Neural networks for calculating Contact-dependent Secondary Structure Propensity
Applications in predicting non-native secondary structures and amyloid fibril formation

    What is NetCSSP?

    NetCSSP is a neural network-based predictor of contact-dependent  secondary structure propensity in protein sequences. The local secondary structure of a protein can change depending on the tertiary environment such as inter-molecular interactions and misfolding. One prominent example is non-native beta- strand formation in most amyloid fibrils.

    We have quantitatively standardized the tertiary interaction energies of various proteins and then trained neural networks (NetCSSP) to predict secondary structures of a protein from the standardized tertiary interaction energies together with a sequence context (7-residue span) . Thus the power of the Net-CSSP method is that it provides insight on the change of secondary structural propensities in the entire range of tertiary interaction energies for a given amino acid sequence without any knowledge of its 3D structure.

     One important application of Net-CSSP is its accurate predictive ability in identifying core sequences of amyloid fibril formation where the switch in local secondary structure occurs.


    Dual artificial neural network for CSSP prediction

    The dual network architecture consists of the two single networks that use distinct output nodes.

    One network has output nodes for a-helix and non-helix, thus predicting helical propensity. The other network has output nodes for b-strand and non-beta strand, thus predicting beta propensity.

    The input layer of single and dual networks contains an additional variable node for the query sequence. This variable node was set to the computed energy term (>(i,i±4) energy) of the query sequence for single network architecture.  However, for dual network architecture, the a-helix- predicting network uses the (i,i±4) energy term of the query sequence as the additional input, while the beta-predicting network uses >(i,i±4) energy term as the additional input.

    During operation, the additional input typically will span the range of scaled energy values from -2.0 to 2.0. Because individual amino acids differ with respect to side-chain length, composition, and hydrophobicity, the interaction energy of an input sequence is pre-normalized by the average interaction energy of the corresponding amino acids (See details in our publications).


    Accuray of CSSP2 for predicting secondary structure at their native tertiary contexts

    Traning set: 440,884 fragments of 7-residue length from SCOP20
    Test  set :    22,707 fragments extracted from 1,629 unique fold SCOP domains.

    Single neural network : 74%
    Dual neural network : 83%

    (Find details in Yoon et al., Comp Biol & Chem, 2007)


    Accuray for predicting aggregation-prone amino acid fragments

    NetCSSP calculates P(helix), P(beta) and P(coil) for each residue of a given sequence
    The amyloidogenic hidden beta propensity (HbP) is calculated using the form,
    In this validation using ROC plots, the calculated "HbP" successfully priotizes aggregration-prone fragments over non-aggregates.

                                             Test set 1 includes 104  fragments
                                             Test set 2 includes   70 fragments
    AUC (Area Under the Curve) represents the predictive power of the method
    Test data sets were retrieved from Fernandez-Excamilla et al. (2004) Nat. Biotechnol., 22, 1302-1306


    Pinpointing the core of amyloid fibril formation

    Top figure:  The CSSP profile of human IAPP4-34 sequence
    Bottom figure: Summary of aggregation data of 28 consecutive overlapping decamers from IAPP (Raw data retrieved from Mazor et al. J. Mol Biol, 2002)

    Energy-dependent fluctuation of beta-propensity (blue) in the entire CSSP profile is highly consistent with experimental data showing the core subsequence for amyloid fibril formation.

    For 28 decamers, the calculated beta propensity is well correlated with the experimental relative aggragation ability, i.e., relative binding to amyloids.

    (Find details in Yoon  & Welsh, Protein Science, 2004 and Proteins, 2005)


    Correlation between predicted "Hidden beta Propensity (HbP)" and experimental
    aggregation rate of AcP mutants

                            Experimental data sets were retrieved from Chiti et al., Nat Struc Biol, 2002 

                               The P
    mut and Pw represent mutant HbP and wild-type HbP respectively, 
                                                     and they were calucalated using the  form,

    (Yoon and Welsh, 2005, Proteins)


    Comparison of CSSP-based beta propensity between amyloidogenic and non-amyloidogenic sub-sequences

          Data from Yoon & Welsh, 2004, Protein Science


    CSSP-based "hidden beta-propensity" in SCOP90 domains and known amyloidogenic proteins

        Data from Yoon & Welsh, 2005, Poteins
       A total of 6,390 domains were retrieved from SCOP90v1.61