src/openbabel/fingerprint.cpp

/**********************************************************************
fingerprint.cpp - Implementation of fingerpring base class and fastsearching

Copyright (C) 2005 by Chris Morley
 
This file is part of the Open Babel project.
For more information, see <http://openbabel.sourceforge.net/>

This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation version 2 of the License.
 
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
GNU General Public License for more details.
***********************************************************************/

#include "config.h"
#include <vector>
#include <algorithm>

#if HAVE_IOSTREAM
#include <iostream>
#elif HAVE_IOSTREAM_H
#include <iostream.h>
#endif

#if HAVE_FSTREAM
#include <fstream>
#elif HAVE_FSTREAM_H
#include <fstream.h>
#endif
#include <vector>

#include "fingerprint.hpp"
#include "oberror.hpp"

using namespace std;
namespace OpenBabel {

OBFingerprint* OBFingerprint::_pDefault; //static variable
const unsigned int OBFingerprint::bitsperint = 8 * sizeof(unsigned int);

void OBFingerprint::SetBit(vector<unsigned int>& vec, unsigned int n)
{
        vec[n/bitsperint] |= (1 << (n % bitsperint));
}

////////////////////////////////////////        
void OBFingerprint::Fold(vector<unsigned int>& vec, unsigned int nbits)
{
        while(vec.size()*bitsperint/2 >= nbits) 
                vec.erase(transform(vec.begin(),vec.begin()+vec.size()/2,
                        vec.begin()+vec.size()/2, vec.begin(), bit_or()), vec.end());
}

////////////////////////////////////////
bool OBFingerprint::GetNextFPrt(std::string& id, OBFingerprint*& pFPrt)
{
        Fptpos iter;
        if(id.empty())
                iter=FPtsMap().begin();
        else
        {
                iter=FPtsMap().find(id);
                if(iter!=FPtsMap().end())
                        ++iter;
        }
        if(iter==FPtsMap().end())
                return false;
        id    = iter->first;
        pFPrt = iter->second;
        return true;
}

OBFingerprint* OBFingerprint::FindFingerprint(string& ID)
{
        if(ID.empty())
                return _pDefault;
        Fptpos iter = FPtsMap().find(ID);
        if(iter==FPtsMap().end())
                return NULL;
        else
                return iter->second;
}

double OBFingerprint::Tanimoto(const vector<unsigned int>& vec1, const vector<unsigned int>& vec2) 
{
        //Independent of sizeof(unsigned int)
        if(vec1.size()!=vec2.size())
                return -1; //different number of bits
        int andbits=0, orbits=0;
        for (unsigned i=0;i<vec1.size();++i)
        {
                int andfp = vec1[i] & vec2[i];
                int orfp = vec1[i] | vec2[i];
                //Count bits
                for(;andfp;andfp=andfp<<1)
                        if(andfp<0) ++andbits;
                for(;orfp;orfp=orfp<<1)
                        if(orfp<0) ++orbits;
        }
    return((double)andbits/(double)orbits);
}

//*****************************************************************8
/*!
          
*/
bool FastSearch::Find(OBBase* pOb, vector<unsigned int>& SeekPositions,
                      unsigned int MaxCandidates)
{
        ///Finds chemical objects in datafilename (which must previously have been indexed)
        ///that have all the same bits set in their fingerprints as in the fingerprint of
        ///a pattern object. (Usually a substructure search in molecules.)
        ///The type of fingerprint and its degree of folding does not have to be specified
        ///here because the values in the index file are used.
        ///The positions of the candidate matching molecules in the original datafile are returned.
        
        vector<unsigned int> vecwords;
        _pFP->GetFingerprint(pOb,vecwords, _index.header.words * OBFingerprint::bitsperint);
        
        vector<unsigned int>candidates; //indices of matches from fingerprint screen
        candidates.reserve(MaxCandidates);

        unsigned int dataSize = _index.header.nEntries;
//      GetFingerprint(mol, vecwords, _index.header.words, _index.header.fptype); 
        
        unsigned int words = _index.header.words;
        unsigned int* nextp = &_index.fptdata[0];
        unsigned int* ppat0 = &vecwords[0];
        register unsigned int* p;
        register unsigned int* ppat;
        register unsigned int a;
        unsigned int i; // need address of this, can't be register
        for(i=0;i<dataSize; i++) //speed critical section
        {
                p=nextp;
                nextp += words;
                ppat=ppat0;
                a=0;
                while(p<nextp)
                {
                        if ( (a=((*ppat) & (*p++)) ^ (*ppat++)) ) break;
                }
                if(!a)
                {
                        candidates.push_back(i);
                        if(candidates.size()>=MaxCandidates)
                                break;
                }
        }

        if(i<_index.header.nEntries) //premature end to search
          {
#ifdef HAVE_SSTREAM
            stringstream errorMsg;
#else
            strstream errorMsg;
#endif
            errorMsg << "Stopped looking after " << i << " molecules." << endl;
            obErrorLog.ThrowError(__func__, errorMsg.str(), obInfo);
          }

        vector<unsigned int>::iterator itr;
        for(itr=candidates.begin();itr!=candidates.end();itr++)
        {
                SeekPositions.push_back(_index.seekdata[*itr]);
        }
        return true;
}

/////////////////////////////////////////////////////////
bool FastSearch::FindSimilar(OBBase* pOb, multimap<double, unsigned int>& SeekposMap,
                                                                                                                 double MinTani)
{
        vector<unsigned int> targetfp;
        _pFP->GetFingerprint(pOb,targetfp, _index.header.words * OBFingerprint::bitsperint);

        unsigned int words = _index.header.words;
        unsigned int dataSize = _index.header.nEntries;
        unsigned int* nextp = &_index.fptdata[0];
        register unsigned int* p;
        register unsigned int i;
        for(i=0;i<dataSize; i++) //speed critical section
        {
                p=nextp;
                nextp += words;
                double tani = OBFingerprint::Tanimoto(targetfp,p);
                if(tani>MinTani)
                        SeekposMap.insert(pair<const double, unsigned int>(tani,_index.seekdata[i]));
        }       
        return true;
}

/////////////////////////////////////////////////////////
bool FastSearch::FindSimilar(OBBase* pOb, multimap<double, unsigned int>& SeekposMap,
                                                                                                                 int nCandidates)
{
        ///If nCandidates is zero or omitted the original size of the multimap is used
        if(nCandidates)
        {
                //initialise the multimap with nCandidate zero entries
                SeekposMap.clear();
                int i;
                for(i=0;i<nCandidates;++i)
                        SeekposMap.insert(pair<const double, unsigned int>(0,0));
        }
        else if(SeekposMap.size()==0)
                return false;

        vector<unsigned int> targetfp;
        _pFP->GetFingerprint(pOb,targetfp, _index.header.words * OBFingerprint::bitsperint);

        unsigned int words = _index.header.words;
        unsigned int dataSize = _index.header.nEntries;
        unsigned int* nextp = &_index.fptdata[0];
        register unsigned int* p;
        register unsigned int i;
        for(i=0;i<dataSize; i++) //speed critical section
        {
                p=nextp;
                nextp += words;
                double tani = OBFingerprint::Tanimoto(targetfp,p);
                if(tani>SeekposMap.begin()->first)
                {       
                        SeekposMap.insert(pair<const double, unsigned int>(tani,_index.seekdata[i]));
                        SeekposMap.erase(SeekposMap.begin());
                }
        }       
        return true;
}/////////////////////////////////////////////////////////
string FastSearch::ReadIndex(istream* pIndexstream)
{
        //Reads fs index from istream into member variables
        // but first checks whether it is already loaded
        FptIndexHeader headercopy = _index.header; 
        pIndexstream->read((char*)&(_index.header), sizeof(FptIndexHeader));

        if(memcmp(&headercopy,&(_index.header),sizeof(FptIndexHeader)))
        {
                pIndexstream->seekg(_index.header.headerlength);//allows header length to be changed

                unsigned int nwords = _index.header.nEntries * _index.header.words;
                _index.fptdata.resize(nwords);
                _index.seekdata.resize(_index.header.nEntries);

                pIndexstream->read((char*)&(_index.fptdata[0]), sizeof(unsigned int) * nwords);
                pIndexstream->read((char*)&(_index.seekdata[0]), sizeof(unsigned int) * _index.header.nEntries);
                
                if(pIndexstream->fail())
                        *(_index.header.datafilename) = '\0';
                
                string tempFP(_index.header.fpid);
                _pFP = OBFingerprint::FindFingerprint(tempFP);
                if(!_pFP)
                {
#ifdef HAVE_SSTREAM
                  stringstream errorMsg;
#else
                  strstream errorMsg;
#endif
                  errorMsg << "Index has Fingerprints of type '" << _index.header.fpid 
                           << " which is not currently loaded." << endl;
                  obErrorLog.ThrowError(__func__, errorMsg.str(), obWarning);
                  *(_index.header.datafilename) = '\0'; 
                }

        }
        return _index.header.datafilename;
}

//*******************************************************
FastSearchIndexer::FastSearchIndexer(string& datafilename, ostream* os, 
                                     std::string& fpid, int FptBits)
{
        ///Starts indexing process
        _indexstream = os;
        _pFP = OBFingerprint::FindFingerprint(fpid);
        if(!_pFP)
          {
#ifdef HAVE_SSTREAM
            stringstream errorMsg;
#else
            strstream errorMsg;
#endif
            errorMsg << "Fingerprint type '" << fpid << "' not available" << endl;
            obErrorLog.ThrowError(__func__, errorMsg.str(), obWarning);
          }

        _nbits=FptBits;
        _pindex= new FptIndex;
        _pindex->header.headerlength = sizeof(FptIndexHeader);
        strncpy(_pindex->header.fpid,fpid.c_str(),15);
        strncpy(_pindex->header.datafilename, datafilename.c_str(), 255);
}

/////////////////////////////////////////////////////////////
FastSearchIndexer::~FastSearchIndexer()
{
        ///Saves index file
        _pindex->header.nEntries = _pindex->seekdata.size();
        _indexstream->write((const char*)&_pindex->header, sizeof(FptIndexHeader));
        _indexstream->write((const char*)&_pindex->fptdata[0], _pindex->fptdata.size()*sizeof(unsigned int));
        _indexstream->write((const char*)&_pindex->seekdata[0], _pindex->seekdata.size()*sizeof(unsigned int));
        if(!_indexstream)
          obErrorLog.ThrowError(__func__,
                                "Difficulty writing index", obWarning);
        delete _pindex;
}

///////////////////////////////////////////////////////////////
bool FastSearchIndexer::Add(OBBase* pOb, streampos seekpos)
{
        ///Adds a fingerprint
        
        vector<unsigned int> vecwords;
        if(!_pFP)
                return false;
        if(_pFP->GetFingerprint(pOb, vecwords, _nbits))
        {
                _pindex->header.words = vecwords.size(); //Use size as returned from fingerprint
                for(unsigned int i=0;i<_pindex->header.words;i++)
                        _pindex->fptdata.push_back(vecwords[i]);
                _pindex->seekdata.push_back(seekpos);
                return true;    
        }
        obErrorLog.ThrowError(__func__, "Failed to make a fingerprint", obWarning);
        return false;

}

/*!
\class OBFingerprint
These fingerprints are condensed representation of molecules (or other objects)
as a list of boolean values (actually bits in a vector<unsigned>) with length 
of a power of 2. The main motivation is for fast searching of data sources
containing large numbers of molecules (up to several million). OpenBabel
provides some routines which can search text files containing lists of molecules
in any format. See the documentation on the class FastSearch.   

There are descriptions of molecular fingerprints at <br>
http://www.daylight.com/dayhtml/doc/theory/theory.finger.html) and <br>
http://www.mesaac.com/Fingerprint.htm <br>
Many methods of preparing fingerprints have been described, but the type supported
currently in OpenBabel has each bit representing a substructure (or other
molecular property). If a substructure is present in the molecule, then a 
particular bit is set to 1. But because the hashing method may also map other 
substructures to the same bit, a match does not guarantee that a particular 
substructure is present; there may be false positives.  However, with proper design,
a large fraction of irrelevant molecules in a data set can be eliminated in a 
fast search with boolean methods on the fingerprints. 
It then becomes feasible to make a definitive substructure search by
conventional methods on this reduced list even if it is slow. 

OpenBabel provides a framework for applying new types of fingerprints without
changing any existing code. They are derived from OBFingerprint and the
source file is just compiled with the rest of OpenBabel. Alternatively,
they can be separately compiled as a DLL or shared library and discovered
when OpenBabel runs.
 
Fingerprints derived from this abstract base class OBFingerprint can be for any 
object derived from OBBase (not just for OBMol). 
Each derived class provides an ID as a string and OBFingerprint keeps a map of 
these to provides a pointer to the class when requested in FindFingerprint.
   
<h4>-- To define a fingerprint type --</h4>
The classes derived form OBFingerprint are required to provide
a GetFingerprint() routine and a Description() routine
\code
class MyFpType : OBFingerprint 
{
        MyFpType(const char* id) : OBFingerprint(id){};
        virtual bool GetFingerprint(OBBase* pOb, vector<unsigned int>& fp, int nbits) 
        {
                //Convert pOb to the required type, usually OBMol
                OBMol* pmol = dynamic_cast<OBMol*>(pOb);
                fp.resize(required_number_of_words);
                ... 
                use SetBit(fp,n); to set the nth bit

                if(nbits)
                        Fold(fp, nbits);
        }
        virtual const char* Description(){ return "Some descriptive text";}
        ...
};
\endcode
//Declare a global instance with the ID you will use in -f options to specify its use.
\code
MyFpType theMyFpType("myfpID");
\endcode

<h4>-- To obtain a fingerprint --</h4>
\code
OBMol mol;
...
        vector<unsigned int> fp;
        OBFingerprint::GetDefault()->GetFingerprint(&mol, fp); //gets default size of fingerprint
\endcode        
or
\code
        vector<unsigned int> fp;
        OBFingerPrint* pFP = OBFingerprint::FindFingerprint("myfpID");
        ...and maybe...
        pFP->GetFingerprint(&mol,fp, 128); //fold down to 128bits if was originally larger
\endcode

<h4>-- To print a list of available fingerprint types --</h4>
\code
        std::string id;
        OBFingerPrint* pFPrt=NULL;
        while(OBFingerprint::GetNextFPrt(id, pFPrt))
        {
          cout << id << " -- " << pFPrt->Description() << endl;
        }
\endcode


Fingerprints are handled as vector<unsigned int> so that the number of bits
in this vector and their order will be platform and compiler
dependent, because of size of int and endian differences. Use fingerprints
(and fastsearch indexes containing them) only for comparing with other 
fingerprints prepared on the same machine.

The FingerprintFormat class is an output format which displays fingerprints
as hexadecimal. When multiple molecules are supplied it will calculate the
Tanimoto coefficient from the first molecule to each of the others. It also
shows whether the first molecule is a possible substructure to all the others,
i.e. whether all the bits set in the fingerprint for the first molecule are
set in the fingerprint of the others. To display hexadecimal information when
multiple molecules are provided it is necessay to use the -xh option.

To see a list of available format types, type babel -F on the command line.
The -xF option of the FingerprintFormat class also provides this output, but due
to a quirk in the way the program works, it is necessary to have a valid input
molecule for this option to work.
*/
        
/*! \class FastSearch
The FastSearch class searches an index to a datafile containing a list of molecules
(or other objects) prepared by FastSearchIndexer.

OpenBabel can also search files for molecules containing a substructure specified
by a SMARTS string, using OBSmartsPattern or from the command line:
\code
babel datafile.xxx -outfile.yyy -sSMARTS
\endcode
But when the data file contains more than about 10,000 molecules this becomes
rather too slow to be used interactively. To do it more quickly, an index
of the molecules containing their fingerprints (see OBFingerprint) is prepared using 
FastSearchIndexer. The indexing process may take a long time but only has to 
be done once. The index can be searched very quickly with FastSearch. Because 
of the nature of fingerprints a match to a bit does not guarantee 
the presence of a particular substructure or other molecular property, so that
a definitive answer may require a subsequent search of the (much reduced) list
of candidates by another method (like OBSmartsPattern).

Note that the index files are not portable. They need to be prepared on the
computer that will acess them.

<h4>Using FastSearch and FastSearchIndexer in a program</h4>

The index has two tables:
- an array of fingerprints of the molecules
- an array of the seek positions in the datasource file of all the molecules

<h4>To prepare an fastsearch index file:</h4>
- Open an ostream to the index file.
- Make a FastSearchIndexer object on the heap or the stack, passing in as parameters:
  - the datafile name, the indexfile stream,
  - the id of the fingerprint type to be used,
  -  the number of bits the fingerprint is to be folded down to, If it is to be left
    unfolded, set fpid to 0 or do not specify it.
  .
- For each molecule, call Add() with its pointer and its position in the datafile.<br>
    Currently the std::streampos value is implicitly cast to unsigned int so that
    for 32bit machines the datafile has to be no longer than about 2Gbyte.
- The index file is written when the FastSearchIndexer object is deleted or goes out
  of scope.

<h4>To search in a fastsearch index file</h4>
 
- Open a std::istream to the indexfile (in binary mode on some systems)
- Make a FastSearch object, read the index and open the datafile whose 
name it provides
\code   
  ifstream ifs(indexname,ios::binary);
  FastSearch fs;
  string datafilename = fs.ReadIndex(&ifs);
  if(datafilename.empty()/endcode
    return false;
  ifstream datastream(datafilename);
  if(!datastream)
    return false;
\endcode

<strong>To do a search for molecules which have all the substructure bits the
OBMol object, patternMol</strong>
\code
  vector<unsigned int>& SeekPositions;
  if(!fs.Find(patternMol, SeekPositions, MaxCandidates))
        
  for(itr=SeekPositions.begin();itr!=SeekPositions.end();itr++)
  {
    datastream.seekg(*itr);
    ... read the candidate molecule
        and subject to more rigorous test if necessary
  }
\endcode

<strong>To do a similarity search based on the Tanimoto coefficient</strong>
This is defined as: <br>
 <em>Number of bits set in (patternFP & targetFP)/Number of bits in (patternFP | targetFP)</em><br>
where the boolean operations between the fingerprints are bitwise<br>
The Tanimoto coefficient has no absolute meaning and depends on
the design of the fingerprint.
\code
  multimap<double, unsigned int> SeekposMap;
  // to find n best molecules
  fs.FindSimilar(&patternMol, SeekposMap, n);
  ...or
  // to find molecules with Tanimoto coefficient > MinTani
  fs.FindSimilar(&patternMol, SeekposMap, MinTani);

  multimap<double, unsigned int>::reverse_iterator itr;
  for(itr=SeekposMap.rbegin();itr!=SeekposMap.rend();++itr)
  {
    datastream.seekg(itr->second);
    ... read the candidate molecule
    double tani = itr->first;
  }
\endcode

The FastSearchFormat class facilitates the use of these routine from the
command line or other front end program. For instance:

<strong>Prepare an index:</strong>
\code
  babel datafile.xxx index.fs
\endcode
or
\code
  babel datafile.xxx -ofs namedindex
\endcode
With options you can specify:
- which type of fingerprint to be used, e.g. -xfFP2, 
-       whether it is folded to a specified number of bits, e.g. -xn128 
        (which should be a power of2)
- whether to pre-select the molecules which are indexed:
        - by structure e.g only ethers and esters, -sCOC
        - by excluding molecules with bezene rings, -vc1ccccc1
        - by position in the datafile e.g. only mols 10 to 90, -f10 -l90
  .
<strong>Substructure search</strong> in a previously prepared index file
\code
  babel index.fs outfile.yyy -sSMILES
\endcode
The datafile can also be used as the input file, provided the input format is specified as fs 
\code
  babel datafile.xxx outfile.yyy -sSMILES -ifs
\endcode
A "slow" search not using fingerprints would be done on the same data by omitting -ifs.
A "slow" search can use SMARTS strings, but the fast search is restricted
to the subset which is valid SMILES.

With the -S option, the target can be specified as a molecule in a file of any format 
\code
  babel datafile.xxx outfile.yyy -Spattern_mol.zzz -ifs
\endcode
These searches have two stages: a fingerprint search which produces
a number of candidate molecules and a definitive search which selects
from these using SMARTS pattern matching. The maximum number of candidates
is 4000 by default but you can change this with an option
e.g. -al 8000  (Note that you need the space between l and the number.)
If the fingerprint search reaches the maximum number it will not
look further and will tell you at what stage it stopped.

<strong>Similarity search</strong> in a previously prepared index file<br>
This rather done (rather than a substructure search) if the -at option is used,
\code
  babel datafile.xxx outfile.yyy -sSMILES -ifs -at0.7
\endcode
for instance
- -at0.7 will recover all molecules with Tanimoto greater than 0.7 
- -at15 (no decimal point) will recover the 15 molecules with largest coefficients.
- -aa will add the Tanimoto coefficient to the titles of the output molecules.

All stages, the indexing, the interpretation of the SMILES string in the -s option,
the file in the -S option and the final SMARTS filter convert to OBMol and apply
ConvertDativeBonds(). This ensures thatforms such as[N+]([O-])=O  and N(=O)=O 
are recognized as chemically identical.     
*/

}//Openbabel

//! \file fingerprint.cpp
//! \brief Definitions for OBFingerprint base class and fastsearch classes
Revision:	819
Committed:	Fri Dec 16 21:52:50 2005 UTC (19 years, 10 months ago) by tim
File size:	20953 byte(s)
Log Message:	changed __FUNCTION__ to __func__ to match C99 standard, and then added an autoconf test to check for __func__ usability. Changed some default compile flags for the Sun architecture