20M
new cancer cases worldwide in 2022
With about 9.7 million deaths. IARC projects more than 35 million new cases a year by 2050.19
NEW Claude flags a previously unnoticed enzyme system in phage DNA
Biology asks why. AI asks again, genome-wide.
WhySquared works where artificial intelligence meets molecular biology and cancer research. We help labs and teams turn sequencing data, protein models and clinical evidence into decisions they can defend.
Why this matters now
The problem is bigger than any single lab.
Cancer cases are rising faster than our capacity to study them, and most drug candidates still fail late, after years of work. AI won't fix that alone, but it changes where the time and money go.
90%
of drug candidates in clinical trials fail
Lack of efficacy and toxicity explain most failures, the bulk of the cost lands at the end of the process.20
10–15
years to bring one drug to patients
At a cost of around US$1 billion per successful drug. Shortening the early, computational stages is where AI is already proving useful.20
Where AI fits
From sequence to patient.
Five stages where models are already changing molecular biology and oncology, what each is used for, and the published result behind it.
What it's used for
Under the hood
Four kinds of model, in plain English.
Most of the progress on this page comes from four families of AI. Knowing which one fits a question is half the work.
Sequence models
Trained to fill in missing letters across billions of DNA or protein sequences. To do that well, they learn which patterns are conserved, which tells you which changes are likely to matter.
e.g. Evo 2, AlphaGenome, ESM
Structure prediction
Learns how a chain of amino acids folds into a 3D shape, and how that shape fits against DNA, RNA or a drug, from structures scientists have already solved.
e.g. AlphaFold 3
Generative design
Starts from noise and refines it, step by step, into a new protein, antibody or binder that meets a target. The same idea behind image generators, applied to molecules.
e.g. Chai-2, ESM3
Research agents
A language model that plans, uses tools, reads the results and decides what to do next, working for hours across databases, code and, increasingly, lab equipment.
e.g. Claude Science
Spotlight
Anthropic moves into the lab.
In twelve months Anthropic went from connecting Claude to lab software to running its own wet lab. The most talked-about result so far: an enzyme system in bacteriophage DNA that Claude's agents flagged, with the caveat that nobody yet knows what it does.
-
Claude for Life Sciences
Connectors for Benchling, BioRender, PubMed, 10x Genomics and Synapse.org, plus a skill for single-cell RNA-seq quality control. Customers named include Sanofi, Broad Institute and Novo Nordisk.3
-
Coefficient Bio acquisition
Anthropic acquires the stealth AI-biotech startup in a deal reported at around US$400 million.6
-
Claude Science and in-house drug programmes
A research product built for reproducibility, which in a demonstration identified drug candidates for phenylketonuria. Anthropic also starts a small number of preclinical programmes for diseases drugmakers tend to neglect.4,7
-
A wet lab run with robots
A Bay Area lab where Claude controls equipment through the Model Hardware Standard, co-developed with HHMI. In an evaluation, Claude Mythos 5.1 produced 12 high-affinity binder candidates with a reported 50% hit rate, against a typical 10–15%.5
-
ART: array-associated reverse transcriptases
Agents screened public sequence data and flagged an enzyme system, mainly in phages, with a CRISPR-like repeat array. Human researchers did the lab work. Pre-print on bioRxiv.1,2
Research radar
The results we build on.
Peer-reviewed papers, trials and major announcements from the last two years. Every figure links to its source.
Genomics · NatureJan 2026AlphaGenome reads the non-coding genome
Google DeepMind's model takes up to a million base pairs and predicts 5,930 human genetic signals at once, from gene expression to splicing.
25 / 26variant-effect benchmarks matched or beaten8
Genomics · NatureMar 2026Evo 2 learns DNA across all domains of life
Arc Institute's fully open model, trained on 128,000+ genomes, can flag disease-causing mutations and design genome-length sequences.
>90%accuracy separating benign from pathogenic BRCA1 variants9
StructureMay 2024AlphaFold 3 models all of life's molecules
Structures of proteins together with DNA, RNA, ligands and ions, available free to researchers through AlphaFold Server.
≥50%better on protein interactions with other molecule types10
Nobel PrizeOct 2024Chemistry Nobel for protein AI
David Baker for computational protein design; Demis Hassabis and John Jumper for AlphaFold's structure prediction.
2M+researchers have used AlphaFold11
Protein design · ScienceJan 2025ESM3 designs a protein nature never made
EvolutionaryScale's model generated esmGFP, a glowing protein its authors compare to 500 million years of evolution.
58%sequence identity to its closest known relative12
AntibodiesJun 2025Chai-2 designs antibodies from scratch
Across 52 targets with no known antibody binders, testing up to 20 designs each, it found binders for half of them in one round.
16%hit rate in a single wet-lab round13
Screening · The LancetJan 2026MASAI: AI-supported mammography
The first randomised trial, with over 100,000 women in Sweden, found fewer cancers missed between screens.
−12%interval cancers, with 44% less reading workload14
Liquid biopsy · ASCOMay 2026NHS-Galleri: a blood test for many cancers
142,250 people in England. The primary endpoint was not met, but stage IV diagnoses fell in a secondary analysis.
4×more screen-detected cancers when added to standard screening16
Virtual cellOct 2026Building data for a virtual cell
Biohub, the US Department of Energy, NIH, Google DeepMind, Isomorphic Labs and Meta commit to open, AI-ready biological data.
$1.8Bin funding, data and compute17
Screening · NHSFeb 2025EDITH: AI across NHS breast screening
A government-funded UK trial testing AI on routine mammograms across 30 sites, backed by £11 million through NIHR.
700Kwomen in one of the largest trials of its kind15
Lab protocolsOct 2025Claude reads lab protocols like a scientist
On Protocol QA, a benchmark of questions about experimental protocols, Claude Sonnet 4.5 edged past the human baseline of 0.79.
0.83score, up from 0.74 for Sonnet 43
Drug discoveryJan 2026Isomorphic Labs heads for the clinic
The AlphaFold spin-out says its first AI-designed drugs should enter clinical trials by the end of 2026, a year later than first planned.
2026target for first human trials18
Model atlas
What goes in, what comes out.
A quick map of the systems on this page by the kind of data they work with. Filter by modality to see your options.
Model / systemDataTakes inGives backYear
What we do
Models are easy to try. Hard to trust.
WhySquared sits between the research and the lab bench: choosing the right model, wiring it into real data, and checking the output the way a diagnostics lab would.
Variant interpretation workflows
Pipelines that combine genome models such as AlphaGenome or Evo 2 with curated databases, so variant calls arrive with evidence attached rather than a bare score.
→ ranked variants, reasons, citations
Structure and binder screening
AlphaFold 3 and design models used to shortlist targets and binders before anything is ordered, with clear criteria for what goes to the bench.
→ shortlist, structures, assay plan
Single-cell and omics analysis
Quality control, clustering and annotation for single-cell RNA-seq and bulk data, delivered as reproducible notebooks your team can rerun.
→ QC report, figures, code
Literature and evidence agents
Agents that search PubMed and pre-prints, extract findings into structured tables and keep every claim traceable to a paragraph in a paper.
→ evidence tables, living reviews
Validation and quality
Model evaluation designed with ISO 15189 and GLP habits: defined acceptance criteria, held-out data, error analysis and documentation an auditor can follow.
→ validation plan, audit trail
Lab data and automation
Getting instrument output, LIMS and electronic notebooks into a shape models can use, and connecting tools like Benchling to AI assistants safely.
→ clean data layer, integrations
How we work
Small pilot first. Proof before scale.
Every engagement follows the same five steps, so you can stop after any of them with something useful in hand.
Scope
Pin down the scientific question and the decision the answer has to support.
→ one-page briefData audit
Check what data exists, its quality, consent and licensing, and what's missing.
→ data readiness reportPilot
Benchmark candidate models on your own data against a simple baseline.
→ comparison and recommendationValidate
Test against acceptance criteria agreed up front, on data the model hasn't seen.
→ validation reportHand over
Code, documentation and training so your team can run and question it.
→ working pipeline
Responsible by design
Powerful tools, careful hands.
Biology AI raises real questions about safety, privacy and over-reliance. These are the rules we work to.
Biosecurity first
We follow the field's lead on dual-use risk: Evo 2's developers, for example, left human-infecting pathogens out of its base training data. We don't take on work aimed at making pathogens more capable.9
Patient data stays protected
Work involving health data follows UK GDPR, uses the minimum data needed and runs inside your own environment wherever possible.
A qualified human signs off
No model output feeds a clinical or experimental decision without review by someone qualified to judge it.
Reproducible by default
Model versions, parameters, code and data snapshots are recorded, so any result can be rerun and checked later.
Honest limits
What the headlines leave out.
Good results come with fine print. We read it before we recommend anything, and we think you should see it too.
!
A discovery is not yet a function
ART's role in phages is still unknown, and only the first experiment, showing the array is expressed as short RNAs, has been reported.1
!
Trials can miss their main goal
NHS-Galleri did not show a significant drop in combined stage III–IV cancers. The stage IV reduction is a secondary finding awaiting peer review.16
!
Timelines slip
Isomorphic Labs expected AI-designed drugs in trials by the end of 2025. Its CEO now says the first trials should come by the end of 2026.18
!
Humans still run the lab
Even in Anthropic's most autonomous work, scientists chose the question, reviewed candidates and did every experiment by hand.1
$ background --lab --diagnostics --ai
Who's behind it
A molecular biologist who has worked the bench.
WhySquared is led by Yılmazer Yılmaz, a molecular biologist with experience in high-throughput clinical diagnostics and hands-on lab research. The view from the bench shapes how we judge AI: a prediction is only useful if a lab can act on it and an auditor can trace it.
- EducationMSc Molecular Biology, Distinction, ranked first in cohort, Coventry University
- DiagnosticsMolecular Biologist, clinical diagnostics, Siemens Healthineers, under ISO 15189 and GLP
- Lab practicePCR and BSL-2 culture work, mammalian cell lines, hospital research assistant
- DeliveryProject coordination between UK and international technical teams
Questions
Things people ask first.
Do we need a huge dataset to get started?
Often not. Many of the strongest models are pre-trained on public data, so a smaller, well-curated set of your own data is enough to test whether they help. Part of the data audit is telling you honestly if what you have is sufficient.
Can we use these models in commercial work?
It depends on the model. Licences vary widely: AlphaFold 3's code and weights, for example, were released for academic use, and AlphaFold Server is intended for non-commercial research. We check the terms for every model before recommending it.10
Will AI replace our wet-lab work?
No. It changes what reaches the bench. Even in Anthropic's ART discovery, every experiment was done by human researchers. The realistic gain is testing fewer, better candidates.1
How do you handle confidential or patient data?
We agree the data handling up front, keep processing inside your environment where possible, and use only what the question needs. Nothing is sent to an external AI service without your explicit approval.
What does a first project look like?
Usually the first two or three steps of our process: a scoping brief, a data audit and a small pilot comparing models on your data. It's designed to give you a clear go or no-go before any larger commitment.
Sources
Every number, traceable.
Primary announcements, journals and trial reports used on this page. Figures reflect what each source reported at the time of writing.
Start a project
Have data and a hard question? Ask why twice.
Tell us what you're working on, the data you have and the decision it needs to support. You'll get a short, honest assessment of whether AI can help, and how we'd check that it does.
Email usyilmazer@whysquared.com