Design of the Comprehensive Fold Recognition Benchmark. Application to SeqFold, Training and Validation

Recent exponential increase of protein sequences creates a challenge for automated annotation methods. When sequence based methods (e.g. PSIBLAST [1]) fail to identify a possible homologue (generally below 25% of protein identity i.e. within so-called twilight zone), fold recognition methods offers additional sensitivity [2,4,5,8]. However, training, validating and comparing fold recognition performance proves to be difficult. This work presents a comprehensive attempt to design a universal, two-tier, balanced fold recognition benchmark that can be used to perform all above mentioned tasks. Also, results of training and comparison of different fold recognition scoring strategies are discussed.