Secondary structure prediction software

From DrugPedia: A Wikipedia for Drug discovery

---

Current revision

The question came up what software exists (preferably free and open) that does secondary structure prediction using neural nets, and more generally, what's the state-of-the-art at all with the problem.

[edit] General information

• Wikipedia:
  • Protein structure prediction
  • Protein structure prediction software
  • especially the table:

Name	Method	Description	Link
NetSurfP	Profile-based neural network	Webserver	server
GOR	Information theory/Bayesian inference	Many implementations	Basic GOR GOR V
Jpred	Neural network assignment	Webserver	server
Meta-PP	Consensus prediction of other servers	Webserver	main page
PREDATOR	Knowledge-based database comparison	Webserver	server
PredictProtein	Profile-based neural network	Webserver	server
PSIPRED	two feed-forward neural networks which perform an analysis on output obtained from PSI-BLAST	Webserver	server
YASSPP	Cascaded SVM-based predictor using PSI-BLAST profiles	Webserver	server

The server software in detail:

• NetSurfP: Ref.: Petersen et al.: A generic method for assignment of reliability scores applied to solvent accessibility predictions [1] From the website and abstract, they use simple feed-forward nets. For the paper, they trained on a commonly used set of sequences known as the CB513 set, but it's not clear which set they use now. The source is available at request.
• JPred: Ref.: Cole, Barber, Barton: The Jpred 3 secondary structure prediction server (paper free) "accuracy 81.5%", predicts from multiple alignment profiles; combined multiple neural networks, which had been trained on the same multiple sequence alignments, but where the alignments were presented to the networks in different ways. Uses HMMs, too. Uses SNNS as library. Uses semi-automatic pipeline to create training data from SCOP and UniProt databases. Not available.
• PSIPRED Ref.: McGuffin, Bryson, Jones: The PSIPRED protein structure prediction server (paper free) feed-forward nets on the output of PsiBLAST, see paper; very efficient due to PsiBLAST frontend, around 80% accuracy; available for non-commercial use at http://globin.bio.warwick.ac.uk/psipred/
• YASSPP: Ref.: Karypis G. "YASSPP: Better kernels and coding schemes lead to improvements in SVM-based secondary structure prediction" [2] Uses support-vector machines on PsiBLAST output. Choice of 4 training sets. Around 80% accuracy. Not available.

Comment: The nearly identical accuracies show that the method is nearly irrelevant. You won't get more except with ab-initio methods.