DeePhy

Reconstructing a phylogenetic tree conventionally means aligning sequences first, then running neighbour joining, maximum parsimony, maximum likelihood, or Bayesian inference on the alignment. Every one of those methods has branch-length regimes where it systematically recovers the wrong topology — long-branch attraction and its relatives — and the alignment step itself introduces choices that propagate into the tree. No method is reliable everywhere.

DeePhy, from Aritra Mahapatra and Jayanta Mukherjee at the Indian Institute of Technology Kharagpur, skips alignment entirely. Each nucleotide sequence is turned into a Genomic Footprint (GFP): a curve on a two-dimensional plane in which each nucleotide advances the walk by a step whose magnitude comes from the self-information of that base under a second-order Markov model, and whose direction comes from a structural grouping of the bases. DeePhy uses the amino-keto grouping, which earlier work found carries the most usable signal. Three such curves — one per operational taxonomic unit — are stacked into a matrix and handed to a convolutional network that classifies which of the three possible rooted triplet topologies is correct.

Working at the triplet is the design choice that makes the supervised framing tractable: true trees are scarce, but triplets can be simulated in bulk, and a rooted tree on n taxa decomposes into a set of triplets that can be reassembled afterwards. This places DeePhy alongside other alignment-free learned approaches in the catalog such as BetaInfer, which transduces unaligned sequences directly into Newick trees, and LucaPhylo, which derives distances from protein language model embeddings.

#Key Features

• No alignment required: The genomic footprint is computed per sequence, so taxon sampling and alignment ambiguity never enter the inference.
• Triplet decomposition and reassembly: The network solves the three-way problem; full trees are built by amalgamating all triplets with the SuperTriplets supertree method.
• A bootstrap that works without alignments: Felsenstein resampling needs aligned columns, so the authors resample by perturbing each unaligned sequence — insertion, deletion, and substitution at 0.33% each, for 1% total variation, matching observed intraspecific mitochondrial divergence — and score clades with BOOSTER transfer distance.
• Steady across branch-length space: Accuracy varies far less across branch-length subspaces than the conventional methods, whose relative ranking flips between zones.
• Validated on real data: The trained model was applied unchanged to 13 protein-coding mitochondrial genes from 20 Gadiformes species.

#Technical Details

The network takes a 2 × 3 × 1,356 input — the x and y coordinates of three footprints, padded to the longest training sequence — through five convolutional layers with ReLU activations, 64 filters in the first and last and 128 in the three intermediate layers, a 2 × 3 kernel in the first layer and 1 × 3 thereafter, average pooling of size 1 × 3, batch normalization after each layer, and dropout at 0.10. Two fully connected layers (64 nodes, then 3) emit the topology class. Training used 6.5 × 10⁵ simulated sequences per substitution model, with Jukes-Cantor and general time-reversible data trained separately: 6 × 10⁵ for training and 3 × 10⁴ for validation, Adam at an initial learning rate of 0.00001 with cross-entropy loss and early stopping, implemented in PyTorch. Evaluation used 100 independent test sets of 3 × 10⁴ triplets each, sampled evenly from 5,995 branch-length subspaces. Mean accuracy was 78.21% (s.d. 0.437), against 72.44% for maximum parsimony, 72.22% for Bayesian inference, 71.21% for maximum likelihood, and 69.32% for neighbour joining; McNemar tests give p < 1 × 10⁻²⁰. On bootstrap replicas the figure is 78.82%. Maximum parsimony additionally failed to resolve 12.29% of cases.

#Applications

DeePhy suits gene-tree reconstruction where alignment is awkward or the branch lengths fall in a zone that trips conventional estimators — mitochondrial gene sets, and sequence collections spanning wide divergence. In the Gadiformes case study, gene trees inferred for each of the 13 genes were amalgamated into a species tree whose Deformity Index against widely accepted subfamily hypotheses was better than trees from the conventional methods.

#Impact

The work is a preprint and has not been peer reviewed, and its scope is honestly bounded. Accuracy holds for sequences up to the 1,356 bp training length; extending to whole mitochondrial genomes by splitting them into overlapping windows and taking a majority vote worked poorly, which the authors read as evidence that evolutionary signal is not distributed uniformly along a genome. The model returns topology only, with no branch lengths. Code and a trained checkpoint are distributed from a GitHub mirror under a BSD 3-Clause licence and from an IIT Kharagpur project page; deriving the genomic footprints requires a separate tool from the same group.