DeepAS – Chemical language model for the extension of active analogue series

Extension of multi-site analogue series with potent compounds using a bidirectional transformer-based chemical language model

“…20,21 MMS are defined as series of compounds that are only distinguished by a chemical modification at a single site 20 and can be identified using a matched molecular pair (MMP) algorithm variant. 21 For the original development of DeepAS, 6 104 627 MMS from 2195 target-based compound activity classes were extracted from ChEMBL. 22 From these MMS, all R-groups were systematically isolated, yielding 3852 unique R-groups, which represented all R-groups occurring in AS covering the entire target space of active compounds.…”

“…The length of a sentence was consistently set to 35 tokens and the total number of label tokens amounted to 3855, including 3852 unique R-groups extracted from the qualifying MMS plus the three special tokens. 6…”

“…Such models can be derived to learn a variety of sequence-to-sequence mappings for diverse design tasks, giving rise to different types of chemical language models (CLMs). [6][7][8][9][10][11][12][13][14] Popular DNN architectures from NLP include recurrent neural networks (RNNs), which were first adapted for applications in chemistry, [4][5][6][7] and different transformer networks that are increasingly used for molecular design and de novo compound generation. [10][11][12][13][14][15] Different deep generative models including transformers and diffusion models have also been developed in structure-based drug design for the discovery and optimization of new active compounds.…”

Chemical language models for molecular design

“…Another CLM was developed for lead optimization in medicinal chemistry. This model was trained to iteratively extend analogue series with new potent compounds [35]. For this purpose, analogue series can be perceived as sequences of R‐groups attached to an invariant core structure, which provides the conceptual basis for series extension via CLMs.…”

“…New R‐groups were then predicted by the dense layer of the model based on conditional probabilities derived from preceding R‐groups. Given the ordering of training series, iterative extension of analogue series using DeepAS was implicitly directed towards compounds with increasing potency, and the model correctly reproduced potent analogues for many test series [35].…”