Which genes does yeast need to live?

In 2002 the Yeast Deletion Project removed almost every gene of baker’s yeast, Saccharomyces cerevisiae, one at a time. About one gene in five turned out to be essential: without it the cell cannot grow. What makes those genes special – and could a computer learn to recognise them from data alone?

This practical is a taste of how a lot of bioinformatics is done now, and increasingly will be: a table of data about thousands of genes, a Python notebook, machine learning, AI models of proteins (a protein language model and AlphaFold) and an AI coding assistant that writes code alongside you. The biology stays in charge: at every step you will check what the computer – and the AI – tells you.

≈ 2½–3 hoursPython in the browser – nothing to installNo programming experience needed

explore the data→test hypotheses→train a model→protein language model→AlphaFold→learn from the mistakes

What you will learn

to explore a table of biological data with pandas and plot it with matplotlib;
to choose a statistical test and report an effect size, not just a p-value;
to train and evaluate a machine-learning model honestly – baselines, cross-validation, ROC AUC – and to spot data leakage;
what a protein language model and AlphaFold add, and what their numbers mean;
to use an AI coding assistant productively – and to catch its mistakes.

How this page works

The instructions are on the left; the workbench is on the right:

Tab	What it is
Notebook	Real Python (with pandas, scikit-learn and matplotlib) running inside this web page. Code goes in cells; run a cell with ▶ or `Shift`+`Enter`.
AI assistant	Ask for code, explanations or help with errors, and insert its code straight into the notebook.
Data	The gene table to browse without code, and what every column means.
3D viewer	AlphaFold’s predicted structures of yeast proteins.

About the AI assistant

The assistant starts in guided mode: its answers to this practical’s questions were generated in advance by an AI model and checked. Like real AI output, some of them contain mistakes on purpose – a wrong label, an unsuitable test, a leaky model, a wrong reference. Finding them is part of the practical, so read every answer as you would a colleague’s work.

You can also press Copy prompt to take a question (with a description of the data and your code) to any AI tool you have access to, or connect a live model with ⚙ in the assistant, using an API key – your own (Google Gemini has a free tier) or one from your lecturer.

First time with Python?

That’s fine. Every step has a ▶ button that puts working code into the notebook, and the assistant can explain any cell. The chapter has a cheat sheet.

Activity 0.1 · Start Python

Look at the Notebook tab: the first time, Python takes 10–30 seconds to download. Wait for Python … ready at the top right of the notebook.
Say hello to the assistant:

Your name or student number (goes on your downloaded answers):

The notebook and the data

≈ 15 minWorkbench: Notebook

A notebook mixes code, results and plots in one document. It remembers what you have run: a variable created in one cell can be used in the next.

Activity 1.1 · Load the data

In the Notebook, run the first cell (click ▶ beside it, or click in it and press Shift+Enter). It reads the file data/yeast_genes.csv into a table called df and shows the first five rows.
Run the next cell, df.shape: the number of rows (genes) and columns.
Run the third cell: how many genes are essential (1) and not (0)?
Not sure what a line does? Click Ask AI → Explain this code beside the first cell.

How many genes are in the table?

5,481 genes (and 19 columns). These are the yeast genes that have a recorded result in the systematic deletion collection – about 90% of the ~6,000 protein-coding genes.

How many are essential?

1,041 essential genes, and 4,440 non-essential.

Activity 1.2 · Ask the assistant

Ask the assistant:
Read its answer, then press Insert and run under the code.
Now type a question of your own in the box at the bottom of the assistant and press Enter – for example How do I look up one gene, for example ACT1?

What percentage of the genes are essential?

19.0% (1,041 of 5,481). Remember this number – it matters a lot when we judge machine-learning models later.

Activity 1.3 · Meet a gene

Look up actin: Then change ACT1 to another gene you know (e.g. CDC28, GAL4, RAD52) and run the cell again.
Open the Data tab and click What the columns mean. Read what each column measures – you will need this.

Which column is the answer we will try to predict (the label), and which are the clues (the features)? Pick two features and suggest why each might differ between essential and non-essential genes.

The label is essential. The features are the other measurements: protein length, GC content, codon bias, protein abundance, interaction partners, transmembrane helices, human homologue, paralogue identity, AlphaFold confidence and disorder (plus location and descriptions). For example: essential proteins may be more abundant (core machinery such as ribosomes and polymerases is made in large amounts); they may have more interaction partners (they sit in large complexes); they are more likely to be conserved in humans (essential functions are ancient); and genes with a close paralogue may be less often essential, because the other copy can take over.

Explore: how do essential genes differ?

≈ 25 minWorkbench: Notebook + AI assistant

Before building any model, look at the data. Here the assistant writes most of the code – your job is to run it, read it, and check that the results make sense.

Activity 2.1 · Compare the two groups

Ask: Insert the code and run it.

Which three features differ most between essential and non-essential genes, and in which direction?

The clearest differences: essential genes have many more interaction partners (median 68 vs 27), much higher protein abundance (36.8 vs 10.7 ppm) and are much more likely to have a human homologue (71% vs 43%). They are also less likely to have a paralogue (23% vs 38%), a little longer, with slightly higher codon bias, higher AlphaFold confidence and less predicted disorder. GC content hardly differs.

Activity 2.2 · A plot to check

Ask for a plot: Insert and run it.
Look critically at the plot. Can you actually see the difference the assistant describes? If not, tell it what is wrong (a suggestion appears under the chat).

What was wrong with the first plot, and why does a log scale fix it?

Protein abundance spans more than six orders of magnitude (≈0.01 to 23,000 ppm) and is very skewed: most proteins are rare, a few are extremely abundant. On a linear axis those few outliers set the scale, so both boxes are squashed flat at the bottom and the “clear” difference the assistant claimed cannot be seen. On a log scale equal steps are equal ratios, which spreads the data out and shows the typical essential protein is about 3½ times more abundant – with a lot of overlap. The assistant’s claim wasn’t supported by the plot it made: always check that a figure shows what the text says.

Activity 2.3 · Another plot to check

Ask: Insert and run it.
Compare every bar with the numbers printed above the plot.

Is the bar chart correct? If not, what exactly is wrong, and how did it happen?

No. The bars are sorted by height, but the labels are taken from the unsorted (alphabetical) table, so five of the eight labels are wrong: the tallest bar (34.8%, the nucleus) is labelled “Cytoplasm”, and “Nucleus” sits under a 9.1% bar (really Golgi, vesicles & vacuole). The assistant’s text (“nuclear proteins are the most likely to be essential”) even contradicts its own plot. The code ran without an error – the only way to catch it was to compare the plot with the numbers. Ask the assistant to fix it, or fix it yourself: take labels and heights from the same sorted Series (plt.bar(order.index, order.values)).

Nuclear genes are the most often essential; genes whose location is unknown the least (under 1%). Suggest a biological reason for the first and a non-biological reason for the second.

Many nuclear proteins run processes every cell needs all the time: DNA replication, transcription by RNA polymerases, mRNA splicing and export, and ribosome assembly in the nucleolus. Genes with no known location are mostly poorly studied – often because deleting them has no obvious effect, so nobody investigated them. That is a study bias: how much we know about a gene is itself linked to whether it is essential. Watch for this again with interaction partners.

Your turn Extension

Ask the assistant (or write the code yourself) for a plot of another feature – for example interaction_partners (also very skewed!) or codon_bias – for the two groups. Check the axes and labels before you believe it.

Is the difference real?

≈ 15 minWorkbench: Notebook + AI assistant

A difference in medians could be chance. A statistical test tells you how surprising the difference would be if there were really none – but it can be the wrong test, and a tiny p-value is not the same as a big effect.

Activity 3.1 · Test it

Ask: Insert and run the code.
Remember what the data looked like in the last chapter. Question the assistant’s choice of test (a suggestion appears under the chat), and run its new code.

Why is the Mann–Whitney U test more suitable than the t-test here? What does the effect size of about 0.67 mean?

The t-test compares means and assumes roughly normal data; protein abundance is extremely right-skewed, so the means (268 vs 149 ppm) are dominated by a few very abundant proteins and the test loses power (p ≈ 8 × 10⁻⁶, compared with p ≈ 7 × 10⁻⁶⁶ for the rank-based test). The Mann–Whitney U test uses ranks, so it is not thrown by skew or outliers. The effect size means: pick one essential and one non-essential gene at random, and there is a 67% chance the essential one makes more protein (50% would be no difference). The groups overlap a lot even though the p-value is minute – with 5,000 genes, even small effects are “significant”.

Activity 3.2 · Many tests at once

Ask: Insert and run it.

Which feature separates essential from non-essential genes best on its own?

codon bias interaction partners GC content protein length

You test 10 features at p < 0.05. If none of them really differed, what is the chance that at least one would look “significant” anyway? What does the Bonferroni correction do about it?

1 − 0.95¹⁰ ≈ 40%. The more tests you run, the more likely a false positive. Bonferroni multiplies each p-value by the number of tests (equivalently, uses α = 0.05 / 10), which keeps the chance of any false positive at or below 5% – at the cost of some power. Here every feature survives the correction, because the effects are real and the sample is large.

Teach a machine to spot essential genes

≈ 30 minWorkbench: Notebook + AI assistant

A classifier learns a rule that turns the features of a gene into a prediction: essential or not. To know whether it works, we must test it on genes it has never seen.

Key ideas

Features (X): the measurements for each gene. Label (y): essential or not.
Training set: genes the model learns from. Test set: genes held back to judge it. Testing on the training genes would reward memorising.
ROC AUC: the chance that a random essential gene gets a higher score than a random non-essential one – 0.5 is guessing, 1.0 is perfect.

Activity 4.1 · A first model

Ask: Insert and run the code (it takes a few seconds).
Is 95% accuracy impressive? Ask:
An AUC near 1 is suspicious for a messy biological question. Ask:

Which feature does the model rely on, and why should it not be used? (Read its description in the Data tab.) Is there a second feature that should not be there?

conditional_mutant_phenotypes – the number of phenotypes recorded for conditional (e.g. temperature-sensitive) or knock-down mutants of the gene. Geneticists make conditional alleles precisely because a gene is essential: you cannot study the deletion of an essential gene, so you make a version that only fails at high temperature. The feature is a consequence of the label – data leakage. The model is not learning biology, it is reading the answer from the history of research, and it would be useless for a gene nobody has studied (or in another organism). chromosome should go too: a chromosome number is a label, not a quantity (chromosome 12 is not “twice” chromosome 6). The assistant also filled missing abundance values with 0, which pretends that unmeasured proteins are absent – filling with the median, inside the pipeline, is safer. And its claim of “excellent” results compared accuracy with nothing: always guessing “not essential” scores 81%.

Activity 4.2 · An honest model

Ask: Insert and run it. The random forest takes up to a minute in the browser.

What cross-validated ROC AUC does the logistic regression reach?

About 0.81 (± 0.015), and the random forest about the same (≈ 0.81). Much lower than 0.98 – but this is an honest estimate for genes the model has not seen, using only information you could measure for an unstudied gene. A simple, interpretable model does as well as a complex one here.

Activity 4.3 · What does 0.81 mean in practice?

Ask:

At a threshold of 0.5 the model finds only about a quarter of the essential genes. Why? If you were choosing genes as possible targets for a new antifungal drug (where essential genes are what you want), which threshold would you use, and what would it cost you?

Essential genes are outnumbered four to one, and the features overlap a lot, so few genes reach a predicted probability above 0.5: the model finds 265 of 1,041 (recall 0.25), with 187 false alarms (precision 0.59). For drug-target discovery you would rather not miss good candidates, so a lower threshold makes sense: at 0.3 it finds 583 (recall 0.56) but with 665 false alarms (precision 0.47). Every candidate would then be checked experimentally – the model only prioritises. Changing the threshold trades false negatives for false positives; the AUC summarises the trade-off over all thresholds.

Too good to be true, part 2 Extension

Language is at the heart of modern AI. Ask: Then ask which words the text model relies on.

The text model scores an AUC above 0.9. Why should you not trust it as a predictor of essentiality?

Its top words include essential, viability and required: the descriptions were written by curators who already knew the deletion phenotypes (“Essential subunit of…”, “required for viability”, “not an essential gene”). That is leakage again, through human-written text. Descriptions also encode decades of experiments on each gene (words such as spliceosome, subunit, rRNA), which a newly discovered gene would not have. When an AI model reads text about the answer, it can look brilliant for the wrong reason.

A protein language model

≈ 20 minWorkbench: Notebook + AI assistant

Large language models learn language by predicting missing words. ESM-2 (Meta AI, 2023) did the same with proteins: it was trained to predict masked amino acids in millions of protein sequences. Along the way it learned a lot about structure, function and evolution – without ever being told about them.

Embeddings

We gave ESM-2 (the 35-million-parameter version) the sequence of every yeast protein. For each protein it produces an embedding: 480 numbers summarising what the model “thinks” of the sequence. Proteins with similar embeddings tend to be similar in structure or function. To keep the file small, the practical uses a compressed version: the first 32 principal components (esm_1 … esm_32).

Activity 5.1 · Load and map the embeddings

Ask: Check the row counts it prints.
Ask:

What do you notice in the map? Why is it remarkable that proteins group by location at all?

The proteins do not fall into neat, separate islands – a 2D map keeps only part of the information, and many proteins work in more than one place – but they are not randomly mixed either. Nuclear proteins and plasma-membrane/cell-wall proteins sit in different parts of the main cloud, and there is a separate arm made almost entirely of short proteins of unknown function (location “Unknown”, median length about 100 amino acids), of which only 1 in about 200 is essential. Essential genes are concentrated in part of the main cloud, too. This is remarkable because ESM-2 only ever saw amino-acid sequences: it was never told about compartments, functions or essentiality. But the signals that send a protein to its compartment – signal peptides, transmembrane helices, mitochondrial targeting sequences, nuclear localisation signals – and the protein families it belongs to are written in the sequence, and the model learned to recognise them simply by predicting masked amino acids.

Activity 5.2 · Neighbours in embedding space

Ask: Then try most_similar('CDC28') or a gene of your choice.

What are HXT1’s nearest neighbours? What has the language model learned?

HXT1’s ten nearest neighbours are seven other hexose transporters (HXT2–HXT5, HXT8, HXT10 and HXT17), the galactose permease GAL2, the glycerol symporter STL1 and the maltose permease MAL31: all sugar or polyol transporters of the same family (the major facilitator superfamily) in the plasma membrane. CDC28’s neighbours are mostly other protein kinases (KIN28, PHO85, FUS3, KSS1 and the PKA subunits TPK1–3), and RPL3’s are other ribosomal proteins. From sequence alone the model has learned protein families – and so, indirectly, function and location. It resembles what a BLAST search finds, but as a fixed-length vector of numbers that a machine-learning model can use. Note that neighbours do not share essentiality: CDC28 is essential, while most of its kinase neighbours are not – similar proteins can have very different roles in the cell.

Activity 5.3 · Does it help the prediction?

Ask:

Compare the three AUCs. What does the result tell you about what the language model has learned from sequence alone?

About 0.81 for the hand-made features, 0.76 for the embeddings alone and 0.83 for both together. With no measurements at all – no abundance, no interaction screens, no homology search – a model reading only the sequence gets a long way towards the features biologists chose, and adding it improves the model a little, so the embeddings capture some information the features miss (such as protein family). But the gain is modest, and the embeddings are hard to interpret: esm_7 has no meaning you can explain to a biologist. One caution: members of the same protein family can end up in both the training and the test folds, which flatters sequence-based models; a stricter test would keep each family together, which is good practice when testing machine-learning models on proteins.

AlphaFold: structure for every protein

≈ 15 minWorkbench: 3D viewer + Notebook

AlphaFold (Google DeepMind; Nobel Prize in Chemistry 2024) predicts a protein’s 3D structure from its sequence. The AlphaFold database holds models for over 200 million proteins – including almost every yeast protein. For every residue it also gives a confidence score, pLDDT (0–100).

pLDDT	Meaning	Colour in the viewer
> 90	very high confidence – side chains usually right	dark blue
70–90	confident – the backbone is usually right	light blue
50–70	low confidence	yellow
< 50	very low – often a disordered region with no fixed structure	orange

Activity 6.1 · Two essential proteins

Load the AlphaFold model of actin (ACT1): It is coloured by confidence. Rotate it (drag), zoom (scroll) and hover over residues.
Now load the nucleoporin NUP1 – also essential:

Describe the two models. Does the long orange “spaghetti” in NUP1 mean AlphaFold failed?

Actin is compact and almost entirely dark blue (average pLDDT 95): a well-folded enzyme-like domain that AlphaFold predicts very confidently. NUP1 has a small confident part and very long low-confidence tails (average pLDDT 44; 83% of residues below 50). Those regions are not an AlphaFold failure: they are intrinsically disordered. NUP1 is an FG-nucleoporin whose flexible phenylalanine–glycine repeats fill the channel of the nuclear pore and form its selective barrier – the disorder is how it works. A low pLDDT means “no single structure”, which can be biology (disorder, flexibility, folding only with a partner) or a genuinely poor prediction.

Activity 6.2 · Disorder across the genome

Back in the notebook, ask:
Ask: Load one in the 3D viewer: type its UniProt accession in the AlphaFold box at the top of the viewer (it says UniProt ID) and press Enter (e.g. P14907 for NSP1).

Are essential proteins more or less disordered, on average? Suggest a reason.

Slightly less disordered: median disorder 0.088 vs 0.124, and median pLDDT 82 vs 79. Much essential machinery consists of stable, well-folded enzymes and complex subunits (polymerases, ribosomal proteins, chaperones), while many non-essential proteins are regulators and signalling proteins with long disordered regions. But the effect is small (effect size ≈ 0.43 for disorder and 0.56 for pLDDT, where 0.5 means no difference) and there are striking exceptions – the essential FG-nucleoporins are among the most disordered proteins in the cell.

Learn from the model’s mistakes

≈ 20 minWorkbench: Notebook + AI assistant + SGD

A model’s mistakes are often more interesting than its successes: they point to biology the features cannot capture – and sometimes to errors in the data.

Activity 7.1 · Confident but wrong

Ask: Insert and run it.

What do many of these “false positives” have in common? Why might the model be misled?

Most have an enormous number of recorded interaction partners – e.g. CCR4 3,041, JHD2 1,928, GIS2 1,209, HEK2 1,149, LAA1 1,115 – and no paralogue. Hundreds or thousands of partners come mostly from high-throughput screens, which also catch “sticky” or abundant proteins and those in RNA–protein assemblies; the model has learned “many partners ⇒ essential” and over-trusts it. Others are core machinery that is important but dispensable, such as the V-ATPase subunit VMA1 (the deletion grows poorly but survives) or the clathrin heavy chain CHC1, whose deletion is lethal in some yeast strains but not in this one. A “non-essential” gene can still matter a great deal.

Activity 7.2 · Missed essential genes

Ask: Run the code and read the descriptions of UTR5 and TRE2 carefully. Look up HIP1 in SGD (Data tab → click the gene → SGD).

Why were UTR5 and TRE2 recorded as essential? What might explain why a histidine transporter (HIP1) is essential in this experiment? What does this tell you about the label we are trying to predict?

UTR5’s description says it was “originally considered essential but deletion overlapped the TATA box of neighboring essential gene HYP2”, and TRE2’s says its “inviability of null mutant in systematic studies is due to proximity to CDC31”: deleting them damaged a neighbouring essential gene. These are errors in the labels – the model was right and the data were wrong. HIP1 illustrates that essentiality depends on the strain and growth conditions: the deletion collection was made in a laboratory strain that cannot make histidine (his3), so the cell depends on taking histidine up from the medium; transporters of nutrients the strain cannot make can become essential. “Essential” is not a fixed property of a gene but of a gene in a genetic background and an environment. No model can be more accurate than its labels.

Activity 7.3 · Backup copies

Ask:
Ask for the literature: Then check the reference in PubMed ↗.

How does the percentage of essential genes change with the identity of the closest paralogue? Explain the pattern.

It falls from 22.7% for genes with no paralogue to 18.5% (up to 40% identity), 11.0% (40–60%), 4.0% (60–80%) and 5.4% (over 80%). A close duplicate can often do the same job, so losing one copy is tolerated – genetic redundancy. Yeast had a whole-genome duplication roughly 100 million years ago, and many of the surviving pairs (such as the ribosomal protein genes RPL…A/RPL…B) buffer each other. Distant paralogues have diverged in function and protect much less.

Is the assistant’s reference correct? What did you find in PubMed?

The authors, title, year, volume and pages are right, but the journal is wrong: the paper is Gu Z et al. (2003) Nature 421:63–66, not Science. AI assistants often produce references that look perfect but are partly or completely invented (“hallucinated”). Never cite a reference you have not opened yourself.

Using AI well in bioinformatics

≈ 15 minWorkbench: AI assistant

Over this practical the assistant saved you a lot of typing – and made at least six mistakes. That is roughly what working with today’s AI assistants is like.

List the assistant’s mistakes you found. For each, say how you caught it (or how you could have).

Unreadable box plot with an unjustified claim – caught by looking at the plot, knowing the data are skewed (use a log scale).
Mislabelled bar chart – caught by comparing the bars with the printed numbers.
Wrong statistical test (t-test on very skewed data) and “highly significant” without an effect size – caught by knowing the test’s assumptions.
Data leakage (conditional_mutant_phenotypes), plus chromosome as a number and missing values filled with 0 – caught because the result was too good to be true, and by checking feature importance and what each column means.
Overclaiming: 95% accuracy called “excellent” without a baseline (81%).
Leaky text model (extension) – caught by looking at which words it used.
Wrong reference (journal) – caught by checking PubMed.

None of these produced an error message. They were caught by domain knowledge, sanity checks and verification – which is why the human in the loop needs to understand the analysis.

A checklist for AI-assisted analysis

Understand every line before you rely on it – ask the assistant to explain anything unclear.
Check results against the numbers: plots vs tables, totals, row counts after merges.
Ask “compared with what?” – baselines, controls, effect sizes.
Be suspicious of results that are too good to be true: look for leakage.
Know where each column of your data comes from.
Verify facts and references in the original source.
Keep a record: save the notebook and note which parts the AI wrote.
Never paste confidential or patient data into an online AI tool.
Follow your module’s rules on using AI in assessed work – and declare it.

Activity 8.1 · Two questions for the assistant

Try Copy prompt in the assistant and read the prompt it builds. What information does it give the AI, and what does it leave out?

A model predicts essential genes with an AUC of 0.99 on held-out genes. What should you do first?

Publish it: the test set was independent Check which features it uses and whether any of them encodes the answer Train a bigger model to push it to 1.0

An AI assistant gives you a reference that supports your argument perfectly. What do you do?

Cite it – the assistant is usually right Find and read the paper before citing it Ask the assistant whether it is sure

Your classes are imbalanced (19% essential). Which single number is least informative about your model?

Accuracy ROC AUC Recall for essential genes

In a few sentences: where do you think AI will help bioinformatics most in the next five years, and what will still need a human expert?

Reference

Python cheat sheet

Task	Code
Load a table	`df = pd.read_csv('data/yeast_genes.csv')`
First rows · size · columns	`df.head()` · `df.shape` · `df.columns`
One column · several	`df['gene']` · `df[['gene', 'essential']]`
Filter rows	`df[df['essential'] == 1]` · `df[df['protein_length'] > 1000]`
Count values	`df['location'].value_counts()`
Summaries by group	`df.groupby('essential')['protein_length'].median()`
Sort	`df.sort_values('disorder', ascending=False).head(10)`
Missing values	`df.isna().sum()` · `df.dropna()`
Histogram · box plot · scatter	`plt.hist(x, bins=40)` · `plt.boxplot([a, b])` · `plt.scatter(x, y, s=5)`
Log axis · labels	`plt.yscale('log')` · `plt.xlabel('…')` · `plt.title('…')`
Rank test	`from scipy import stats; stats.mannwhitneyu(a, b)`
Cross-validated AUC	`cross_val_score(model, X, y, cv=5, scoring='roc_auc')`

Prompts that work

Be specific about the data: name the DataFrame and columns (“using df, plot protein_abundance for essential 0 and 1”).
Say what you want back: “one code cell”, “explain each step”, “label the axes with units”.
Give the error when something fails – the assistant’s Explain this error button does it for you.
Ask for checks: “how could I check this is right?”, “what could go wrong with this approach?”
Iterate: tell it precisely what is wrong with its answer.

The live assistant

In the assistant’s settings (⚙) you can connect a live model: Google Gemini, Anthropic (Claude) or any OpenAI-compatible service, with an API key. The Gemini API has a free tier: make a key with a Google account at aistudio.google.com/apikey (you must be 18 or over). On the free tier Google may use what you send to improve its products and human reviewers may read it, and only a few requests per minute and per day are allowed. Your messages, the code in your notebook and – when you ask about a cell – the start of its output or error message are sent from your browser directly to that service; the data files are not. Keys stay in this browser tab. The ▶ buttons in the instructions still give the practical’s prepared answers; questions you type go to the live model. Without a key, Copy prompt prepares a question for any AI tool you can use.

Data sources

Genes, sequences, essentiality and phenotypes: Saccharomyces Genome Database (SGD, yeastgenome.org; Engel SR et al. (2025) Genetics 229:iyae185), reference genome S288C. Essential = “inviable” null mutant in the S288C systematic deletion collection (Giaever G et al. (2002) Nature 418:387–391 and later SGD curation); 15 genes with conflicting records were removed.
Main cellular compartment: the compartment with the most Gene Ontology cellular-component annotations in SGD, preferring curated, low-throughput evidence, with terms grouped using the GO ontology (Ashburner M et al. (2000) Nat Genet 25:25–29; The Gene Ontology Consortium (2026) Nucleic Acids Res 54:D1779–D1792).
Protein abundance: PaxDb v6.1, integrated whole-organism dataset for S. cerevisiae (Huang Q et al. (2026) Nucleic Acids Res 54:D427–D439).
Physical interactions: BioGRID release 5.0.261 (Oughtred R et al. (2021) Protein Sci 30:187–200).
Homologues and paralogues: DIAMOND (Buchfink B et al. (2021) Nat Methods 18:366–368) against the UniProt human reference proteome and all yeast proteins.
Structures: AlphaFold Protein Structure Database, model version 6 (Bertoni D et al. (2026) Nucleic Acids Res 54:D358–D362; Varadi M et al. (2024) Nucleic Acids Res 52:D368–D375); AlphaFold2: Jumper J et al. (2021) Nature 596:583–589.
Protein language model: ESM-2 35M (Lin Z et al. (2023) Science 379:1123–1130), mean-pooled embeddings reduced to 32 principal components.
UniProt (The UniProt Consortium (2025) Nucleic Acids Res 53:D609–D617) for accessions and transmembrane helices.
Duplicate genes: Gu Z et al. (2003) Nature 421:63–66.

Software

Python in the browser: Pyodide (MPL-2.0) with pandas, NumPy, SciPy, scikit-learn, matplotlib and seaborn. Code editor: CodeMirror (MIT). 3D viewer: NGL Viewer (Rose AS et al. (2018) Bioinformatics 34:3755–3758; MIT).