Servers go dark.
A survey of 2,396 bioinformatics web tools found that half of those published in 2010 were dead by 2020. ToxClassifier, a well-known toxin classifier, was reported unavailable in 2024.
offlineVenom protein classifier. Early access.
Is this protein a toxin?
Paste a sequence. Get toxic or atoxic. From a tool built to stay online.
ToxinClass is a convolutional neural network trained on animal venom proteins. It reads the chemistry of every amino acid and makes the call. We are opening early access to researchers first.
93.1% accuracy on the TOXIFY and ToxClassifier benchmark setThe quiet problem
Toxin classifiers keep disappearing.
You find the paper. You click the link. The server is gone.
Installs break.
Some tools still depend on Python 2. If you are a biologist and not a developer, that is a wall.
offlineAnswers arrive without reasons.
Most tools return a label and nothing else. Hard to defend in a methods section.
offlineWe want to be the one that is still here.
Sources: Lifetime of bioinformatics web services, Nucleic Acids Research 2020. Web of venom, GigaScience 2024.
How it works
Sequence in. Verdict out.
Three steps.
-
01
Paste a protein sequence.
Plain amino acid text or a FASTA record.
-
02
We translate it into chemistry.
Each amino acid becomes five numbers, the Atchley factors, which capture properties such as polarity, size and charge. Your sequence becomes a signal the network can read.
-
03
The network makes the call.
A convolutional neural network, the model family behind image recognition, scans that signal and returns toxic or atoxic.
Try the idea
Use standard amino acid letters, at least 6.
WAITING
Illustration with sample data, not a real prediction.
Fun fact: we borrowed the network design from LeNet-5, a model built to read handwritten digits. Sequences, it turns out, have shapes too.
Trained on 5,896 toxic and 5,896 atoxic UniProtKB sequences, up to 500 amino acids long.
Results
93.1%. And here is where that sits.
Measured on the benchmark set used by TOXIFY and ToxClassifier, so the comparison is fair.
- ToxClassifier 99.7%
- TOXIFY 96.0%
- ToxinClass 93.1%
Bars start at 0, with gridlines at 25, 50 and 75. ToxClassifier and TOXIFY figures are as reported in their own studies. ToxinClass figure from our research project report.
We are not the highest number on this chart, and we will not pretend to be. ToxinClass takes a different route, a convolutional network on Atchley-encoded sequences, and builds a more usable tool around it. A higher score is the next job. A tool that is online and documented is the one you can use.
The 20 amino acids, shaded by polarity (Atchley factor 1). Five such numbers per residue feed the network.
The research
Read the dissertation behind it.
Machine Learning Approaches for Classifying Animal Venom Proteins
69 pages on the data, the Atchley encoding, the network and the results. Free to read, free to cite.
Who it is for
For people who work with venom.
Drug discovery teams.
Venom has produced real medicines, such as ziconotide for chronic pain, from cone snail venom. Triage candidate sequences before the expensive lab work begins.
Venom researchers.
Annotate proteins from a newly sequenced species, or compare how venom changes across populations.
Students and educators.
A readable example of machine learning on biological sequences.
Roadmap
What we plan to build next.
Based on what researchers say is missing from existing tools.
-
Planned
Reasons, not just labels.
Show which parts of a sequence drove the call.
-
Planned
Beyond toxic or atoxic.
Researchers want to know what kind of toxin it is, not only whether it is one.
-
Planned
Batch and API.
Screen many sequences and feed results into a pipeline.
Early access
Free for academics. Paid plans coming.
Early access is open. Founding users help shape what we build.
Academic
Free
For researchers and students.
- Toxic or atoxic verdict from a sequence
- Early access, limits to be confirmed
- .edu or .ac.uk email to verify
Professional
Coming soon
For research labs and small biotech teams.
- Batch upload (Planned)
- API access (Planned)
- Result export (Planned)
Enterprise
Talk to us
For pharma and contract research teams.
- Volume screening (Planned)
- Priority support (Planned)
- Tell us what you need
Academic access requires a verified .edu or .ac.uk email.
Be among the first to try it.
Tell us who you are and which plan fits. No credit card.
Questions? Write to [email protected]