷⁵).

Many signs come from genome-wide expression screens, often made under stress, so the camps could echo the experiments rather than the cell. The controls:

Reading it: the monotone structure survives every control. With only targeted signs, or without ZNF1, the network stays far less frustrated than its shuffles. The reading of the camps as growth and stress does not survive. ZNF1 alone has 830 signed links, and 824 of its records come from one study (PMID 38331422). Without its links the camps no longer sort growth from stress, while leaving out any one of the next 19 busiest regulators changes little. Until other signs confirm them, the camps' names are labels for the two sides of the split.

Loops

Feed-forward loops X on Y, Y on Z, and X on Z

Among all regulation links, 48,118, against 54,813 ± 1,744 in networks rewired with every gene's links kept (10 rewirings). In 365 all three links are signed: 294 are coherent, the direct link having the sign of the path through Y, and 71 incoherent. That is 81% coherent, against 57.9% ± 2.3% with the signs shuffled (Mangan & Alon 2003, doi:10.1073/pnas.2133841100; Kaplan et al. 2008, doi:10.1038/msb.2008.43).

X on Y	Y on Z	X on Z		Loops
activates	activates	activates	coherent	161
activates	represses	represses	coherent	61
represses	activates	represses	coherent	51
represses	represses	represses	incoherent	41
represses	represses	activates	coherent	21
activates	represses	activates	incoherent	12
activates	activates	represses	incoherent	10
represses	activates	activates	incoherent	8

Shown in a gene's panel, as the number of loops it is in.

Feedback loops

Pairs that regulate each other: 41, against 26 ± 5 rewired; 1 with both links signed. Three genes in a cycle: 97, against 104 ± 12; 2 with all three signed.

Reading it: an incoherent loop can make a response rise and then fall as its signal grows, but telling whether one does takes measured responses, which the atlas does not have. It calls no loop non-monotone.

The bow-tie

The control network, its regulation, modification and acts-on links with their direction, has the shape of a bow-tie: a core of 394 genes that all reach one another along the links, a few genes that feed it, and the many it reaches.

Reading it: the shape depends on the evidence. Without the links that only binding assays show, the core has 310 genes and many more genes fall outside the network, so the view says which evidence it draws.

Modules

The contact network (contact and co-complex pairs) and the coupling network (aggravating pairs) are each split into modules by the Louvain method, at resolution 1 with a fixed seed. Modules of fewer than 10 genes are left out, and their genes join the tray. Each module is named for the GO-slim process most over-represented among its genes, against the genes of its network (a hypergeometric p, not corrected for the many processes tried, with at least 5 of the module's genes in the process).

Reading it: a module's name is its most over-represented process, which may hold only a small share of its genes. Another method or resolution would split the networks differently.

Roles

Almost every gene is in some network; what differs is its role. Each gene takes the first role in this list that holds. The last two columns give each role's share of the cell's protein-building budget and of its protein mass (copies × length).

Matter

What the networks act on, lumped: the biomass of the metabolic model, weighed from its metabolites' formulas. Each class of the biomass is what its pseudoreaction takes in less what it gives back; a charged tRNA less the free tRNA, for instance, is a protein's residue. An input with no formula is opened up through the pseudoreaction that makes it (the lipids' backbones and chains), and a chain's point of attachment weighs nothing. Per gram of dry weight: protein 0.465, carbohydrate 0.383, RNA 0.064, lipid 0.035, cofactors 0.005, DNA 0.004, ions 0.003, 0.958 g in all (0.00148 mmol of the lipid has no formula to weigh).

Shown in the matter band above the control network's bow-tie and in the compartments in the cell, each disc's area following its mass.

What makes each pool mg per gram of dry weight per hour

In each growth condition of the metabolic model the cell lays each pool down at its weight times the growth rate. Its makers are found in the flux balance solution's mass flow graph, which joins two reactions when one makes a metabolite the other uses, weighed in grams (Beguerisse-Díaz et al. 2018, doi:10.1038/s41540-018-0067-y). From the pseudoreaction that gathers the pool, each unit is followed back to the reactions that make it, in proportion to how much each makes:

• a reaction with genes is a maker, and the trail stops there;
• a reaction with no genes (a transport the model leaves unassigned, the lipid bookkeeping) passes its share back to the inputs that end up in its products;
• a carrier (the 9 kinds of nucleotide, FAD, FMN, NAD, coenzyme A, folates and glutathione) counts as one pool with its family: its makers are the reactions that make its skeleton, not the ones that recharge it. The NAD+ in new biomass was NADH moments before, but NAD synthase made it.

Pool	g per g dry weight	respiring	fermenting	Largest makers, respiring
protein	0.465	37.7	94.6	9% lysyl-tRNA synthetase (KRS1); 9% leucyl-tRNA synthetase (CDC60); 8% glutamyl-tRNA synthetase (GUS1)
wall glucans	0.160	13.0	32.5	75% 1,3-beta-glucan synthase (FKS1, FKS3, GSC2, RHO1); 25% 1,6-beta-glucan synthase (KRE6, SKN1)
mannan	0.114	9.2	23.2	100% dolichyl-phosphate-mannose--protein mannosyltransferase (PMT1, PMT2, PMT3, PMT4 …)
RNA	0.064	5.2	13.1	29% orotidine-5'-phosphate decarboxylase (URA3); 25% GMP synthase (GUA1); 22% CTP synthase (NH3) (URA7, URA8)
glycogen	0.058	4.7	11.8	100% glycogen (starch) synthase (GLG1, GLG2, GSY1, GSY2)
trehalose	0.047	3.8	9.5	100% trehalose-phosphatase (TPS1, TPS2, TPS3, TSL1)
membrane lipids	0.021	1.7	4.2	49% C-s24 sterol reductase (ERG4); 15% phosphatidyl-N,N-dimethylethanolamine methyltransferase (1-16:0, 2-18:1), ER membrane (OPI3); 8% phosphatidyl-N,N-dimethylethanolamine methyltransferase (1-16:0, 2-16:1), ER membrane (OPI3)
acyl chains	0.011	0.91	2.3	27% phosphatidyl-N,N-dimethylethanolamine methyltransferase (1-16:0, 2-18:1), ER membrane (OPI3); 15% diacylglycerol acyltransferase (1-16:1, 2-16:1, 3-16:1), ER membrane (ARE1, ARE2, DGA1); 13% phosphatidyl-N,N-dimethylethanolamine methyltransferase (1-16:0, 2-16:1), ER membrane (OPI3)
cofactors	0.005	0.40	1.00	89% NAD synthase (QNS1); 8% riboflavin synthase (RIB5); 3% dephospho-CoA kinase (CAB5)
chitin	0.005	0.39	0.98	100% chitin synthase (CHS1, CHS2, CHS3)
DNA	0.004	0.32	0.79	30% thymidylate synthase (CDC21); 29% ribonucleoside-triphosphate reductase (ATP) (TRX2); 21% ribonucleotide reductase (RNR1, RNR2, RNR3, RNR4)
storage lipids	0.003	0.23	0.57	91% acyl-CoA:sterol acyltransferase (oleoyl-CoA:ergosterol), ER membrane (ARE1, ARE2); 9% diacylglycerol acyltransferase (1-16:1, 2-16:1, 3-16:1), ER membrane (ARE1, ARE2, DGA1)
ions	0.003	0.20	0.51	69% oxaloacetate/sulphate antiport, mitochondrial (OAC1); 9% sulfate uniport (SUL1, SUL2); 6% Mn(2+) transport (ALR1)

Shown as ribbons with mass flow on, and in a pool's and a gene's panel.

By age of each reaction's oldest enzyme

Dated as the ATP ledger dates reactions, by their oldest enzyme: of the mass laid down, 65% by families shared with prokaryotes, 35% by LECA-born ones and 0.0% by younger ones. 51% of the dry mass is not protein.

The mass budget mg per gram of dry weight per hour

What the cell takes in and gives out under each condition, from the model's exchanges and the new biomass by class, weighed by element. Given out less taken in, it closes to within 1.3% respiring and 0.5% fermenting; most of the gap is between the gram of new cell the model counts and the 0.958 g its formulas weigh.

Condition	Growth per hour	Taken in	Given out and built	Closure
respiring	0.081	D-glucose 180, oxygen 75.3, ammonium 9.6, phosphate 2.2	carbon dioxide 117, H2O 69.4, protein 37.7, carbohydrate 31.1, RNA 5.2, lipid 2.8	-1.3%
fermenting	0.204	D-glucose 1,802, oxygen 32.0, ammonium 24.2, phosphate 5.5	carbon dioxide 785, ethanol 723, protein 94.6, H2O 78.9, carbohydrate 78.0, glycerol 72.2	-0.5%

Shown in a pool's panel, as three columns joined by bands.

Outside the networks

The genes no link of a network reaches, and their share of the protein mass and of the building budget:

Network	Genes	Protein mass	Budget
control	277	0.18%	0.19%
contact	452	0.01%	0.01%
coupling	627	0.06%	0.06%

Small molecules acting on proteins

Matter acts back on the enzymes that make it. BRENDA's records of activators and inhibitors for S. cerevisiae (BRENDA (Hauenstein et al. 2026, doi:10.1093/nar/gkaf1113)) give 717 links from 181 of the model's metabolites to the enzymes of 275 genes: 568 inhibit, 138 activate and 11 do both. 41% come from carriers (nucleotides, NAD(P), coenzyme A and the rest), and 42 are called allosteric or feedback in BRENDA's comments.

How a record is placed

Of BRENDA's 5,625 records for the organism, 872 are kept. A record's protein goes to its genes by its UniProt accessions (77 links) or, when BRENDA gives none, by its EC number (640 links), if no more than 3 genes carry that number in their UniProt protein names. A record's compound goes to a metabolite of the model by its name, after setting aside case, a leading D- or L- and punctuation, or through a short list of synonyms (NAD+ for NAD, CoA for coenzyme A), and is kept only if that metabolite is organic.

Records dropped because	Records
the compound is not in the model	4,516
inorganic (a metal, a salt or water)	343
no gene for the protein	91
its EC number is carried by more than 3 genes	41

The molecules acting on the most enzymes

Metabolite	Enzymes	Activates	Inhibits	Both
ATP	35	9	24	2
AMP	28	4	23	1
ADP	27	7	20	0
glutathione	24	10	13	1
L-cysteine	20	9	11	0
NADH	14	0	14	0
NAD	14	2	12	0
acetate	14	7	7	0
citrate	13	4	9	0
coenzyme A	12	0	12	0
L-methionine	12	1	11	0
NADP(+)	11	0	11	0
UDP	11	0	11	0
Oxalate	10	0	10	0
UMP	10	0	10	0

Shown as squares in the matter band with small molecules, in a molecule's panel, and in a gene's panel.

Reading it: BRENDA records what was seen on the enzyme, mostly purified, in the test tube: an effect at some concentration, not always one the cell reaches. Reducing agents that assays add, glutathione and cysteine, are among the most frequent molecules for that reason. A record placed by its EC number goes to every gene that carries the number, which can credit an isozyme with its partner's regulation. Many links close a loop on the matter: CTP inhibits CTP synthase, UDP the glucan and chitin synthases, glucose 6-phosphate activates glycogen synthase.

Reading it: almost every gram of protein sits inside the networks; the matter outside them is the half of the dry mass that is not protein. A pool's makers are the model's: one optimal routing, with a reaction's isozymes sharing its grams equally, and the model's own bookkeeping (it builds RNA from nucleotide monophosphates, for instance). The biomass's make-up is fixed in the model, so the conditions differ in how fast each pool is laid down, not in who makes it.

Flows compared

Three things flow through the cell's genes: information, energy and matter. Each is measured gene by gene, in one unit per flow, and set beside the others by era, by contact module and by role. Dot size can follow each flow (see Size); in the networks view a gene's panel gives its three flows, and the introduction draws each flow's shares by era.

Information

In the contact network (5,593 proteins, 158,396 links: contacts and co-complex associations), information is current flow (Missiuro et al. 2009, doi:10.1371/journal.pcbi.1000350). The network is an electrical circuit, each link a resistor of one ohm; a unit of current is sent from every gene to every other in turn, and a gene's information flow is the current that passes through it, summed over the pairs, as a share of all of them. Unlike counting shortest paths, it counts every path. It is computed exactly, and again on direct contacts alone (4,872 proteins, 27,434 links).

In the control network, information is what a gene acts on: its links out, and every gene downstream of it. 763 proteins act on others; half of them reach at least 5,874 of the 6,021 proteins downstream, and 397 reach at least half of them: reach saturates, so the comparisons use the links out.

Shown as size: by information flow, and in a gene's panel in the networks view, with its place among genes of about as many contacts.

Energy mmol ATP per gram of dry weight per hour

ATP that flows: the ATP and GTP that join a protein's residues as its copies are built, four per peptide bond (1.36 respiring and 3.40 fermenting in all), and the ATP a gene's reactions make or spend in the metabolic model (18.5 respiring and 75.5 fermenting, made and spent). Making the amino acids is counted in the enzymes that make them. This is not by cost, which counts, as Wagner does, also the energy forgone by turning carbon into amino acids.

Shown as size: by energy (building and reactions together), and in a gene's panel.

Mass grams per gram of dry weight per hour

The matter a gene's reactions process (8.2 g respiring and 46.6 g fermenting in all). A carrier counts only as far as a reaction uses it up, as the mass flow traces carriers (see Matter): ADP made into ATP, or cytochrome c reduced, is no mass, while the glucose that UDP-glucose gives to glycogen is. Water and protons do not count.

Shown as size: by mass flow, and in a gene's panel.

From a reaction to its genes

A reaction's ATP and mass pass through its genes as the mass flow shares them out: each isozyme an equal share, every gene of an isozyme (a complex needs all its parts) all of it. That is what a gene's panel and its dot show. For an era's, a module's or a role's share of a flow, each isozyme's share is split among its genes instead, so that the totals are kept.

Each flow by era

Reading it: three flows run through genes of three ages. In the metabolic model (respiring), genes shared with prokaryotes carry 78% of the matter processed (95% fermenting) and 65% of the ATP through reactions. LECA-born genes carry 47% of the current in the contact network and 51% in direct contacts, though they are 35% of the proteins. 63% of the control network's links come from genes younger than LECA, 96% of those from transcription factors. By each reaction's oldest enzyme, as the ATP ledger dates it, 99% of the ATP goes through reactions shared with prokaryotes; split among the genes, 30% of it reaches genes younger than LECA, 90% of that the small accessory subunits of ATP synthase, whose calls may be too young (see What not to read into it).

Against degree, rewired networks and abundance

Information flow follows a gene's number of contacts closely. Each gene is compared with genes of about as many contacts (10 bins of equal size; a percentile of 0.5 means the era carries what its degrees give) and with the same gene in networks rewired so that every gene keeps its number of contacts (3 rewirings of 10 swaps tried per link; over 1, more current than the rewired networks carry).

Reading it: the two disagree: in the contact network the genes shared with prokaryotes carry the least current for their degree, in direct contacts alone the most. Co-complex associations join a bait to every prey it pulls down, so each complex is a dense clique that spreads a member's current over its partners; beyond degree, the answer turns on whether they count. Against genes of their degree, within each of the 5 bands of abundance, LECA-born genes still carry more current for their degree than those shared with prokaryotes (the report has the table). Abundant proteins have more detected contacts and cost more to build, which is why building energy follows information (Spearman 0.67).

Do the same genes carry the flows?

A flow's carriers are the fewest genes that carry 50% of it. Two flows' carriers, against two random sets of the same sizes:

Reading it: across all proteins information hardly follows ATP or mass (Spearman 0.06 and 0.07), and the control network's carriers share nothing with the metabolic ones. Where information and mass meet (10 genes, where chance gives 3.1), it is mostly in the cell's most abundant enzymes: ATP6, CDC19, ATP1, ATP3, PGK1, TPI1, TDH3, PFK1, TDH1, TDH2, FBA1, MDH1, GPM1, ATP4.

By module and role

The heat network against the unconditioned one

The only heat-conditioned study in SGD's regulation, Venters et al. 2011 (ChIP-chip at 37 °C), is set against a ChIP-chip study in rich medium, MacIsaac et al. 2006, for the 32 regulators both profile, so that the two differ in the condition and the laboratory but not in the assay. They find 7,934 and 1,542 links, and agree on 185 (Jaccard 0.02).

Reading it: no sign of heat in the difference. The targets bound only in heat are 24% shared with prokaryotes, like the genome (23%), and they are no more often induced by heat: 21.2% of the links seen only in heat lead to a gene heat induces, against 21.7% of those seen only in rich medium (odds ratio 0.97, p 0.69). The two studies differ mostly as studies: the only under heat evidence of the networks view marks one study, not the heat network.

Signs and dynamics

SGD signs 4,172 of the regulation links and leaves 37,017 unsigned, most of them seen only by a binding assay. Two experiments say what a regulator does to a gene: deleting it and watching every other gene, or inducing it and following every other gene through a time course (Kemmeren et al. 2014, Cell 157:740, doi:10.1016/j.cell.2014.02.054; Hackett et al. 2020, Mol Syst Biol 16:e9174, doi:10.15252/msb.20199174).

What a deletion says

The deleteome measures every gene's expression in about 1,500 deletion mutants. A gene that falls when a regulator is deleted is one the regulator activates; one that rises is one it represses. A change counts at p below 0.05 and a fold change of 1.7 or more, the deleteome's own cut-offs. That gives 44,263 effects, 67% of them repression.

What an induction says

IDEA induces one transcription factor at a time and fits each gene's time course, either as a rise that stays up or as a rise that falls back. The fitted asymptote gives the sign, its half-max time gives the speed, and the falling fit marks a transient response. That gives 44,110 responses reaching the same fold change.

Together

83,438 regulator-and-target pairs in all, from 1,120 regulators onto 5,382 genes. Where a deletion and an induction both see a pair (1,693), they agree 90.5% of the time; against SGD's curated signs, 78.2% of the 825 pairs both give.

Shown as evidence: sign measured in the networks view, which draws each measured link with the sign the perturbation found, and in a gene's panel.

Binding is not regulation

Of the 37,017 unsigned links, 28,563 have a regulator that was deleted or induced. Only 16.2% of those move the target enough to count, which signs 4,638 links and grows the signed network from 4,172 to 8,810.

The camps, on signs no curator wrote

Step 9 found the control network close to monotone and its camps lined up with growth and stress, but only through one regulator's genome-wide signs (see The two camps). The same split, run on each set of signs, settles it:

Reading it: the camps are growth and stress after all. On the measured signs alone, which ZNF1 hardly touches, they line up with the stress response at odds 9.6 (p 7 × 10⁻³⁰), where the curated signs without ZNF1 give 1.4 (p 0.29). The measured signs are the more frustrated: a perturbation carries every indirect effect with it, which curation leaves out.

The shape and speed of a response

Reading it: the age of the gene that answers says little about how it answers. Among the 2,797 genes that at least 5 factors move, the median half-max time runs 23 to 30 minutes across the eras, and transient responses are rare everywhere. The flows differ sharply by era (see Flows compared); the timing does not.

Readout, key and introduction

The readout

The first line gives the chosen stop, its date and what it means. The second gives what the cell holds there: the compartments that exist, the genes drawn, the complexes present (with how many are complete), the reactions present, and the links whose two genes are both drawn, of every kind.

The key

The eras with their numbers of genes (or, with the colour off, one entry for all), grey for no call or undated (385 genes), the two shapes (crossed out when that kind is hidden), and how genes are placed in each compartment.

The introduction

The panel with nothing selected: the atlas's counts, how to read the drawing, and a paragraph for each layer with its main finding, then the datasets and their licences.

The data

Datasets and licences

The atlas is built from 40 public datasets, each pinned by its address and, where its host does not version it, by its bytes. "Used for" links to where this page describes what each one feeds.

The atlas adapts these datasets: it joins them on SGD's gene names, filters them and dates every element. Mistakes in that are the atlas's, not the sources'. YGOB states no licence and asks to be cited. NCBI places no restrictions on GEO data.

Choices made by hand

Some inputs are written by hand rather than downloaded, each with its reasons and sources:

• The dates of the stops, from TimeTree 5 (retrieved 2026-09-18).
• The compartments: which GO terms and model compartments each one takes, what it is drawn inside, and when it appears (the table under Place).
• The ages of RNA classes (under Age).
• The amino-acid costs and the four bonds per peptide bond (under Cost).
• The growth conditions of the metabolic model (under Energy).
• The processes left out of the systems, and the smallest system (under Complexes, pathways and processes).
• The species and the chance threshold for dating interactions (under Connections back in time).
• The cut-offs of the heat layer and the frontier (under Heat and The frontier).
• The networks' choices: which experiments make each kind of physical link, which regulation assays are binding and which genome-wide, which marks are taken off, the settings of the searches, shuffles and modules, the pools the matter is lumped into and where each is drawn, the carrier families, and how BRENDA's records are placed: the synonyms, the organic-only rule and the limit on genes per EC number (under The networks).
• The flows' choices: which links make the contact network and which the direct contacts, the rewirings, the bins of degree and abundance, the carriers that cycle besides the carrier families, and the two studies of the heat comparison (under Flows compared).
• The signs' cut-offs: what counts as a change in a deletion mutant or an induction time course, how many factors must move a gene before its timing is summarised, and which regulator the camps are tested without (under Signs and dynamics).

What not to read into it

• Ages are of gene families. A gene's age is its deepest detectable relative, not the age of its function, and a call can be too young when relatives went undetected. The flags mark the calls to handle with care.
• One place per dot. Many proteins move between compartments; each is drawn once. The distance from a compartment's centre shows age, not position.
• A compartment's size shows its number of genes, not its volume or its protein mass.
• Absence dates nothing. An interaction not seen in another species may simply be unmapped there.
• The fluxes are one optimal solution of a model, not measurements.
• The heat data are four experiments, in different strains and at different temperatures.
• Translated is not functional. The frontier's ORFs are read by ribosomes; most have no known role.
• Small dots can be unmeasured. Under every size choice a gene without the quantity is drawn at the smallest size.
• Colour groups stops. Five eras colour twelve stops; the stop itself is in the tooltip and the panel.
• Most regulation has no sign. 37,038 of the 41,267 regulation pairs have none, and a sign is what one experiment saw.
• The camps' names rest on one study's signs. The split into two camps is robust; its reading as growth against stress is not (see The two camps).
• Binding is not regulation. A regulator seen at a promoter may do nothing to the gene there.
• The frustration index is an upper bound, from a heuristic search.
• Modules are one method's answer, at one resolution.
• The mass flow is a model's. Who makes a pool comes from one optimal flux solution, with a reaction's isozymes sharing its grams equally.
• A small molecule's effect is a test tube's. BRENDA's activators and inhibitors were mostly seen on purified enzymes, at concentrations the cell may never reach.
• Information flow is a circuit's. Every contact is the same resistor, and contacts are as the assays found them: abundant proteins are found in more of them.
• Small subunits may be dated too young. A short, fast-changing protein can lose its detectable relatives, which makes old machines' accessory subunits look fungal; their share of the energy and mass flows is a share of those calls.
• The heat comparison is two laboratories. No study in the data measured the same regulators with and without heat.
• A measured sign is an effect, not a mechanism. Deleting or inducing a regulator moves genes through everything it touches, not only the genes it binds.
• An induction is not the cell's own dose. IDEA drives a factor far above its usual level, and a deletion removes it altogether.

Glossary

ORF, systematic name

An open reading frame; SGD names each gene by its place, for example YGL008C (chromosome VII, left arm, 8th ORF, on the Crick strand).

Dubious ORF

An ORF SGD judges unlikely to encode a protein. The gene table leaves them out.

LECA

The last eukaryotic common ancestor. A gene "at LECA" is in every major group of eukaryotes.

Opisthokonts, Dikarya, Ascomycota, Saccharomyceta

Nested groups on yeast's side of the tree: fungi with animals; the higher fungi; the sac fungi; the budding yeasts with the filamentous sac fungi.

Whole-genome duplication, ohnolog

An ancestor of S. cerevisiae doubled its genome; the surviving pairs of copies are ohnologs.

Ortholog, least-diverged ortholog, interolog

The corresponding gene in another species; of several, the one closest to the yeast gene; an interaction seen again between two orthologs.

ppm

Parts per million.

~P, ATP equivalent

One high-energy phosphate bond, the currency of the costs.

Flux balance analysis, parsimonious

Finding a steady state of a metabolic network that maximises growth; parsimonious: the one with the least total flux.

psup

The proportion of a protein in the supernatant after centrifugation: how much stays in solution.

Environmental stress response

The genes that every stress switches on or off together (Gasch et al.).

Hsf1, Msn2/4

Heat-shock factor 1, which drives the chaperones; Msn2 and Msn4, which drive the general stress response.

Anchor-away

Removing a nuclear protein by tethering it outside the nucleus with rapamycin.

Ribosome profiling, iRibo

Sequencing the RNA fragments that ribosomes protect, to see what is translated; iRibo pools many such experiments.

Transient ORF, de novo ORF

A translated ORF not preserved by selection; one that arose from sequence that did not code before.

GO slim

A small set of broad Gene Ontology terms that SGD maps every gene to.

Configuration model

A random network in which every node keeps its number of links; the null model of the interaction dates.

Monotone network, camp

A signed network whose genes split into two camps, with every activation inside a camp and every repression between them.

Frustration index, frustrated link

The fewest links that disagree with the best split into two camps; those links.

Rewiring, sign shuffle

The two nulls: links swapped between pairs of genes so that every gene keeps its numbers of links in and out; signs permuted among each regulator's own links.

Feed-forward loop, coherent, incoherent

X acts on Y and on Z, and Y on Z; coherent when the direct link has the sign of the path through Y, incoherent when it has the other.

Bow-tie, strongly connected

A directed network's shape: a core in which every gene reaches every other (strongly connected), what reaches it, and what it reaches.

Louvain method

A way of splitting a network into modules with more links inside them than between them.

Epistasis: aggravating, alleviating

Two losses together hurting more, or less, than expected from each alone.

ChIP, binding assay

Chromatin immunoprecipitation: finding where on the genome a protein sits.

Mass flow graph

Reactions joined when one makes a metabolite that another uses, each join weighed by the grams that pass along it in a flux solution.

Allosteric

Acting on an enzyme at a site other than where it works on its substrate; feedback inhibition is the end product of a pathway inhibiting an enzyme early in it.

Current flow, information flow

The current through a node when a network of unit resistors carries a unit between every pair of nodes, summed over the pairs; unlike shortest-path betweenness, it counts every path.

Deletion profile, induction time course

Every gene's expression measured in a mutant missing one gene; or followed through time after one factor is switched on.

Transient response

A response that rises and falls back while the factor stays on, rather than rising and staying up.

Carriers of a flow

The fewest genes that together carry a given share of it.

Carrier family

A nucleotide, NAD(P), coenzyme A, a flavin, a folate or glutathione, whatever it carries: made once, then recharged many times.

Pseudoreaction, pool

A metabolic model's bookkeeping reaction that gathers a class of the biomass (protein, lipid and so on); the lumped matter the atlas draws from it.

Credits and citing

The atlas is a working draft. If you use what it shows, cite the datasets behind it (above), and the atlas as: Whole-cell atlas of budding yeast, thinkingembryo.com/whole_cell_atlas, build of 2026-09-29.

This page and the atlas are generated by the same build. A number here that disagrees with the atlas is a bug.