Yang Shengyong's team at Sichuan University discovered effective selective RIPK1 inhibitors through deep learning

Retrieving hit/lead compounds using novel scaffolds is an important but challenging task
during early drug development.
Various generative models have been proposed to create drug-like molecules
.
However, the ability of these generative models to design wet laboratory validation and target-specific molecules using novel scaffolds has received little validation
.

On November 12, 2022, Shengyong Yang's team from Sichuan University published a report entitled "Generative deep learning enables the discovery of a potent and selective RIPK1 inhibitor" in Nature Communications The study proposes a generative deep learning (GDL) model, a distributed learning conditional recurrent neural network (cRNN) for generating a tailored library of virtual compounds for a given biological target
.
The GDL model is then applied to RIPK1
.

Virtual screening and subsequent bioactivity assessment against the resulting custom compound library resulted in the discovery of a potent selective RIPK1 inhibitor with previously unreported scaffold RI-962
.
The compound has shown potent in vitro activity in protecting cells from necrotizing apoptosis and has good in vivo efficacy
in both inflammatory models.
Taken together, these findings demonstrate the ability of the researchers' GDL models to generate hit/lead compounds with unreported scaffolds, highlighting the great potential
of deep learning in drug discovery.

Identifying new starting active compounds with chemical structures that differ significantly from those already on the market or under development is a critical step
in the early stages of drug development.
This task is accomplished primarily through physical or virtual high-throughput screening of a set of known chemical libraries
.
However, due to the limited structural diversity of existing chemical libraries and repeated screening by various companies and institutes, it is becoming increasingly difficult
to search for active compounds and establish intellectual property rights with new scaffolds.
De novo molecular design has been proposed to generate new molecules with desired properties computationally as a solution to
this problem.
These methods are also primarily based on fragments, and the quality and diversity of the molecules generated depend largely on the fragment library and the algorithms
used for fragment assembly.

Recently, generative deep learning (GDL) has emerged as a promising de novo molecular design approach
.
Deep neural networks are used as generative models
.
This approach is a completely data-driven de novo molecular design strategy that does not require clear design rules and avoids the fragmentation problems
described above.
It has attracted a lot of attention, and several GDL models have been established to generate molecules, and a detailed description or comparison of these models can be found
in several recent reviews.
Among these models, the RNN-based model is the most widely used, and several recently proposed RNN-based GDL models have achieved impressive results
in generating new molecules due to the maturity of the RNN theoretical system.

While many GDL models, including RNN-based models, show good performance in generating molecules, most of them aim to generate optimal molecules to meet predefined goals (goal-oriented).

These goal-oriented models rely strongly on the objective function, which can lead to numerators that are numerically
superior but actually useless.
Although most GDL models have been shown to be theoretically effective, few have been validated by laboratory experiments
.
To solve these problems, the researchers propose a GDL model
based on distributed learning cRNNs.
A combination of transfer learning, regularization enhancement, and sampling enhancement enables the generation of molecules
with previously unreported and multiple chemical scaffolds.
The model is then applied to the discovery of receptor-interacting protein kinase 1 (RIPK1) inhibitors, followed by comprehensive in vitro and in vivo validation
.

RIPK1 is a serine/threonine protein kinase involved in various signaling pathways
for cell survival.
In particular, RIPK1 is a key regulator of programmed cell necrosis (necrotizing apoptosis), which is closely related to
the occurrence and progression of various inflammatory and immune diseases.
Mechanically, RIPK1 will be activated
first when necrotizing apoptosis is triggered by stimuli such as a family of tumor necrosis factors.
The activated RIPK1 then binds to and phosphorylates with its downstream protein RIPK3, which subsequently recruits and phosphorylates pseudokinase mixed lineage kinase domain (MLKL) phosphorylated MLKL to form oligomers and transfer to cell membranes to perform necrotizing apoptosis
.
Due to its central role in necrotizing apoptosis, RIPK1 is considered a promising target for the treatment of necrotizing
apoptosis-related diseases.

In this work, a GDL model based on distributed learning cRNNs is proposed, which is then applied to the discovery
of RIPK1 inhibitors.
The proposed GDL model is first introduced, and then the model is applied to generate a custom virtual compound library for RIPK1 and virtually screened
against the library.
The search
for potent and selective RIPK1 inhibitors (RI-962) is described next.
The X-ray crystal structure
of the RIPK1 with the RI-962 complex was then described.
Subsequently, the in vitro and in vivo effects of RI-962 and its pharmacokinetic properties and safety evaluation
were introduced.
Finally, a discussion and a detailed description
of the methodology used.

RI-962 reduces inflammation in acute DSS-induced colitis (Source: Nature Communications)

Overall, the discovery of a lead compound with previously unreported RIPK1 scaffolds using the GDL model proposed by the investigators witnessed the successful application
of deep neural networks in early drug discovery.

Informational message:

https://doi.
org/10.
1038/s41467-022-34692-w

： ，
。   ， ，