Biological models

318 model records in release 2026-10-23-5347695500ed (release manifest, gzip). Showing 289–312; page 13 of 14.

Evaluated configurations, methods and pipelines are listed separately. Names alone do not establish equivalent models or checkpoints.

• SpliceBERT

  DNA and genomes

  BERT-style RNA language model trained by masked token prediction on precursor RNA from 72 vertebrate species, used for splice site, branchpoint and splicing variant effect prediction.

  Record: lit-genomics-20261011-geneb-model-splicebert

• SpTransformer

  DNA and genomes

  Task-specific transformer for tissue-specific splice site prediction.

• STATE

  Single-cell omics

  State separates cellular representation learning from prediction of responses to perturbation.

• SweetNet

  Glycans

  SweetNet predicts glycan properties and produces learned representations from glycan graphs.

• TAPE Bepler

  Protein function

  The TAPE Bepler comparison uses a protein representation that combines bidirectional language modelling with supervised structural pretraining.

• TAPE LSTM

  Protein function

  The TAPE LSTM baseline represents protein sequences using recurrent networks that read residues in both directions.

• TAPE ResNet

  Protein function

  The TAPE ResNet baseline represents protein sequences using residual convolutional blocks before a task-specific prediction head.

• TAPE Transformer

  Protein function

  The TAPE Transformer learns contextual protein representations using masked-residue pretraining and a task-specific prediction head.

• TAPE Unirep

  Protein function

  The TAPE UniRep baseline uses a multiplicative recurrent network to represent protein sequences for downstream tasks.

• text-embedding-3-large

  Cells and tissues

  OpenAI text embedding model that maps a text string to a 3,072-dimensional vector.

• TranceptEVE

  Proteins and complexes

  Hybrid of a family-agnostic autoregressive transformer and family-specific alignment-based models that relies more on the alignment models as alignment depth grows, scored in ProteinGym as a zero-shot substitution baseline.

• Tranception

  Proteins and complexes

  Autoregressive transformer protein language model that can retrieve homologous sequences at inference to predict the fitness of substitutions, multiple mutants and indels; ProteinGym scores it with and without retrieval as a zero-shot substitution baseline.

• Transformer

  Protein function

  Transformer encoder trained from scratch on the task's sequences.

• TREDNet

  DNA and genomes

  Two-phase CNN for enhancer prediction and variant prioritisation, developed by the group of the senior author of the comparison that evaluated it.

• trRosettaRNA

  RNA and transcriptomes · Protein structure

  trRosettaRNA: MSA and pair representations fed into the RNAformer transformer, followed by 3D folding.

• UCE (Universal Cell Embeddings)

  Cells and tissues

  Single-cell foundation model producing cell embeddings from protein-embedded gene tokens.

• Uni-RNA

  RNA and transcriptomes

  Uni-RNA RNA foundation model (L8, L12 and L16 checkpoints).

• Unirep

  Proteins and complexes

  Deep learning representation of protein sequences learned from unlabeled amino-acid sequences, on which simple models predict stability and mutant function, scored in ProteinGym as a zero-shot single-sequence substitution baseline.

• Unirep evotuned

  Proteins and complexes

  Evo-tuned variant of UniRep, a protein sequence representation learned from unlabeled amino-acid sequences, which ProteinGym lists as using a multiple sequence alignment for evo-tuning and scores as a zero-shot substitution baseline.

• UniSpec

  Proteins and complexes

  UniSpec attention-based fragment dictionary model (about 12M parameters).

• UTR-LM

  DNA and genomes

  Language model pre-trained on 5' untranslated region sequence.

• VenusREM

  Proteins and complexes

  Retrieval-enhanced protein language model that captures local amino acid interactions for mutation effect prediction, scored in ProteinGym as a zero-shot substitution baseline from alignment and structure inputs.

• VESPA

  Proteins and complexes

  Logistic regression ensemble that predicts single amino acid variant effects without alignments from protein language model embeddings, combining predicted conservation, BLOSUM62 scores and language-model mask reconstruction probabilities; scored in ProteinGym as a zero-shot substitution baseline.

• VespaG

  Proteins and complexes

  Missense variant effect predictor that feeds protein language model embeddings to a minimal deep learning model trained on GEMME predictions for 39 million human variants, scored in ProteinGym as a substitution baseline.

This index reflects a dated catalogue, not an exhaustive census. Source checking does not mean independent reproduction; compare results only under compatible protocols, datasets and metrics.