AlphaDrug: protein target specific de novo molecular generation

Abstract

Traditional drug discovery is very laborious, expensive, and time-consuming, due to the huge combinatorial complexity of the discrete molecular search space. Researchers have turned to machine learning methods for help to tackle this difficult problem. However, most existing methods are either virtual screening on the available database of compounds by protein-ligand affinity prediction, or unconditional molecular generation which does not take into account the information of the protein target. In this paper, we propose a protein target-oriented de novo drug design method, called AlphaDrug. Our method is able to automatically generate molecular drug candidates in an autoregressive way, and the drug candidates can dock into the given target protein well. To fulfill this goal, we devise a modified transformer network for the joint embedding of protein target and the molecule, and a Monte Carlo Tree Search (MCTS) algorithm for the conditional molecular generation. In the transformer variant, we impose a hierarchy of skip connections from protein encoder to molecule decoder for efficient feature transfer. The transformer variant computes the probabilities of next atoms based on the protein target and the molecule intermediate. We use the probabilities to guide the look-ahead search by MCTS to enhance or correct the next-atom selection. Moreover, MCTS is also guided by a value function implemented by a docking program, such that the paths with many low docking values are seldom chosen. Experiments on diverse protein targets demonstrate the effectiveness of our methods, indicating that AlphaDrug is a potentially promising solution to target-specific de novo drug design.

Chemical language models for molecular design

“…Furthermore, both RNN and transformer models have been extensively used for chemical reaction prediction [18][19][20][21], representing a growth area for CLM applications. Furthermore, transformer variants have been employed to predict new compounds from protein sequences [11,[22][23][24]. These applications translating protein sequences into molecular strings are closely related to others focusing on the prediction of drug-target interactions [24][25][26].…”

Motif2Mol: Prediction of New Active Compounds Based on Sequence Motifs of Ligand Binding Sites in Proteins Using a Biochemical Language Model

“…In a few studies, RNNs or transformers have also been applied to associate protein and ligand representations. Specifically, three studies have attempted to generate new small molecule ligands from target protein sequences via language models [22][23][24]. Hence, in these cases, the machine translation task required the derivation of models to construct SMILES representations encoding new compounds from amino acid sequences of targets.…”

“…Therefore, a DNN was employed to generate protein sequence vector embeddings [26] that served as input for an RNN comprising multiple LSTM units to generate SMILES strings of new compounds via reinforcement learning [22]. In addition, two methodologically distinct studies trained transformer networks to directly associate protein sequences with SMILES of known compounds and generate new molecules [23,24]. Therefore, a transformer architecture with an attention mechanism was adapted [14,23].…”

“…Therefore, a transformer architecture with an attention mechanism was adapted [14,23]. Furthermore, a transformer variant was developed based on the Lmser network [27] to combine embeddings of protein sequences and molecule strings and generate new compounds conditioned on given protein sequences via Monte Carlo tree search over intermittent strings [24]. Both transformer-based approaches used conventional proteinligand docking scores to assess or the guide compound design [23,24], which increased the intrinsic uncertainty of the design approach (that is, one hypothetical model was employed to guide another).…”

“…Since the transformer relates amino acid sequences to molecular structure, it is termed a "biochemical language model". Different from the earlier studies [22][23][24] that learned from complete protein sequences, we use sequence motifs defining ligand binding regions as input to avoid redundancy or noise of sequence information that might not be relevant for ligand binding. In addition, we apply a compound evaluation scheme that does not rely on hypothetical scoring, but directly assess the ability of a model to produce active compounds.…”