A new species of Dictyanthus from the state of Michoacán, Mexico, is described and illustrated: Dictyanthus reflexiflorus sp. nov. It grows in tropical deciduous forest located in the municipality of Tacámbaro. A phylogenetic analysis based on molecular data (plastid regions trnL-trnF and rps 16) was conducted to explore the relationships of the new species. Results place the new species in Dictyanthus, congruent with its morphology. Characters that support this relationship are the presence of uncinate hairs, a reticulate corolla, and a gynostegial corona basally fused to the stipe of the gynostegium and partially or completely fused to the corolla tube. The species has unique characteristics within the genus, such as a reflexed corolla and the presence of slits along the corolla lobes. A key to distinguish Dictyanthus reflexiflorus from congeners in the state of Michoacán is included.

Phylogenetic analysis—To assess the phylogenetic position of Dictyanthus, we employed sequences which had been used successfully in previous phylogenetic studies of Gonolobinae (Liede-Schumann et al. 2005; Krings et al. 2008; Mangelsdorff et al. 2016). The sequences were aligned separately using MAFFT v. 7 (Katoh and Standley 2013) with FFT–NS–i iterative refinement method, and the rest of the parameters selected by default. Subsequently, the alignments were manually edited using PhyDE v. 0.9971 (Müller et al. 2010). For the substitution model of each marker, we used Model Finder (Kalyaanamoorthy et al. 2017) employing default parameters and chosen under Akaike information criteria (AIC). Maximum Likelihood Analysis (ML) was performed in IQ-TREE web server (Trifinopoulos et al. 2016) with ultrafast bootstrapping (Hoang et al. 2018) of 1000 replicates and the rest of parameters selected by default. Consensus tree and bootstrap (BS) were visualized and edited in Fig Tree v. 1.4.3 (Rambaut 2016). Bayesian Inference (BI) was performed using Mr. Bayes v. 3.2.6 (Ronquist et al. 2012) with two independent runs and four Markov chains Monte Carlo (MCMC) each. Each chain ran for 10 million generations and was sampled every 1,000 generations. The convergence of the chains was checked using Tracer v. 1.7.1 (Rambaut et al. 2018), and according to the performance of the chains, 25% of the trees were eliminated as burn-in. The remaining topologies were used to build a majority rule consensus tree, which was viewed using FigTree v. 1.4.3 (Rambaut 2016).