ACS Denver 2024: Generative chemistry with deep learning models

Generative deep learning networks can be used to train an encoder/decoder sequence for converting a molecular graph into a numeric vector and back again. Not only do these vectors make very good descriptors for QSAR models and similarity comparisons, but they bring additional capabilities: it is possible to tweak them in search of better properties, and decode them into hypothetical molecules that represent novel chemical entities. We will describe some derived use cases, such as (1) using the molecular vectors to effect bioisostere-like transformations, (2) performing a gradient-assisted walk through chemical space, and (3) further flattening a dataset in order to produce a visual contour which can be used to interactively search for regions that have improved activity and ADME properties.

1. Alex M. Clark
Some of the doors we unlock
by describing molecules using
generative models
[email protected]

2. Generative Models
❖Bottleneck is the latent vector (λ):
✦highly orthogonalized fixed-length vector descriptor
✦and any set of values can be used to recreate corresponding molecule
❖Most models use SMILES input & output
✦linear representation is a good fit for the technology
✦ours also uses graph inputs and fingerprint outputs
2
λ
molecular
graph input
encoder decoder
molecular
graph output

3. Navigation of Chemical Space
❖λ as a multi-dimensional
coordinate
3
❖e.g. if molecule A has good activity and B has good ADME, is there
something good in between?
A
B
?

4. Too Many Dimensions
❖λ384 necessary to get most of ChEMBL ➾ SMILES & ECFP
❖Many dimensions hard to visualize: reduction from 384 → 2
4
❖Medium sized datasets (several thousand molecules) work well
❖Very simple neural network construction
❖Transformation to μxy is reversible
encoder decoder
λ λμxy
flattenunflatten

5. Flattening Model
❖λ384 vectors reversibly projected onto 2D, displayed interactively
❖Fun fact: deep learning model is trained in-browser, native JavaScript
5
DHFR, 500 molecules

6. Activity Contours
❖Create a "gradient boost" model for λ384 → activity (fast to train)
❖Empty grid: unflatten μxy → λ384 and feed into model... colour the grid
6

7. Exploring
7
μxy λ384
unflatten
reconstruct
Nc1ncc(-c2cccc(N(CCC(=O)c3ccccc3)C(=O)C=CCONC(=O)c3ccccc3OCCC(=O)c3nc(N)ncc3Cl)c2)c(N)n1
NC1=NC(NC(=O)c2ccccc2)C(OCCCNC(=O)c2cccc(OCCC(=O)c3ccccc3N)n2)=N1
NC1=NC(NC(=O)c2ccccc2)C(=O)C=C1OCCCNc1cccc(C(=O)NCCC(=O)c2cccc(O)c2ONC(=O)c2ccccc2)n1
NC(N)=Nc1cccc(C(=O)N=C2N=C(N)N(C=CCCCOc3ccccc3-c3cccc(N)n3)C2=O)c1
N=C(N)NCCC=C1N=C(N)C(C=C(C(=O)c2cccc(Cl)c2)c2ccccc2)Oc2c(C(=O)NCCON=C(N)N)cccc21
NC1=NC(NC(=O)c2cccc(OCCCN3C=CC(CCOC(=O)c4ccccc4)N=C3N)c2)=NC1=O

8. Bio-isosteres
❖Modify the structure to:
✦keep the biological activity
✦re-roll the dice on everything else
❖Conventionally implemented as scaffold replacements
❖Pain points:
✦where to get suitable bio-isostere transforms?
✦what if you need something a bit less cleanly defined?
8
transform
AB C D
λa λb
δab
λc
λd
δab
+
encode encode encode
decode
derive latent vector offset
from known bioisostere
apply latent vector offset
decode vector to molecule

9. Quasi-bioisosteres?
9
❖Hypothesis: a translation in latent space can achieve similar effects to a
scaffold replacement
❖derive latent vector offset
❖from known bioisostere
❖apply latent vector offset
❖decode vector to molecule
AB C D
λa λb
δab
λc
λd
δab
+
encode encode encode
decode
indication of how similar transforms are to each other, in latent space

10. Why?
❖Propose a structural transformation:
✦has bioisostere-like properties
✦but not necessarily a well defined
substructure swap
10
❖Molecules A and B have good activity, but B has better ADME
❖Define transform δ as λ(A) <z λ(B) in latent vector space
❖Apply transform as λ(C) + δ to improve ADME...?
A
Bδ
C
?δ
passes the
eyeball test, but
needs to be
developed

11. Conclusions & Future Work
❖Operating in latent vector space enables new drug design tools...
❖... currently in early prototype stage
❖Generative models: shrink latent vectors? alternatives to SMILES?
❖Exploring multidimensional space via gradient
❖Contours: plot multiple activities and ADME/tox
❖Quasi-bioisosteres: can it enrich datasets?
❖TODO: get some of these tools into the hands of scientists...
11

12. Questions?
❖Contact:
✦Alex M. Clark [email protected] (Collaborative Drug Discovery)
❖Thanks to the Vault & Research Informatics teams
12