Ancestral Sequences Cannot be Accurately Reconstructed via Interpolation in a Variational Autoencoder’s Latent Space

{"Abstract":[0],"Standard":[1],"methods":[2,102],"for":[3,12,187],"ancestral":[4],"sequence":[5,18,72],"reconstruction":[6],"(ASR)":[7],"rely":[8],"on":[9],"substitution":[10],"models":[11,38],"the":[13,27,61,75,118,157,166,170,175,178,188],"residues":[14],"in":[15,50,130,156],"a":[16,51,90,151],"biological":[17],"and":[19,88,100,112,123],"assume":[20],"independent":[21],"evolution":[22],"across":[23],"these":[24,58],"sites,":[25],"ignoring":[26],"epistatic":[28,131],"interactions":[29],"that":[30,54,117,144],"shape":[31],"molecular":[32],"evolution.":[33],"In":[34,79],"contrast,":[35],"deep":[36],"learning":[37],"like":[39],"variational":[40],"autoencoders":[41],"(VAEs)":[42],"can":[43],"learn":[44],"low-dimensional":[45],"representations":[46],"(“embeddings\\")":[47],"of":[48,63,106,153,191],"sequences":[49,183],"protein":[52,107],"family":[53],"may":[55],"implicitly":[56],"handle":[57],"dependencies,":[59],"raising":[60],"possibility":[62],"performing":[64],"more":[65],"accurate":[66],"ASR":[67,92],"by":[68,86,126],"interpolating":[69],"between":[70],"extant":[71],"embeddings":[73],"within":[74],"VAE’s":[76,179],"latent":[77,158],"space.":[78],"this":[80,84,95,145],"study,":[81],"we":[82,115],"test":[83],"hypothesis":[85],"developing":[87],"evaluating":[89],"VAE-based":[91,119],"pipeline.":[93],"Benchmarking":[94],"approach":[96,120],"against":[97],"established":[98],"likelihood-based":[99],"parsimony":[101],"using":[103],"various":[104],"simulations":[105],"evolution,":[108],"including":[109],"scenarios":[110],"with":[111,184],"without":[113],"epistasis,":[114],"find":[116],"is":[121,147,169],"consistently":[122],"significantly":[124],"outperformed":[125],"standard":[127],"methods,":[128],"even":[129],"regimes":[132],"where":[133],"it":[134],"was":[135],"hypothesized":[136],"to":[137,150,174],"have":[138],"an":[139],"advantage.":[140],"We":[141],"further":[142],"show":[143],"failure":[146],"not":[148],"due":[149],"lack":[152],"phylogenetic":[154],"structure":[155],"space,":[159],"which":[160],"does":[161],"contain":[162],"evolutionary":[163],"signal.":[164],"Rather,":[165],"primary":[167],"limitation":[168],"information":[171],"loss":[172],"inherent":[173],"autoencoding":[176],"process:":[177],"decoder":[180],"cannot":[181],"generate":[182],"sufficient":[185],"fidelity":[186],"precise":[189],"demands":[190],"ASR.":[192]}

Authors

Institutions

Publication Details

Journal
Bulletin of Mathematical Biology
Published
2026-09-19
DOI
https://doi.org/10.1007/s11538-026-01749-6
Primary Topic
Genomics and Phylogenetic Studies
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
article

Ancestral Sequences Cannot be Accurately Reconstructed via Interpolation in a Variational Autoencoder’s Latent Space

Mengze Tang, Claudia Solís‐Lemus, Hailey Bruzzone, Evan Gorstein
Bulletin of Mathematical Biology
Genomics and Phylogenetic Studies
article

Ancestral Sequences Cannot be Accurately Reconstructed via Interpolation in a Variational Autoencoder’s Latent Space

Mengze Tang, Claudia Solís‐Lemus, Hailey Bruzzone, Evan Gorstein
article en

Abstract

Abstract Standard methods for ancestral sequence reconstruction (ASR) rely on substitution models for the residues in a biological sequence and assume independent evolution across these sites, ignoring the epistatic interactions that shape molecular evolution. In contrast, deep learning models like variational autoencoders (VAEs) can learn low-dimensional representations (“embeddings") of sequences in a protein family that may implicitly handle these dependencies, raising the possibility of performing more accurate ASR by interpolating between extant sequence embeddings within the VAE’s latent space. In this study, we test this hypothesis by developing and evaluating a VAE-based ASR pipeline. Benchmarking this approach against established likelihood-based and parsimony methods using various simulations of protein evolution, including scenarios with and without epistasis, we find that the VAE-based approach is consistently and significantly outperformed by standard methods, even in epistatic regimes where it was hypothesized to have an advantage. We further show that this failure is not due to a lack of phylogenetic structure in the latent space, which does contain evolutionary signal. Rather, the primary limitation is the information loss inherent to the autoencoding process: the VAE’s decoder cannot generate sequences with sufficient fidelity for the precise demands of ASR.

Bulletin of Mathematical BiologyVol. 88(10)
University of Wisconsin–Madison (US), Wisconsin Institutes for Discovery (US)
Openalex Percentile: Top 18%
Genomics and Phylogenetic Studies
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.