Ancestral sequence reconstruction (ASR) is a foundational task in evolutionary biology, providing insights into the molecular past and guiding studies of protein function and adaptation. Conventional ASR methods rely on a multiple sequence alignment (MSA), a phylogenetic tree, and an evolutionary model. However, alignments and trees are often uncertain, and the evolutionary models typically assume a single, simplified process of substitutions and insertion–deletion (indel) events for all sites. Here, we introduce BetaReconstruct, a generative approach to ASR that builds on recent advances in natural language processing (NLP) and uses a hybrid transformer–state-space architecture. The model was first trained on large-scale simulated datasets with known ancestral sequences and then fine-tuned on empirical protein sequences. Reconstruction requires neither an MSA nor a phylogenetic tree; to reconstruct an internal node, the user only specifies which sequences descend from it (ingroup) and which do not (outgroup). On simulated data, BetaReconstruct achieved accuracy comparable to that of maximum-likelihood and Bayesian methods and inferred fewer false ancestral insertions. Its accuracy remained similar under evolutionary models not seen during training. When different regions of the sequence evolved under different processes, BetaReconstruct was more accurate than maximum-likelihood methods and close to the Bayesian method. In a structural analysis of mammalian proteins, the ancestors reconstructed by BetaReconstruct were, on average, more similar to their descendants than to the outgroup, whereas maximum-likelihood reconstructions were pulled toward the outgroup. This work provides a scalable, alignment-free strategy for ASR that can be adapted to new evolutionary processes by changing the simulations used for training.
Illustration of ASR using BetaReconstruct. (a) The ‘true’ evolutionary dynamics, in which the ancestral sequence “AAMM” evolved along a phylogenetic tree. The leaf sequences are “AAL”, “AYM”, and “ATMMM”. The target internal node, “AAM”, is the ancestor of the ingroup sequences “AAL” and “AYM”, and the sequence "ATMMM" serves as the outgroup. (b) The BetaReconstruct pipeline: (Ⅰ) the unaligned protein sequences are provided as input; (Ⅱ) the sequences are concatenated into a single string, in which the outgroup sequence is followed by a “#” character, ingroup sequences are separated by a “|” character, and a special “” token marks the start of the prediction; (Ⅲ) the model processes the input; (Ⅳ) the model generates the ancestral sequence. (c) A typical ASR pipeline, which runs multiple independent tools: (Ⅰ) the unaligned protein sequences are provided as input; (Ⅱ) an alignment program infers an MSA; (Ⅲ) the resulting MSA, including newly introduced gap characters; (Ⅳ) a tree inference program uses the MSA to infer the phylogeny; (Ⅴ) the inferred tree, shown in a Newick format; (Ⅵ) specialized ASR software uses both the MSA and the tree as inputs to perform the reconstruction; (Ⅶ) the final output is the reconstructed ancestral sequence. Joint methods such as BAli-Phy combine steps (Ⅱ)–(Ⅵ).
We have released three pre-trained models on the Hugging Face Hub, available for public download. The three models share the same architecture (a hybrid transformer–state-space model, Zamba) and tokenizer, and were first trained on large-scale simulated datasets.
- Internal-node reconstruction model
This model reconstructs the ancestral sequence at an internal node: the ancestor of the ingroup sequences, given one or more outgroup sequences. The input is each outgroup sequence followed by #, then the ingroup sequences separated by |, and the <RECONSTRUCT> token (for example, ATMMM#AAL|AYM<RECONSTRUCT>). It was trained with two outgroup sequences and 10–14 sequences in total. Use it with align_infer_or_reconstruct.py --reconstruct (see Align-Infer-Reconstruct pipeline).
- Mammalian model
This model was further fine-tuned on mammalian proteins from OrthoMaM (version 12; Allio et al., 2024), with the same internal-node input format. In each training example, the model received eight sequences of one mammalian clade (the ingroup) and two marsupial sequences (the outgroup), and was trained to predict a held-out sequence of the same clade. Use it to reconstruct ancestral mammalian proteins. For families with many species, use predict_ancestral_hierarchical_approach.py (see Mammalian proteins (OrthoMaM)).
- Multi-task model (alignment, tree inference and root reconstruction)
This model was trained jointly on three tasks, given a set of unaligned sequences: multiple sequence alignment (--align), phylogenetic tree inference (--infer), and reconstruction of the ancestral sequence at the root (--reconstruct). The input is the sequences separated by |, followed by the task token (<ALIGN>, <INFER> or <RECONSTRUCT>); no outgroup is given. Use it with align_infer_or_reconstruct.py.
[IMPORTANT] Note on Generalization: While we have verified the models' ability to generalize to out-of-distribution data (as detailed in the main text), performance is highest when inference is performed on data distributions similar to those used during training.
The models utilize a data-driven tokenizer specifically optimized for evolutionary related protein sequences. Instead of relying on standard single-amino-acid characters, this approach allows the model to learn and represent frequently occurring sequence patterns as distinct tokens. For a comprehensive technical breakdown of the tokenizer architecture and its training process, please refer to the main text of our publication.
Related Resources
- BiologicalTokenizers repository: You can find the source code and training scripts for biological tokenizers.
- Research paper: For detailed methodology and performance benchmarks, please see our paper: "Effect of tokenization on transformers for biological sequences" (2024; Bioinformatics).
This script runs a ZAMBA-based language model for performing one of three tasks in molecular phylogenetics:
- Alignment of unaligned sequences: See also: Multiple sequence alignment as a sequence-to-sequence learning problem (2023; ICLR) and BetaAlign: a deep learning approach for multiple sequence alignment (2025; Bioinformatics)
- Inference of phylogenetic relationships
- Reconstruction of ancestral sequences
The model works with FASTA or CSV inputs. The models were trained on simulated data, containing 10 - 14 species. Input examples are provided in example_inputs_for_align_infer_or_reconstruct
Argument Description
- --model_path (str, Required) Path or name of the pretrained ZambaForCausalLM model.
- --align (flag) Use this flag to perform sequence alignment.
- --infer (flag) Use this flag to perform phylogenetic inference.
- --reconstruct (flag) Use this flag to reconstruct ancestral sequences.
- --input (str, Required) Path to input file or folder. Must be either:
- a. a csv file with unaligned_seqs and num_species columns
- b. a fasta file
- c. a folder containing .fasta files
- --output (str, Required) Path to the output file (for CSV or FASTA) or folder (for batch FASTA processing).
The sequences named in --outgroup (a .txt file with one name per line, or comma-separated names) form the outgroup, and all other sequences form the ingroup. The ancestor of the ingroup is reconstructed.
export BETA_RECONSTRUCT="<PATH_TO_BETA_RECONSTRUCT>"
conda activate $BETA_RECONSTRUCT/python_env/
export HF_DATASETS_CACHE="$BETA_RECONSTRUCT/python_env/cache/"
export HF_HOME="$BETA_RECONSTRUCT/python_env/cache/"
export MODEL_PATH="dotan1111/BetaReconstruct_Configuration1_internal_nodes"
export INPUT="$BETA_RECONSTRUCT/example_inputs_for_align_infer_or_reconstruct/internal_node_example.fasta"
export OUTGROUP="$BETA_RECONSTRUCT/example_inputs_for_align_infer_or_reconstruct/internal_node_outgroup.txt"
export OUTPUT="$BETA_RECONSTRUCT/outputs/internal_node_example_prediction.fasta"
cd $BETA_RECONSTRUCT
mkdir -p "$(dirname "$OUTPUT")"
python "align_infer_or_reconstruct.py" \
--model_path $MODEL_PATH \
--input $INPUT \
--output $OUTPUT \
--outgroup $OUTGROUP \
--reconstruct
Each row is one instance. The script uses the seq_x column (or the outgroup and ingroup columns) and adds a prediction_reconstruct column. The test sets in data/ can be used the same way.
export INPUT="$BETA_RECONSTRUCT/example_inputs_for_align_infer_or_reconstruct/internal_node_test.csv"
export OUTPUT="$BETA_RECONSTRUCT/outputs/internal_node_predictions.csv"
python "align_infer_or_reconstruct.py" \
--model_path $MODEL_PATH \
--input $INPUT \
--output $OUTPUT \
--reconstruct
export BETA_RECONSTRUCT="<PATH_TO_BETA_RECONSTRUCT>"
conda activate $BETA_RECONSTRUCT/python_env/
export HF_DATASETS_CACHE="$BETA_RECONSTRUCT/python_env/cache/"
export HF_HOME="$BETA_RECONSTRUCT/python_env/cache/"
export MODEL_PATH="dotan1111/BetaReconstruct_Configuration1"
export INPUT="$BETA_RECONSTRUCT/example_inputs_for_align_infer_or_reconstruct/test.csv"
export OUTPUT="$BETA_RECONSTRUCT/outputs/trees.csv"
cd $BETA_RECONSTRUCT
mkdir -p "$(dirname "$OUTPUT")"
python "align_infer_or_reconstruct.py" \
--model_path $MODEL_PATH \
--input $INPUT \
--output $OUTPUT \
--infer
export BETA_RECONSTRUCT="<PATH_TO_BETA_RECONSTRUCT>"
conda activate $BETA_RECONSTRUCT/python_env/
export HF_DATASETS_CACHE="$BETA_RECONSTRUCT/python_env/cache/"
export HF_HOME="$BETA_RECONSTRUCT/python_env/cache/"
export MODEL_PATH="dotan1111/BetaReconstruct_Configuration1"
export INPUT="$BETA_RECONSTRUCT/example_inputs_for_align_infer_or_reconstruct/example1.fasta"
export OUTPUT="$BETA_RECONSTRUCT/outputs/example1_aligned.fasta"
cd $BETA_RECONSTRUCT
mkdir -p "$(dirname "$OUTPUT")"
python "align_infer_or_reconstruct.py" \
--model_path $MODEL_PATH \
--input $INPUT \
--output $OUTPUT \
--align
export BETA_RECONSTRUCT="<PATH_TO_BETA_RECONSTRUCT>"
conda activate $BETA_RECONSTRUCT/python_env/
export HF_DATASETS_CACHE="$BETA_RECONSTRUCT/python_env/cache/"
export HF_HOME="$BETA_RECONSTRUCT/python_env/cache/"
export MODEL_PATH="dotan1111/BetaReconstruct_Configuration1"
export INPUT="$BETA_RECONSTRUCT/example_inputs_for_align_infer_or_reconstruct/folder_with_fasta"
export OUTPUT="$BETA_RECONSTRUCT/outputs/ASR_results_folder"
cd $BETA_RECONSTRUCT
mkdir -p "$OUTPUT"
python "align_infer_or_reconstruct.py" \
--model_path $MODEL_PATH \
--input $INPUT \
--output $OUTPUT \
--reconstruct
The mammalian test set is in data/mammalian_validation/: 40 protein families from OrthoMaM version 12 (Allio et al., 2024), the mammalian test families of the paper. In each family, the task is to reconstruct the ancestor of the placental mammals:
- Ingroup: the placental species (Placentalia), about 164 per family.
- Outgroup: the marsupial species (orders Dasyuromorphia, Didelphimorphia, Diprotodontia, Microbiotheria (Dromiciops gliroides) and Peramelemorphia), up to 8 per family.
- The monotremes (platypus and echidna), which are neither placental nor marsupial, were removed.
Each family has its own folder with
ingroup.fastaandoutgroup.fasta(unaligned sequences, named by species). The folder also containsmammalian_test_families.csv(all families in the input format of the simulated test sets),family_counts.csv(number of ingroup and outgroup species per family), and the script that created these files from OrthoMaM. Families 283297 and 5535 contain no marsupials, so their ingroup ancestor cannot be reconstructed. In the paper, the reconstructed ancestors were evaluated through their predicted structures (Supplementary Information S5). Because each family has about 170 species, the ancestors are reconstructed hierarchically withpredict_ancestral_hierarchical_approach.py(see Inferring ancestral sequence for many species), which reads the folder directly.
Reconstruct the ancestor of the placental mammals for one family (2331), then for all families:
export BETA_RECONSTRUCT="<PATH_TO_BETA_RECONSTRUCT>"
conda activate $BETA_RECONSTRUCT/python_env/
export HF_DATASETS_CACHE="$BETA_RECONSTRUCT/python_env/cache/"
export HF_HOME="$BETA_RECONSTRUCT/python_env/cache/"
export MODEL_PATH="dotan1111/BetaReconstruct_Mammals_Configuration1"
export RESULTS_FOLDER="$BETA_RECONSTRUCT/results/mammalian/"
export SPECIES_LIST_FOLDER="$BETA_RECONSTRUCT/species_lists/"
cd $BETA_RECONSTRUCT
mkdir -p "$RESULTS_FOLDER"
# One family
python "predict_ancestral_hierarchical_approach.py" \
--model_path $MODEL_PATH \
--fasta_files_folder "$BETA_RECONSTRUCT/data/mammalian_validation/2331_NT_AL_AA/" \
--results_folder $RESULTS_FOLDER \
--species_list_folder $SPECIES_LIST_FOLDER
# All families (283297 and 5535 are skipped: no outgroup)
python "predict_ancestral_hierarchical_approach.py" \
--model_path $MODEL_PATH \
--fasta_files_folder "$BETA_RECONSTRUCT/data/mammalian_validation/" \
--results_folder $RESULTS_FOLDER \
--species_list_folder $SPECIES_LIST_FOLDER
Each family gives results/mammalian/<family>.json. The reconstructed placental ancestor is in its prediction field:
python -c "import json; print(json.load(open('results/mammalian/2331_NT_AL_AA.json'))['prediction'])"
Edo Dotan, Elya Wygoda, Asaf Schers, Iris Lyubman, Yonatan Belinkov and Tal Pupko, Ancestral sequence reconstruction using generative model, bioRxiv 2026; https://www.biorxiv.org/content/early/2026/01/21/2026.01.18.700141
@article{dotan_ancestral_2026,
title = {Ancestral sequence reconstruction using generative models},
url = {https://www.biorxiv.org/content/early/2026/01/21/2026.01.18.700141},
doi = {10.64898/2026.01.18.700141},
abstract = {Ancestral sequence reconstruction (ASR) is a foundational task in evolutionary biology, providing insights into the molecular past and guiding studies of protein function and adaptation. Conventional ASR methods rely on a multiple sequence alignment (MSA), a phylogenetic tree, and an evolutionary model. However, the underlying alignments and trees are often uncertain, and existing models typically focus on substitutions and do not explicitly account for insertion-deletion (indel) processes. Here, we introduce BetaReconstruct, a novel generative approach to ASR that harnesses recent advances in natural language processing (NLP) and hybrid transformer architectures. Our model was initially trained on large-scale simulated datasets with gold-standard ancestral sequences and subsequently on real-world protein sequences. The reconstruction requires neither MSAs nor phylogenetic trees. We demonstrate that BetaReconstruct generalizes robustly across diverse evolutionary scenarios and reconstructs ancestral sequences more accurately than maximum-likelihood-based pipelines. We additionally provide evidence that the generative-model ASR approach is also more accurate when analyzing empirical datasets. This work provides a scalable, alignment-free strategy for ASR and highlights the ability of data-driven models to capture evolutionary signals beyond the reach of traditional methods.},
journal = {bioRxiv},
author = {Dotan, Edo and Wygoda, Elya and Schers, Asaf and Lyubman, Iris and Belinkov, Yonatan and Pupko, Tal},
year = {2026},
note = {Publisher: Cold Spring Harbor Laboratory
\_eprint: https://www.biorxiv.org/content/early/2026/01/21/2026.01.18.700141.full.pdf},
}