Fine-Tuning a Transformer Model for METTL3 Lead Optimization

Transformers are
machine learning models originally developed
to
translate between natural languages. Recently, a transformer model
was trained on knowledge of medicinal chemistry, i.e., matched molecular
pairs of nearly a million bioactive compounds from the ChEMBL database.
Here, we customize (i.e., fine-tune) the pretrained model to enhance
the affinity and/or metabolic stability of a series of inhibitors
of methyltransferase-like protein 3 (METTL3). We first fine-tune the
transformer model using a data set of about 500 METTL3 inhibitors
with known binding affinities and validate it by retrospective analysis.
Then, we fine-tune the original transformer model to simultaneously
optimize binding affinity and metabolic stability in a prospective
application. Two of the five METTL3 inhibitors predicted by the multiobjective
optimized model show low-nanomolar potency and higher stability than
the lead compound of the chemical series used for fine-tuning.