GLARE algorithm using RDKit with python3

Now version of RDKit has many tools. And I interested in the Glare algorithm.

This algorithm is used for good quality library generation from large set of reagents.
In the method, key point is pre calculation of reagent properties and sum the value for product. So, It does not need calculate product property on the fly.
Details are described following article

I tried to use the algorithm in my PC, but it did not work. Because of my environment is python 3.5.
There are some changes form python2 to python3. And the code has specific function for python2.
So, I added some changes to original code.

First, defined cmp function because there is not used cmp function in python3.
Second, changed sort key argument from cmp to cmp_to_key( cmp ).
3rd devision was changed. In the python2 9/8 return 1, but 9/8 return 1.125 in python3. So, I changed following code.
#sizes = [ (libIdx, max(rg.count()/total_num_partitions_per_rgroup, 1))
# for libIdx, rg in enumerate(self.rgroups) ]
sizes = [ (libIdx, max(rg.count()//total_num_partitions_per_rgroup, 1))
for libIdx, rg in enumerate(self.rgroups) ]

Finally modified code is following.

from __future__ import print_function
import random, operator, itertools, math
# in the python3, need to call reduce from function
from functools import reduce, cmp_to_key
# in python3, cmp function is not used.
# xrange is equal to range.
def cmp( a,b ):
    if a == b: return 0
    return -1 if a < b else 1

Glare Algorithm

Some nomenclature:

 A Libary is made of RGroups
 RGroups are a collection of sidechains (the paper uses Fragments)
  that can populate the rgroup position.

 We desire to optimize the Library so that we have a good chance
  of making the desired products.

 Example From the testing code, using Fake data:

    r1 = RGroups(makeFakeSidechains("aldehydes", num=1000))
    r2 = RGroups(makeFakeSidechains("boronic_acids", num=1500))

    lib = Library([r1,r2])
    props = [
        Property("mw",    propIdx=0, minValue=0, maxValue=500),
        Property("alogp", propIdx=1, minValue=-2.4, maxValue=5),
        Property("tpsa",  propIdx=2, minValue=0, maxValue=90)

    glare = Glare()
    # optimize the library...
    glare.optimize(lib, props)

class Property:
    def __init__(self, name, propIdx, minValue, maxValue, scaffoldoffset=0.0):
        """name, propIdx, minValue, maxValue, scaffoldoffset -> initial a Property
        name is the name of the property.
         propIdx:  the index of the property in the property vector
         minValue: the minimum acceptable value for the property
         maxValue: the maximum acceptable value for the property
         scaffoldoffset: any offset from the reaction scaffold (defaults to 0)
        """ = name
        self.propIdx = propIdx
        self.minValue = minValue
        self.maxValue = maxValue
        self.offset = scaffoldoffset

    def evaluate(self, sidechains):
        """sidechains -> Evaluate a list of sidechains to see if they
        pass the property values.

        Each sidechain must have a property vector e.g. (s.props for s in sidechains)
        which is a vector of values where s.props[propIdx] is the property
        being inspected
        product = self.offset
        propIdx = self.propIdx
        for s in sidechains:
            product += s.props[propIdx]
        return self.minValue <= product <= self.maxValue

class Sidechain:
    """Holds the name (identifier) and property list for the
    given sidechain/fragment.  Properties are assumed to
    be numerical values"""
    def __init__(self, name, props, goodCount=0):
        """name, props, goodCount=0 -> initialize a Sidechain
        initialize a sidechain.
        name: the unique name for the sidechain
        props: the property vector (see Properties class for details)
        goodCount: the number of times this reagent belongs to
            a good product, where good is a product that is in the desired
            property space.
        """ = name
        self.props = props
        self.good_count = goodCount # shared variable
        self.dropped = False        # shared variable

    def __str__(self):
        return "Sidechain %s(%s, goodCount=%s)"%(,
                                                 self.props, self.good_count)
    def __repr__(self):
        return "Sidechain(%r, %r, %s)"%(, self.props, self.good_count)

class RGroups:
    """Holds a collection of sidechains for the given RGroup"""
    def __init__(self, sidechains):
        """Sidechains -> RGroups
         sidechains: the list of Sidechains that make up the potential
                     sidechains at this rgroup position"""
        self.sidechains = sidechains

        self.rejected = []  # list of rejected sidechains
        self.initial_size = len(sidechains)

    def count(self):
        """Returns the number of possible sidechains"""
        return len(self.sidechains)

    def randomize(self):
        """Randomly shuffles the sidechains and reset the goodness counts"""
        for s in self.sidechains:
            s.good_count = 0

    def effectiveness(self):
        """-> return the current effectiveness of this collection
        effectiveness is the number of items left divided by the
        initial amount"""

        return len(self.sidechains)/float(self.initial_size)

    def chunk_size(self, num_chunks):
        """num_chunks -> return the number of sidechains in each chunk
        if the sidechains are split into num_chunks chunks"""
        return int(math.ceil(float(len(self.sidechains))/num_chunks))

    def chunk(self, chunk_idx, num_chunks):
        """chunk_idx, num)chunks -> RGroups
        return the chunk_idxth chunk given num_chunks total chunks"""
        assert chunk_idx >=0  and chunk_idx < num_chunks, "%s %s"%(
            chunk_idx, num_chunks)

        n = self.chunk_size(num_chunks)
        return RGroups(self.sidechains[chunk_idx*n:(chunk_idx+1)*n])

    def prune(self, fractionToKeep):
        """fractionToKeep -> Sort the sidechains from the most often
        found if good products to the least, and keep the best
        fractionToKeep percentage"""
        assert 0 < fractionToKeep <= 1.0, "fractionToKeep: %s"%fractionToKeep

        self.sidechains.sort(key=cmp_to_key(lambda x,y: -cmp(x.good_count, y.good_count)))
        fragment_index = int(len(self.sidechains) * fractionToKeep + 0.5)

        # update rejected set
        self.rejected += self.sidechains[fragment_index:]
        self.sidechains = self.sidechains[:fragment_index]

class Library:
    """A library is a collection of RGroups that need to be combinitorially
    def __init__(self, rgroups):
        """rgroups -> Initialize the Library.
        rgroups: the list of possible RGroups that is combinitorially
                 combined to make the library"""
        self.rgroups = rgroups

    def isValid(self):
        """If we have an empty set for any rgroup, return False"""
        for rg in self.rgroups:
            if len(rg.sidechains) == 0:
                return False
        return True

    def randomize(self):
        """randomize the order of the sidechains"""
        for rg in self.rgroups:

    def getSidechainsPerPartition( self, total_num_partitions_per_rgroup ):
        """total_num_partitions -> [num_fragments/partition for rgroup1,
                                    num_fragments/partition for rgroup2]
        return the number of sidechains in a partition
        for each rgroup"""
        #sizes = [ (libIdx, max(rg.count()/total_num_partitions_per_rgroup, 1))
        #          for libIdx, rg in enumerate(self.rgroups) ]

        sizes = [ (libIdx, max(rg.count()//total_num_partitions_per_rgroup, 1))
                  for libIdx, rg in enumerate(self.rgroups) ]

        # "optimially" apportion the partitions according the
        #  the glare paper see Appendix eq (8) and (9)
        # sort by size
        sizes.sort(key=cmp_to_key(lambda x,y: cmp(x[1], y[1])))
        last_size = 1
        opt_sizes = []
        for libIdx, current_size in sizes[:-1]:
            opt_sizes.append( (libIdx,
                               current_size - (current_size % last_size)) )
            last_size = current_size

        # From the Glare paper:
        # the last library size is set equal to the second to last
        # From Table 3, it is easy to understand that, if the fourth dimension
        # was split in 24 instead of 12, a factor of 2 would be gained from the
        # reduced size of the sublibraries. However, twice as many sublibraries
        # would be needed, and the net speedup would be null, hence, the decision to
        # set p4=p3. (p4 here is the last library)
        libIdx, current_size = sizes[-1]
        opt_sizes.append((libIdx, last_size))
        # back to the original library order
        res = [size for libIdx, size in opt_sizes]
        return res

    def chunk(self, num_partitions):
        """num_partitions -> [Library(..), Library(...)]

        Return new libraries that are chunks of this one.
        These are the libraries that get sampled to see of
        sidechains participate in good products.
        partitions = self.getSidechainsPerPartition(num_partitions)
        max_subsets = max(partitions)

        enumeration_indices = []
        # change xragen to range because there is no xragen function in python3.
        for i in range(max_subsets):
            combinations = []
            for size in partitions:
                combinations.append( i % size )
            enumeration_indices.append( combinations )

        library_sets = []
        for subset_index, combinations in enumerate(enumeration_indices):
            libs = []
            partitioned_rgroups = []
            for lib_index, libpart_index in enumerate(combinations):
                lib = self.rgroups[lib_index]
                num_chunks = partitions[lib_index]
                partitioned_rgroups.append( lib.chunk(chunk_idx=libpart_index,
            lib = Library(partitioned_rgroups)
            if lib.isValid():

        return library_sets

    def effectiveness(self):
        """-> returns the average effectiveness of this library set"""
        sum = 0.0
        for rg in self.rgroups:
            sum += rg.effectiveness()
        return sum/len(self.rgroups)

    def evaluate(self, props):
        """props -> num_good_enumerations, total_enumerations

        props: a list of Property evaluators for the fragments.

        returns the number of good enumerations and the total number of
        enumerations for this Library"""
        frags = [rg.sidechains for rg in self.rgroups]
        good = 0
        bad = 0
        for i,frag in enumerate(itertools.product(*frags)):
            for p in props:
                if not p.evaluate(frag):
                    bad += 1
                good += 1
                for sidechain in frag:
                    sidechain.good_count += 1
        return good, i+1

class Glare:
    """Glare Algorithm.  Implementation of

    GLARE: A New Approach for Filtering Large Reagent Lists in
           Combinatorial Library Design Using Product Properties
    Jean-Francois Truchon* and Christopher I. Bayly

       # somehow make sidechains1/2 with props [mw, alogp, tpsa]
       r1 = RGroups(sidechains1)
       r2 = RGroups(sidechains2)
       lib = Library([r1, r2])
       props = [
         Property("mw", 0, 0, 500),
         Property("alogp", 1, -2.4, 5),
         Property("tpsa", 2, 0, 90)

      glare = Glare()
      glare.optimize(lib, props)
    def __init__(self,
                 rgroupScale=6.0, # None if no scaling
                 initialFraction=None,#None=auto -100.,
        self.fractionGood = self.desiredFinalGoodness = desiredFinalGoodness
        self.maxIterations = maxIterations
        self.rgroupScale = rgroupScale

        if initialFraction is not None:
            self.initialFraction = initialFraction/100.
            self.initialFraction = initialFraction
        self.numPartitions = numPartitions

    def optimize(self, library, props):
        """library, props
        Given a Library and the list of Propery evaluators,
        optimize the library.
        The library is modified in place by removing building blocks
        (sidechains) that are not likely to pass the property
        # attempt to generate report like glare application
        print ("------- PARAMETERS: --------------")
        print ("GOOODNESS THRESHOLD : %s%%"%(self.desiredFinalGoodness * 100))
        print ("MIN PARTITION SIZE : %s"%self.numPartitions)
        if self.initialFraction is None or self.initialFraction > 0.999:
            print ("INITAL FRACTION TO KEEP : %s%%"%(self.initialFraction*100))
        print ("Actual SIZE : %s = %s"%(
            " x ".join([str(len(rg.sidechains)) for rg in library.rgroups]),
            reduce(operator.mul, [len(rg.sidechains) for rg in library.rgroups])

        running_total = 0.0
        Gt = self.desiredFinalGoodness

        for iteration in range(1, self.maxIterations+1):
            # chunk of the total library into smaller more managable sets
            #  and run combinitorial analysis on the sub libraries
            #  each of these records the number of times a sidechain is used
            #  in a successful enumeration which is then used to prune the
            #  library at the end
            for rg in library.rgroups:

            good = total = 0.0
            chunked_libs =  library.chunk(self.numPartitions)
            # for each chunk, do the combinitorial check to see
            #  if reagents make good products
            for libidx, chunk in enumerate(chunked_libs):
                g,t = chunk.evaluate(props)
                good += g
                total += t
            running_total += total
            Gi = good/total # current goodness

            if Gi < 1e-12:
                # I think we're done here :)
                fraction = 0.0
            elif iteration == 1:
                G0 = Gi # Goodness at first iteration

                # the first time, use the initalFraction or a "good enough"
                #  value
                if self.initialFraction is not None:
                    fraction = K0 = self.initialFraction
                    # auto choose the fraction based on the current good percentage
                    #  and the desired
                    fraction = K0 = min(-1.1 * ( Gt - G0) + 1.2,
                # the second time, gradually eliminate reagents slowing
                #  down as the number of iterations increases
                #  see equation (5) in reference
                if abs(Gt-G0) < 1e-4:
                    Ki = 1.0
                    Ki = (1.0 - K0) * (Gi - G0) / (Gt - G0) + K0;
                fraction = min(1.0, Ki)

            # prune the library to keep the highest occuring sidechains
            #  note that even if all sidechains are acceptable,
            #  some will always get pruned

            max_size = float(max([len(rg.sidechains) for rg in library.rgroups]))
            for rg in library.rgroups:
                scale = 1.0
                if self.rgroupScale is not None:
                    # scale differently size rgroups via equation (6)  in paper
                    numSidechains = len(rg.sidechains)
                    numer = 1.0
                    denom = 1.0 + math.exp(-self.rgroupScale *
                                           ((numSidechains/max_size) - 0.5))
                    scale = numer/denom
                fraction_to_reject = (1.0 - fraction) * scale
                # keep the best fraction...
                rg.prune(1.0 - fraction_to_reject)

            print ("-------------- ITERATION : %s ----------------------"%iteration)
            print ("GOODNESS      : %s%%"%(Gi * 100))
            print ("NUMBER EVAL   : %s"%(total))
            print ("CUMUL EVAL    : %s"%(running_total))
            print ("KEPT IN STEP  : %s%%"%(fraction*100.))
            if not iteration:
                print ("GOODNESS THRESHOLD : %s"%self.desiredFinalGoodness)
                print ("MIN PARTITION SIZE : %s"%self.numPartitions)
                print ("INITIAL FRACTION TO KEEP : ")
                if self.fractionToKeep > 0.999:
                    print ("AUTOMATIC")
                    print ("%s%%"%self.fractionToKeep)

            print ("Actual SIZE : %s = %s"%(
                " x ".join([str(len(rg.sidechains)) for rg in library.rgroups]),
                reduce(operator.mul, [len(rg.sidechains) for rg in library.rgroups])
            print ("EFFECTIVENESS : %s%%"%(library.effectiveness()*100.))

            # stopping critieria
            if iteration and Gi < 1e-12:
            elif abs(Gi - self.desiredFinalGoodness) < 0.001 or \
                 Gi > self.desiredFinalGoodness:

# testing codes
def makeFakeProps():
    mw = random.randint(10,500)
    alogp = random.randint(-10,10)
    tpsa = random.randint(0,180)
    return [mw, alogp, tpsa]

def makeFakeSidechains(lib, num):
    res = []
    for i in range(num):
        res.append(Sidechain(lib + "_" + str(i), makeFakeProps()))
    return res

def testGlare():
    a = RGroups(makeFakeSidechains("aldehydes", 1000))
    b = RGroups(makeFakeSidechains("boronic_acids", 1500))

    lib = Library([a,b])
    props = [
        Property("mw", 0, 0, 500, 230.1419),
        Property("alogp", 1, -2.4, 5, 2.212749),
        Property("tpsa", 2, 0, 90, 24.5)

    glare = Glare()
    glare.optimize(lib, props)

if __name__ == "__main__":

OK Let’s test.

# call library and run the test code.

import glare4py3 as glare

------- PARAMETERS: --------------
Actual SIZE : 1000 x 1500 = 1500000
-------------- ITERATION : 1 ----------------------
GOODNESS      : 0.292%
NUMBER EVAL   : 25000.0
CUMUL EVAL    : 25000.0
KEPT IN STEP  : 15.821200000000001%
Actual SIZE : 385 x 297 = 114345
-------------- ITERATION : 2 ----------------------
GOODNESS      : 0.9463233012721067%
NUMBER EVAL   : 6446.0
CUMUL EVAL    : 31446.0
KEPT IN STEP  : 16.40277864502602%
Actual SIZE : 78 x 89 = 6942
EFFECTIVENESS : 6.866666666666667%
-------------- ITERATION : 3 ----------------------
GOODNESS      : 7.183908045977011%
NUMBER EVAL   : 1740.0
CUMUL EVAL    : 33186.0
KEPT IN STEP  : 21.94689739642575%
Actual SIZE : 23 x 23 = 529
EFFECTIVENESS : 1.9166666666666665%
-------------- ITERATION : 4 ----------------------
GOODNESS      : 48.39319470699433%
NUMBER EVAL   : 529.0

Works fine!
I pushed the code to my repo.

I will try to calculate my reagent set. ;-)


Published by iwatobipen

I'm medicinal chemist in mid size of pharmaceutical company. I love chemoinfo, cording, organic synthesis, my family.

One thought on “GLARE algorithm using RDKit with python3

  1. I believe this site has got some really fantastic information for everyone :D. “The public will believe anything, so long as it is not founded on truth.” by Edith Sitwell.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

This site uses Akismet to reduce spam. Learn how your comment data is processed.

%d bloggers like this: