Total RNA was extracted from Populus using the RNA extraction kit (Tiangen, Beijing, China) following the manufacturer’s instructions. We used UEIris Ⅱ RT-PCR System for First-Strand cDNA Synthesis (with dsDNase) kit (Biorigin, Beijing, China) according to the manufacturer’s instruction. The full-length sequence of PtoCOMT was amplified by PCR with 25μL 2 × Phanta Max Master Mix, 2μL forward primers and reverse primers, 2μLcDNA and 19μL double distilled water at The PCR reaction was carried out in a total volume of 50 μL 95℃ for 3 min; 34 cycles of 95℃for 30 s, 56℃for 30 s and 72℃for 90 s. The PCR product was separated on an agarose gel (1.5%).

The recovered bands was cloned into the pEASY vector according to the manufacturer’s protocols and sequenced. Then the cDNA was PCR-amplified with specific primers harboring XbaI and KpnI restriction sites from pEASY-PtoCOMT and was then ligated into the PBI121vector. Recombinant PBI121-PtoCOMT was transformed into populus by Agrobacterium tumefaciens-mediated transformation.

Endogenous melatonin was extracted uasing PDS in plant leaves. After extraction the centrifuged extract from leaves was sed for quantification using Melatonin ELISA Kit (Ruixin Biotech, Quanzhou, China) as the manufacturer’s instruction described.

COMT protein sequences of multiple species were collected from NCBI (https://www.ncbi.nlm.nih.gov/), and then MEGA (https://www.megasoftware.net/) was used to perform multiple sequence alignment and draw phylogenetic evolutionary trees. Then, positive selected amino acid sites were analyzed by EasyCodeML (https://www.java.com/zh-CN/). And the library of potential mutants was constructed according to the positive selected amino acid sites. We used alphafold (https://alphafold.ebi.ac.uk/) to predict protein structure.

In order to predict protein 3D structures, we utilized ColabFold, which is accessible as a set of Jupyter notebooks on Google Colaboratory. ColabFold mainly consists of two parts. The first is an MMseqs2-based homology search server to build diverse MSAs and to find templates. The input sequence(s) was efficiently aligned against the databanks including UniRef100, PDB70 and an environmental sequence set. The second portion is a Python library that interacts with the MMseqs2 search engine, gets the input features ready for structure inference (as single chains or complexes), and displays the outcomes.

The results show two confidence measures. One is the predicted local-distance difference test (pLDDT), which reliably predicts the Cα local-distance difference test (lDDT-Cα) accuracy of the corresponding prediction. The other is predicted aligned error (PAE) indicating that ColabFold predicts well-defined relative positions and orientations for residue pairs x, y from two different domains, when it is generally low for them (https://alphafold.com/). Finally, we uploaded top-ranked model to Mol* 3D Viewer of RCSB PDB for visualization (https://www.rcsb.org/3d-view/).