This document discusses using genetic algorithms and other machine learning techniques to explore the vast chemical space of possible molecules. It notes that while there are over 10^60 possible small molecules, only around 10^8 have been made so far, leaving most of chemical space unexplored. It provides examples of how genetic algorithms with additive scoring functions have been able to rediscover specific target molecules and molecules with desired properties like light absorption or protein docking. With continued improvements in scoring functions, these techniques may eventually be able to efficiently search the entire chemical space to discover new molecules with useful applications.

1. Chemical Space Exploration
Jan H. Jensen
University of Copenhagen
The game of Go has 10170 possible positions, yet
computers can now beat grandmasters.
Can we use similar approaches for chemistry?
Chemical Space
1060 possible molecules
(1023 stars in the observable universe)
108 molecules made so far
Almost all of chemical
space is unexplored but
how do we search such
a large space?

2. GAN 2014,
VAE 2013,
RL 2013
VAE
applied to
molecules
Oct. 2016

3. The Fundamental Challenge
1060 106 100 1
AI?
Recurrent NNs
autocomplete
for molecules
Autoencoders
molecules as
vectors
Genetic Algorithms
Evolving new
molecules

4. “to be or not to be that is the question”
27 characters and 39 positions
2739 = 6.7 x 1055 possible sentences
Yet a genetic algorithm can consistently find the correct
sentence by considering only 50,000 sentences
How?
A Simple Example from Shakespeare
1055 104 1

5. to be or not to be that is the question
ll hczcoanysflshfkeoomatsinswqm ld jpzn
pssogzosqrnapy ywuwqakdvrs snibjoqmziwx
ll hczcoanysflshf + keoomatsinswqm ld jpzn
pssogzosqrnapy yw + uwqakdvrs snibjoqmziwx
pssogzosqrnapy ywkeoomatsinswqm ld jpzn
pssogzosqrnapy ywketomatsinswqm ld jpzn
score = 1
score = 1
score = 2
Genetic Algorithm
score = 3
Generate 100 random sequences
Score sequences
Pick pair of sequences based on score
Mate/crossover
Mutate
Score
Mate
Mutate

6. 1-(26/27)39 or 77% of the
6.7 x 1055 possible sequences have
at least one character placed correctly
77% of sequences have score ≥ 1
Sequence Space
path
Maria H.
Rasmussen

7. Need Additive and Semi-Continuous Scores

8. Is it possible to find one specific molecule among 1060?
Rediscovery
Score = Tanimoto Similarity
OH
H2N OH
Tanimoto = 0.33

9. Is it possible to find one specific molecule among 1060?
Rediscovery
Score = Tanimoto Similarity
OH
H2N OH
Tanimoto =
3 in common
9 total

10. O
HN
O
S
O
O
OH
Can we find Troglitazone?
(55 unique fragments)

11. CC1=C(O)C(C)=C2CCC(C)(COC3=CC=C(CC4SC(=O)NC4=O)C=C3)OC2=C1C
So what’s the problem?
String can easily be matched with GA, but …
Scoring requires sequence to correspond to real molecule
Most matings/mutations fail, i.e. many fewer paths
*Starting population
Tanimoto score between 0.23 - 0.32
Only one fragment not represented
Rediscovery using SMILES fails, despite a lot of help*
O
HN
O
S
O
O
OH
CC(C + OC = CC(COC
Emilie
Henault

12. Success Using Graph-Based Methods
Molecules are more
like crossword puzzles
crossover
Chem. Sci. 2019, J. Chem. Inf. Comput. Sci. 2004, JACS 2013
github.com/jensengroup/GB-GA
Emilie
Henault

13. O
HN
O
S
O
O
OH
O
S
O
N
NF
F
F
NH2
O
S
O
N
S
N
N
Some molecules are harder to find
Missing
fragments

14. Finding Chromophores using Genetic Algorithms
(molecules absorbing at 300-500 nm are removed from starting population)
(Computed using xTB-STDA//MMFF, population = 20)
score = λ-score + f -score
Chemical Space
Emilie
Henault

15. Finding Chromophores using Genetic Algorithms
These molecules absorb strongly round 400 nm

16. Docking using Genetic Algorithms
These molecules have better docking scores than native ligand
(Target = 𝛽2-adrenergic receptor, minimizing Glide htvs_ds score)
(Population = 400, 50 generations, 20 GA searches)
Casper
Steinmann
(Aalborg U)
native ligand

17. Docking using Genetic Algorithms
native ligand
Casper
Steinmann
(Aalborg U)

18. Is it possible to find 1 specific molecule among 1060?
Yes, if
the property of interest is cumulative function of structure
most building blocks can be identified beforehand
Because there are any many paths to the molecule
Most properties of interest have many solutions, each with many paths
Chemical Space Chemical Space
path

19. Future Directions
1060 10x 100 1
How small can we make x?
Smaller x, better scoring function