From chemistry-request@server.ccl.net Fri Apr 5 10:17:36 2002 Received: from mr4.cc.ic.ac.uk ([155.198.5.114]) by server.ccl.net (8.11.6/8.11.0) with ESMTP id g35FHZp26588 for ; Fri, 5 Apr 2002 10:17:35 -0500 Received: from origin.ch.ic.ac.uk ([155.198.224.252] helo=chcmga.ch.ic.ac.uk) by mr4.cc.ic.ac.uk with esmtp (Exim 3.35 #1) id 16tVSz-0002va-04 for chemistry@ccl.net; Fri, 05 Apr 2002 16:17:29 +0100 Received: from hbtower.ch.ic.ac.uk ([155.198.226.132] helo=ic.ac.uk) by chcmga.ch.ic.ac.uk with esmtp (Exim 2.02 #1) id 16tVTk-00361E-00; Fri, 5 Apr 2002 16:18:16 +0100 Sender: krzys@ic.ac.uk Message-ID: <3CADB191.9BD4243@ic.ac.uk> Date: Fri, 05 Apr 2002 15:15:45 +0100 From: Krzysztof Radacki Organization: Imperial College, London X-Mailer: Mozilla 4.78 [en] (X11; U; Linux 2.4.7-10smp i686) X-Accept-Language: en MIME-Version: 1.0 To: CCL Subject: intel fortran Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Dear CClers I've tried to compile mopac 5 with intel fortran but during linking I've got following errors: finish.o(.text+0x178): undefined reference to `ieee_flags_' getdat.o: In function `getdat_': getdat.o(.text+0x13a): undefined reference to `iargc_' getdat.o(.text+0x15f): undefined reference to `getarg_' mopac.o: In function `main': mopac.o(.text+0x46): undefined reference to `ieee_handler_' mopac.o(.text+0x20d): undefined reference to `free_' mopac.o(.text+0x2a9): undefined reference to `free_' mopac.o(.text+0x39e): undefined reference to `malloc_' mopac.o(.text+0x1c98): undefined reference to `free_' mopac.o: In function `zero_': mopac.o(.text+0x1df7): undefined reference to `sleep_' readmo.o: In function `readmo_': readmo.o(.text+0xae6): undefined reference to `fdate_' second.o: In function `second_': second.o(.text+0x10): undefined reference to `etime_' writmo.o: In function `writmn_': writmo.o(.text+0x434): undefined reference to `fdate_' When I tried to use portability library (-lPEPCF90) the compilation was succesful but binary wasn't working. I'm not realy familiar with fortran so if sombody could help me ... :) regards Krzysztof -- Dr. K.Radacki Department of Chemistry Catalysis and Advanced Materials Section Imperial College, London From chemistry-request@server.ccl.net Fri Apr 5 15:09:36 2002 Received: from boris.3dp.com ([207.106.4.70]) by server.ccl.net (8.11.6/8.11.0) with ESMTP id g35K9Zp31822 for ; Fri, 5 Apr 2002 15:09:36 -0500 Received: from 3dp.com (dhcp-1-46.3dp.com [172.26.1.46]) by boris.3dp.com (SGI-8.9.3/8.9.3) with ESMTP id PAA65640 for ; Fri, 5 Apr 2002 15:12:46 -0500 (EST) Message-ID: <3CAE053A.3403E216@3dp.com> Date: Fri, 05 Apr 2002 15:12:42 -0500 From: Ed Jaeger X-Mailer: Mozilla 4.78 [en] (Windows NT 5.0; U) X-Accept-Language: en MIME-Version: 1.0 To: chemistry@ccl.net Subject: MACC Meeting on April 17th in Princeton Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit I ask CCL's indulgence for the broadcast nature of this meeting notice of local interest. The Mid-Atlantic Computational Chemistry discussion group will be holding its next meeting on April 17th from 6-9 P.M. Professor Curt Breneman of Renssalear Polytechnic Institute will be speaking on "New Developments in Molecular Property Descriptors, Machine Learning and Computational ADME" The schedule is: Social Hour 6-7 P.M. Seminar 7-8 The venue is Kresge Auditorium (Room 120 Frick) at Princeton University. For R.S.V.P. and information please email to jaeger@3dp.com -- Ed Jaeger 3-Dimensional Pharmaceuticals, Inc. ph:610-458-6052 jaeger@3dp.com 665 Stockton Drive, Exton, PA 19341 fx:610-458-8249 From chemistry-request@server.ccl.net Fri Apr 5 08:52:55 2002 Received: from mail.fqspl.com.pl ([212.244.147.20]) by server.ccl.net (8.11.6/8.11.0) with SMTP id g35Dqgp24403 for ; Fri, 5 Apr 2002 08:52:54 -0500 Received: (qmail 32228 invoked from network); 5 Apr 2002 13:52:44 -0000 Received: from test.fqspl.com.pl (HELO test) (10.10.10.50) by mail.fqspl.com.pl with SMTP; 5 Apr 2002 13:52:44 -0000 Message-ID: <00e901c1dca9$3f922290$320a0a0a@test> From: "Victor Anisimov" To: Subject: MESP visualization Date: Fri, 5 Apr 2002 15:53:21 +0200 MIME-Version: 1.0 Content-Type: text/plain; charset="koi8-r" Content-Transfer-Encoding: 7bit X-Priority: 3 X-MSMail-Priority: Normal X-Mailer: Microsoft Outlook Express 5.50.4133.2400 X-MimeOLE: Produced By Microsoft MimeOLE V5.50.4133.2400 Dear CCLers, I'd like to make nice 3D visualization of a molecular electrostatic potential calculated for 120 000 atoms protein system on Connolly surface. My MESP file is a textual file with three columns corresponding to Cartesian coordinates of a point and fourth column containing MESP value, i.e. each line corresponds to a separate point on the Connolly surface. Number of points on one Connolly surface is about one million. No cube format is possible I guess by size of the molecule and number of the mesh points one would need to store in the cube file. Can someone suggest me a free software that can read this file and draw nice 3D map with colours corresponding to MESP value? My hardware limitation is 1GB RAM and Windows/Linux single Athlon CPU machine. I'd greatly appreciate any hints. Thank you in advance, Victor. From chemistry-request@server.ccl.net Fri Apr 5 06:43:32 2002 Received: from biochim.ro ([193.231.158.224]) by server.ccl.net (8.11.6/8.11.0) with ESMTP id g35Bfhp21545 for ; Fri, 5 Apr 2002 06:41:47 -0500 Received: from localhost (amilac@localhost) by biochim.ro (8.9.3/8.9.3) with ESMTP id OAA15763 for ; Fri, 5 Apr 2002 14:52:48 -0200 Date: Fri, 5 Apr 2002 14:52:48 -0200 (GMT+2) From: Adina Milac To: Computational Chemistry List Subject: Re: CCL:neural networks problem, details In-Reply-To: <002801c1dbe1$9f75b320$6602050a@mahindrabt.com> Message-ID: MIME-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII Thank you all for answering me so promptly. Several of the messages I have received ask me for details about the network I have used, therefore I will give additional information about what I have done. To be more clear, perhaps I should mention that the software I use for neural networks is Matlab, version 4.2c.1 (if you can recommend me other neural network software, especially running under UNIX, I would be most grateful). I have normed my input data in order to be able to use various transfer functions (I have always been aware of the fact that using sigmoid transfer functions the network will never be able to predict an output equal to 8). I have normed both in [-1,1] interval and in [0,1] interval. Results were slightly better in the latter case. I haven't tried normalization in [-0.9;0.9], I have to try this, thanks for the suggestion. With the toolbox included in Matlab, I have first tried to use a simple perceptron, but, as expected, it didn't do much (it couldn't even learn the data properly). After that I developed a 2-layered feed-forward network. The network is able to learn but the prediction has a HUGE error if I use 7 input neurons and still a very high error (>200% for some vectors) if I use 40 neurons in the input layer. In this case the problem might be with the respective vectors because there are vectors for which the error is below 4%. The next step was to add layers of neurons. I have tried many combinations of numbers of neurons in the input and hidden layer and of transfer functions and finally I have found an "optimal" architecture of 10 (input), 15 (hidden), 1 (output). The training algorithm I have used is an improved version of backpropagation (Levenberg-Marquardt algorithm) because simple backpropagation did not give satisfying results. Practically, the code I have used to design, train and test the network is the following (I will add some explanatory notes in case you use another software; if you use Matlab also and identify mistakes in the commands I gave, please correct me): P - matrix of 40 training vectors (each vector has 4 fields) T - (line) matrix of target vectors [w1,b1,w2,b2,w3,b3] = initff(P,10,'logsig',15,'logsig',T,'logsig') initializes a network with 10 neurons in the input layer and 15 neurons in the hidden layer; output layer has 1 neuron tp = [50 1000 0.01] 1000 training epochs; display error at each 50 epochs; error goal is 0.01 [w1,b1,w2,b2,w3,b3] = trainlm(w1,b1,'logsig',w2,b2,'logsig',w3,b3,'logsig',P,T,tp) the network is trained (the longest training time was about 500 epochs, it never reached 1000) y = simuff(P,w1,b1,'logsig',w2,b2,'logsig',w3,b3,'logsig') simulates the system. Instead of P I have also used the matrix of vectors in the testing set - unseen by the network (vectors used for prediction). As transfer functions I have also tried tansig and purelin, but not with signifficant differences in network performance. ************ What is even more intriguing for me is that it seems to be important the order of the fields in the training vectors. For example, if instead of v1 = [f1 f2 f3 f4] I train with v1` = [f3 f4 f1 f2] I get completely different results. Should it be this way? *********** Thank you all and good luck, ADINA MILAC Ph.D Student, Research Assistant Institute of Biochemistry, Romanian Academy Splaiul Independentei 296, 77700 Bucharest 17, Romania Phone: (+401).223.90.69; FAX: (+401).223.90.68 e-mail: amilac@biochim.ro