“Large language model” means two quite different things in biology, and conflating them is the source of most confusion about what this technology can do.
The first is a biological sequence model — ESM, ProtBert, Evo, Nucleotide Transformer — trained on protein or genomic sequence rather than human text. The second is a general text model such as GPT or Claude, applied to scientific literature and research workflows. They are built on similar architectures and are useful for almost entirely different things.
Protein language models: the useful half
A protein language model is trained by masking residues in millions of protein sequences and learning to predict them. To do that well it has to internalise which substitutions are tolerated where, which residues co-vary because they contact each other in the folded structure, and which positions are functionally constrained.
Nobody tells it about structure or function. Both fall out of the statistics of evolution, because the sequences it trains on are the ones that survived selection.
Embeddings are the workhorse
The practical output is an embedding: a numeric vector per residue or per protein capturing what the model has learned. These are genuinely useful, and this is where most real value sits.
Take a few hundred labelled proteins, embed them, and train a small classifier on those vectors. This routinely outperforms models built on hand-engineered features, and works with training set sizes that would be hopeless for a model learning from raw sequence. Typical applications include subcellular localisation, enzyme class prediction, thermostability, solubility and binding-site identification.
The reason it works is transfer: the expensive representation learning was done once, on hundreds of millions of sequences, and your small dataset only has to learn the mapping from that representation to your label.
Variant effect prediction
Because these models assign a likelihood to every residue at every position, they score how surprising a mutation is given everything the model has learned about that protein family. Low-likelihood substitutions correlate with deleterious effects.
This is zero-shot — no training on variant data at all — and it is competitive with established variant effect predictors. For a protein with no experimental mutational data, it is often the best available signal. It should not be read as a clinical determination: correlation with pathogenicity is good in aggregate and unreliable for any individual variant.
Structure prediction without alignments
ESMFold predicts structure directly from a single sequence, skipping the multiple sequence alignment that AlphaFold2 depends on. It is generally somewhat less accurate but dramatically faster, which is what makes proteome-scale prediction feasible. For orphan proteins with few homologues, where alignments are thin, the gap narrows considerably.
Text models: useful, and easy to misuse
General LLMs applied to biomedical literature are valuable for a narrower set of tasks than the enthusiasm suggests.
Where they work well: extracting structured information from unstructured text — pulling gene-disease relationships, experimental conditions or reported effect sizes out of thousands of papers into a table. Named entity recognition and normalisation to ontology identifiers. Drafting and refactoring analysis code. Summarising a body of literature you will then read properly.
Retrieval-augmented generation is the pattern that makes these systems trustworthy enough to deploy. Rather than asking the model to recall facts, you retrieve relevant passages from a controlled corpus — your papers, your protocols, your internal documents — and ask the model to answer using those passages, with citations. The answer is grounded in text you can check.
Where they fail: anything relying on recall of specific facts. Fabricated citations remain a real failure mode, and plausible-sounding but wrong gene names, accession numbers and statistics are common. A model that does not know it is wrong is dangerous in exactly the situations where you most want help.
A test worth applying
Before building anything on an LLM, ask: if the model is confidently wrong, will anyone notice?
For code generation, yes — it fails to run, or the tests catch it. For extraction with citations, yes — you can check the source passage. For embeddings feeding a classifier, the question barely applies, because you evaluate on held-out data like any other model.
For “summarise the state of research on X” or “which gene causes Y”, often no. Those are the applications that generate impressive demonstrations and unreliable science.
Practical guidance
- Start with embeddings, not generation. Most real value in biological LLM work comes from representations feeding conventional models, not from text output.
- Fine-tune only with enough data. Below a few thousand labelled examples, a classifier on frozen embeddings usually beats fine-tuning and is far cheaper.
- Benchmark against a simple baseline. BLAST-based homology transfer, or a model on classical features, is often competitive. If the language model does not beat it, use the simpler method.
- Split data by homology, not randomly. This is the most common evaluation error in the field. Random splits put close homologues in both training and test sets, and the reported accuracy is inflated. Cluster by sequence identity and split at the cluster level.
- Ground every generative claim. If a system produces text a researcher will act on, it needs retrieval and citations, not recall.
What has actually changed
The realistic summary is that protein language models have made a class of prediction problems tractable with small datasets, and text models have made literature and code work faster. Neither has replaced experiments or domain knowledge, and neither is close to doing so.
The teams getting the most from this technology are not the ones building chat interfaces. They are the ones quietly using embeddings to make a 300-sample dataset behave like a 30,000-sample one.
