Deep reinforcement learning yields two new-scaffold anti-influenza compounds: at 100 mg/kg they protected 65% and 100% of infected mice
Synopsis
A deep reinforcement learning molecule-design platform, given no information about known active molecules, grew 60,291 unique compounds from a benzene ring, and 86.5% scored higher on its reward than oseltamivir. After filtering for synthetic accessibility, drug-likeness, docking and MM/PBSA, nine compounds were synthesized; DS-22-inf-009 and DS-22-inf-021 inhibited influenza A and B viruses in cells with EC50 values from 0.29 to 2.31 μM, and in mice at 100 mg/kg protected 65% and 85% against H1N1 pdm09 and 65% and 100% against Yamagata-lineage influenza B.
Interpretation
A sizeable share of the compounds generated by a deep reinforcement learning platform outscored the existing anti-influenza drug oseltamivir on physicochemical reward, and these molecules bore almost no chemical resemblance to oseltamivir. Earlier reinforcement-learning antiviral design stayed computational, and the designed molecules had not been tested in vitro or in vivo; most current neuraminidase-inhibitor efforts also concentrate on sialic-acid-containing compounds. Three independent runs generated 60,291 unique structures, of which 86.5% had a higher reward than oseltamivir's 0.69; the 31 selected compounds had an average pairwise Tanimoto similarity of 0.258 and an average similarity to oseltamivir of only 0.13.
Two selected compounds showed activity against influenza A and B viruses at the enzyme, cell and mouse levels, including against the oseltamivir-resistant H275Y mutant. Among the nine synthesized and tested compounds, DS-22-inf-009 and DS-22-inf-021 inhibited influenza A/WSN/33 neuraminidase with IC50 values of 24.8 μM and 11.2 μM, influenza B neuraminidase at 21.2 μM and 9.1 μM, and the H275Y mutant neuraminidase at 18.2 μM and 14.7 μM. In cells, DS-22-inf-009 gave EC50 values of 0.56, 2.16 and 2.31 μM against A/St.Petersburg/63/2020, A/Vladivostok/2/2009 and B/Samara/32/2018, and DS-22-inf-021 gave 0.4, 0.29 and 0.74 μM, with CC50 above 50 μM for both; at 100 mg/kg in mice, survival protection was 65% and 85% against H1N1 pdm09 and 65% and 100% against Yamagata-lineage influenza B.
The agent builds molecules from scratch without any external compound dataset and without information about a reference ligand. Conventional generative models usually learn a distribution from known active molecules or large compound libraries and tend to stay close to existing chemical scaffolds. The reinforcement learning is framed as a Markov decision process whose state is a molecular graph and whose actions are atom- and fragment-level edits, with a reward formed by multiplying docking binding energy, a normalized synthetic-accessibility score and QED drug-likeness; the policy network is an Attentive FP graph neural network combined with Double Q-Learning and prioritized experience replay.
Perspective
The demonstrated scope is computational design against influenza A and B neuraminidase, enzyme-inhibition assays on nine compounds, MDCK cell assays on two compounds, and survival and lung-titer experiments in BALB/c mice. The intended readers are researchers in antiviral drug discovery and AI molecular design; the stated next step is further optimization to improve efficacy and identification of the full spectrum of antiviral targets engaged beyond neuraminidase inhibition.
The compounds' antiviral activity in cells exceeds their neuraminidase inhibition (EC50 below IC50), which the authors take as a sign of mechanism(s) beyond neuraminidase inhibition, though these are not yet identified. In mice the compounds were dosed at 100 mg/kg, five times oseltamivir's 20 mg/kg, and oseltamivir reduced lung viral titers at lower concentrations. The authors also note the compounds came from a single design iteration, so later optimization may raise activity.
