Why toxin prediction tools disappear, and what to look for in one

What is the best tool to predict whether a protein is a toxin?

There is no single answer, and anyone who gives you one is selling something. The honest version is that the best tool is the one that is still online when you need it, reports a benchmark you can check, and can be run again in three years when a reviewer asks you to repeat the analysis. Raw accuracy matters, but accuracy you cannot reach is worth nothing.

Of the classifiers that get cited most in venom protein work, two of the historical names cannot realistically be used today. ClanTox, published in Nucleic Acids Research in 2009, served predictions from a university web server that has since stopped answering. ToxClassifier, published in PeerJ Computer Science in 2016, was reported as unavailable as of April 2024 in a GigaScience review of venom data resources. [2] Both are still cited as baselines. Neither is a working option.

How often do bioinformatics web tools go offline?

Often enough that it should change how you choose. The 2020 study "On the lifetime of bioinformatics web services" tracked 2,396 tools and reported 25.7% unreachable against 74.3% working at the initial snapshot. [1]

The age gradient is the part that matters for anyone reading an older methods section. Tools published in 2019 and 2020 showed roughly 90% availability. For tools published in 2010 that figure dropped to about 50%. [1] Put plainly, a decade is enough to lose half the field.

Intermittent failure is just as common as permanent failure. Across 133 days of repeated checks, 31% of the tools were consistently accessible, 48.4% were occasionally reachable and 20.6% were never accessible. [1] A tool that answers on Monday and times out on Thursday is a real problem for a pipeline run you need to finish.

The study also found that tools accessible during testing received approximately five times higher mean citation counts than permanently unavailable services from the same publication period. [1] Whatever else that correlation reflects, staying online appears to be part of how a method gets used.

What happened to ToxClassifier and ClanTox?

ToxClassifier combined support vector machines, generalised linear models, BLAST searches against curated venom positives and HMMER profile matching. Methodologically it was a reasonable hybrid, and its profile hits gave a small amount of biological traceability. Its public code repository was last updated in 2016. A licence request opened on that repository in January 2022 has sat unanswered for over four years. [5]

The 2023 paper describing CSM-Toxin was blunt about the practical consequence. It reported that it could not obtain predictions from some prior methods because of inaccessible servers, and it described ToxClassifier as depending on outdated and no longer supported tools such as Python 2. [3] That single dependency is enough to stop most biologists on a modern machine, and Python 2 reached end of life in 2020.

ClanTox is older. It was a support vector machine meta-classifier aimed at short, cysteine-rich animal toxins, and for its time it was a sophisticated design. It was delivered as a web form on a university host over plain HTTP. Our own check of that address in April 2026 found the connection refused. Recent papers continue to cite ClanTox as a historical baseline while quietly not benchmarking against a live server.

The pattern is not specific to toxin classifiers. The same GigaScience review noted that Trinotate was no longer under active development or support as of March 2024, that ArachnoServer is at times unavailable, that SCORPION2 has become obsolete, and that VenomKB, described as a comprehensive database useful in translational research, was discontinued. [2]

Why do these tools disappear?

Almost never because the science was wrong. The causes are mundane and structural.

• The project was a paper, not a product. A classifier built for a publication has a funded lifetime of a grant or a degree. Nothing in the incentive structure pays for year six.
• The host was personal or departmental. A machine under a specific lab's control follows that lab's staffing. When the maintainer moves on, the hostname outlives the person responsible for it.
• The dependencies rot. Python 2, an exact R version, TensorFlow 1.x. A pinned environment that worked in 2016 is a build failure in 2026, even when the code is public and correct.
• Nobody can be reached. A preprint on the decay of scientific email addresses found that 49% become invalid within ten years, and roughly 18% of author contact addresses in MEDLINE are already invalid. [6]

There is a hopeful footnote. When the lifetime study contacted the authors of 47 recently published non-functional tools, 57.4% replied and 51.1% of the tools were restored to working order. [1] Decay is often neglect rather than refusal. It is just that neglect compounds, and the odds get worse the longer you wait.

Why does a dead tool matter for your paper?

Three reasons, in increasing order of annoyance.

First, reproducibility. If your methods section says a sequence was screened with a classifier whose server no longer answers, nobody can repeat that step, including you. Later publications that build on the original analysis inherit the problem.

Second, comparison. Benchmarks drift when competitors cannot be run. The authors of CSM-Toxin reported that the predictive performance of several prior methods deteriorated significantly from their published results, which they read as a sign of overfitting and poor generalisability. [3] You cannot audit a claim against a tool you cannot execute.

Third, silent partial failure, which is the worst of the three. In the CSM-Toxin evaluation, TOXIFY was unable to make predictions for the whole blind set and processed 1,796 of 2,540 sequences. [3] A tool that returns fewer rows than you submitted without raising an alarm is harder to catch than a tool that is plainly offline.

What should you look for in a toxin prediction tool?

Six checks, in the order we would run them. Each one takes a few minutes and tells you something a results table does not.

One extra check that costs nothing: submit a sequence you already know the answer for, count the rows that come back, and confirm the count matches what you sent. That one habit would have caught the partial blind set run described above.

Which toxin classifiers can you still run today?

A short and deliberately non-exhaustive map, as of late 2026. ToxinPred3, from the Raghava group, is hosted and actively maintained, with a peptide focus. CSM-Toxin, from the Biosig lab, is hosted and is the rare example with a documented REST API. VISH-Pred and ToxDL 2.0 are both recent and available, the latter bringing predicted structure into the picture. TOXIFY is installable as a command line tool and is the closest methodological relative of ToxinClass, but its TensorFlow 1.x dependency makes installation harder every year, and it has no hosted interface at all.

For historical context, TOXIFY's own 2019 benchmark reported accuracy of 0.68 for ClanTox, 0.77 for ToxClassifier and 0.86 for TOXIFY, with ToxClassifier taking around 100 seconds and 6.8 GB of memory against TOXIFY's 4 seconds and 293 MB. [4] The lighter tool was also the more accurate one, which is a reminder that heaviness is not rigour.

Where does ToxinClass fit?

ToxinClass started as a University of Surrey final year project: a convolutional neural network that classifies animal venom proteins as toxic or atoxic from primary sequence, using Atchley factor encoding of each amino acid. It was trained on 5,896 toxic and 5,896 atoxic UniProtKB sequences of up to 500 amino acids, and it reached 93.1% accuracy on the same benchmark set used by TOXIFY and ToxClassifier. [7]

That is below TOXIFY's reported 96.0% and below ToxClassifier's reported 99.7% on their own benchmarks, and we say so on the landing page and in the paper. The method is one contribution. The reason the project continued past the degree is the other half of this article: the classifiers people cite most in this niche are the ones they can no longer run.

So the commitments are narrow and checkable rather than ambitious. Hosted access with no install. The dissertation published in full, free to read and cite. Honest benchmark framing that names the tools we lose to. Versioned models, so a result from today can be reproduced later. Explanations of a verdict, batch screening and an API are planned, not shipped, and are labelled that way wherever they appear.

None of that is a guarantee of permanence. Nobody who ran a server in 2010 planned to stop. It is a set of design choices that make the failure mode recoverable instead of final, which is the most any tool in this field can honestly offer.

Frequently asked questions

Is ToxClassifier still available?

A 2024 review in GigaScience reported ToxClassifier as unavailable as of April 2024. [2] Its code repository was last updated in 2016, and the CSM-Toxin paper noted its dependency on Python 2, which is no longer supported. [3] The method remains citable, but most researchers can no longer run it.

How many bioinformatics web tools go offline?

In a 2020 study of 2,396 bioinformatics web services, 25.7% were unreachable at the initial snapshot. Availability was around 90% for tools published in 2019 and 2020 and about 50% for tools published in 2010. Over 133 days, 31% were consistently accessible, 48.4% occasionally reachable and 20.6% never accessible. [1]

Does a tool going offline affect how often it is cited?

The same study found that tools accessible during testing received approximately five times higher mean citation counts than permanently unavailable services from the same publication period. [1]

What should I look for in a toxin prediction tool?

Hosted access with no install, an offline fallback such as open code or downloadable weights, a documented API or command line interface, a benchmark on a named test set with the comparison tools named, a versioned model so results can be reproduced later, and visible maintenance activity.

Which toxin classifiers can you still run today?

ToxinPred3 and CSM-Toxin are hosted and actively maintained, VISH-Pred and ToxDL 2.0 are recent and available, and TOXIFY is installable as a command line tool although its TensorFlow 1.x dependency makes installation harder each year. ClanTox and ToxClassifier are cited often but are no longer practically runnable.

Can I just email the author of a dead tool?

Sometimes. The lifetime study contacted authors of 47 recently published non-functional tools, got a 57.4% response rate and saw 51.1% of those tools restored. [1] The odds fall with age, since 49% of scientific author email addresses become invalid within ten years. [6]

Read the method, not just the claim

The ToxinClass dissertation covers the data, the Atchley encoding, the network design and the results in full, including where it loses to other tools.

Read the dissertation (PDF, 2.7 MB) or read the companion explainer, how venom proteins are classified with CNNs and Atchley factors. Questions about choosing a tool? Email [email protected].

References

1. Kern F., Fehlmann T., Keller A. "On the lifetime of bioinformatics web services." Nucleic Acids Research, 2020. pmc.ncbi.nlm.nih.gov/articles/PMC7736811
2. "Web of venom: exploration of big data resources in animal toxin research." GigaScience, 2024. academic.oup.com/gigascience/article/doi/10.1093/gigascience/giae054
3. "CSM-Toxin: a web server for predicting protein toxicity." 2023. pmc.ncbi.nlm.nih.gov/articles/PMC9966851
4. Cole T. J., Brewer M. S. "TOXIFY: a deep learning approach to classify animal venom proteins." PeerJ, 2019. peerj.com/articles/7200
5. Gacesa R., Barlow D., Long P. F. "ToxClassifier." PeerJ Computer Science, 2016, and project repository. peerj.com/articles/cs-90, github.com/rgacesa/ToxClassifier
6. "Revisiting the decay of scientific email addresses." bioRxiv. biorxiv.org/content/10.1101/633255v1
7. Van Sebroeck C. "Machine Learning Approaches for Classifying Animal Venom Proteins." University of Surrey, Department of Computing, 2020. Read the PDF

Figures attributed to other tools are as reported in their own publications. ToxinClass figures come from the dissertation above.