From SMian@lbl.gov Mon May 15 14:36:27 2000 Return-Path: Sender: saira@lbl.gov Date: Mon, 15 May 2000 14:36:07 -0700 From: Saira Mian X-Accept-Language: en To: Kevin Karplus Subject: Different libraries/T87 Content-Type: text/plain; charset=us-ascii Dear Kevin, Have you considered making sets of HMM libraries derived from the current one that differ only in terms of which columns of the alignment are designated as match states? This would change the likelihoods of the different model given the same data but I think it may allow (more automated) detection of more remote homologues. My intuition is that by making the models shorter (fewer nodes), structural homologues that are considerably shorter/longer than the training set or those at the borderline may move to being significant matches. The sets of HMMs I had in mind are the following (i) secondary structure: only residues i.e. columns that are part of a helix or strand are treated as match states (everything else is insert). (ii) physical core: only residues that the physical core of the structure are match states (this is akin to scoring against a set of "threading-like" models). (iii) structural core: only residues in helices/strands that are in the structural core of the fold are match states. In a TIM barrel, for example, this would correspond to the 8 strands and 8 helices. I had a quick look at T87: my suggestions are casp4/t87/saira/README. I didn't know whether everyone is free to add to files in the main directory so I made a work space for myself. -saira -- I. Saira Mian Life Sciences Division (Mail Stop 74-197) E-mail: SMian@lbl.gov Lawrence Berkeley National Laboratory Tel: (510) 486-6216 1 Cyclotron Road Fax: (510) 486-6949 Berkeley, California 94720 From karplus@cse.ucsc.edu Mon May 15 15:51:01 2000 Return-Path: Date: Mon, 15 May 2000 15:50:59 -0700 From: Kevin Karplus To: SMian@lbl.gov CC: karplus@cse.ucsc.edu In-reply-to: <39206DC7.C023090@lbl.gov> (message from Saira Mian on Mon, 15 May 2000 14:36:07 -0700) Subject: Re: Different libraries/T87 I have not experimented with building HMMs for reduced numbers of columns, though I have thought about it occasionally. My intuition is that there is no advantage on traget models (where structure is unknown), but there may be some advantage for template models, where one could restrict the model to core regions, and use FIMs for large inserts. The hard part is in figuring out which columns to keep and which to discard in building the model. For a small number of hand-built models, expert intuition is usable, but for the 2500 or more needed for a template library, it has to be automatic. The closest I've come is for some of the SCOP domains, which are non-contiguous, using a FIM for the inserted other domain. I'm having enough trouble figuring out how to set parameters for the whole-chain library---adding extra complexity in the form of choosing the right subset of the columns would probably overwhelm me right now.