design of biologically active small is intrinsically
complicated and hard, and it is getting harder and more
complicated every day. For one thing, it is inherently multi-objective
with improvement in one attribute - e.g., potency -
often coming at the expense of another - e.g.,
bioavailability. Moreover, attributes such as half-life
exhibit optima as well as trends, requiring a careful
balancing act between being cleared too quickly and
being cleared too slowly.
Given that complexity, it makes sense to bring as many
disciplines to bear on the problem as possible; cranking
blindly through black box models or a synthetic strategy
of methyl, ethyl, butyl, futyl ... will not get the job
done. That is why drug discovery and development has
evolved into a team activity, with people from many
disciplines working together.
Biochemical Infometrics was founded in 2008 as a way to
make my drug design experience available to smaller
companies and it closed down in January 2010 when I
became Director of Life Sciences at Simulations
. My time there afforded me many great
opportunities to apply my inter- and trans-disciplinary
perspective to challenging modeling problems in DMPK and
predictive toxicology. Retiring from Simulations Plus on
April Fool's Day in 2022 has given me more freedom to
write, to collaborate, and to explore new approaches to
old problems. My Adjunct Professorship in Informatics at
provides some excellent
opportunities for doing that, but I am also resurrecting
Biochemical Informatics as a scientific consulting group
- albeit a nonprofit one, at least for the time being."
20 August 2022
Fall National ACS Meeting in Chicago
Coupling high-throughput pharmacokinetic
simulation (HTPK) to multi-objective molecular
evolution of triazolopyrimidine antimalarial leads
De novo drug design is a
long-standing challenge in computational chemistry.
Much recent work has centered on using generative
neural networks and related technologies to create
chemical libraries that reproduce property &or
structural distributions that mimic those of
compounds known to be active against a selected
target. The AI-driven Drug Design (AIDD) module in
ADMET Predictorฎ takes a different approach: in
AIDD, one or more molecules are submitted to
randomly chosen SMIRKS transformations, yielding
products that are then evaluated against a panel of
structural SMARTS filters and objective functions.
Those that pass the filters and are Pareto optimal
with respect to the objectives and to the prior
generation of analogs are carried forward as the
next generation, which is itself subject to
additional transformations, and so on.
The case study to be presented here involves a
series of triazolopyrimidines (TzPs) that inhibit
the dihydroorotate dehydrogenase of the malarial
parasite, Plasmodium falciparum (PfDHODH)
. Examples from very early in the synthesis
program were input to AIDD as structural seeds from
which to evolve several series of analogs optimized
against four objective functions: a general
artificial neural net ensemble (ANNE) model of PfDHODH
inhibition with a single hidden layer of neurons
; bioavailability estimates obtained by
high-throughput pharmacokinetic (HTPK) simulation;
synthetic difficulty estimates; and a set of
absorption, distribution, metabolism, excretion and
toxicity (ADMET) property estimates encoded as a
weighted set of rules (AIDD Risk).
The output analogs obtained afforded good structural
coverage of the activity space, were novel but
synthetically reasonable. They included several that
were actually synthesized later in the TzP program.
Those subsequently synthesized analogs included
several that were not in the data set used to build
the ANNE activity model. Hence this
case study is largely prospective in nature
despite being drawn from the literature.
1. M.A. Phillips et al., J.
Med. Chem. 2008, 51, 36493653.
2. Clark et al., J. Comput.-Aided Drug Des.
2020, 34, 1117-1132.
Hedge Your Bets to Come
If you watch the
people who make a living at racetracks, you will see
that they do not pick winners. Instead, they
typically bet on several horses to Win, Place or
Show. In many ways, drug development is a
similar game, in that you are operating with limited
and imperfect information and chance plays a big part
in how your corporate gambles turn out.
Intelligent molecular design can add value - a lot of value - by
helping you spread your risk rationally across leads
and candidates. Most people already appreciate
the importance of cultivating structural diversity and
synthetic accessibility when picking compounds for
follow-up acquisition or synthesis, but there is as
much or more to be gained by diversifying across
likely ADMET and PK risks as well. Being able to
make relevant and reliable predictions of the
corresponding properties within the realm of
well-explored chemistry is critical. Such
predictions are still something of an art when one
ventures into new areas, but that is where the value
ฉ 2022 Biochemical Infometrics; all rights reserved.
QSAR & statistics