AbLang-2

What is AbLang-2?

AbLang-2 is an antibody-specific language model designed to predict non-germline (NGL) residues in antibody sequences. Developed by Tobias Olsen and colleagues at the Oxford Protein Informatics Group, it addresses a fundamental limitation of earlier antibody language models: their tendency to predict germline residues while ignoring the somatic mutations that make antibodies functional.

Antibody sequences are heavily biased toward their germline templates. During somatic hypermutation, only a small fraction of residues mutate to create binding specificity against a target. When trained on natural antibody repertoires, models like AbLang-1, AntiBERTy, and Sapiens learn to predict germline residues with overwhelming probability (84-87%) because that's what most residues are.

How does AbLang-2 work?

The model architecture consists of 12 transformer layers with 480-dimensional embeddings (45 million parameters). Training proceeds in two phases: pre-training on 35.6 million unpaired VH and VL sequences from the Observed Antibody Space (OAS) database, followed by fine-tuning on 1.26 million paired heavy/light chain sequences.

The key innovation is the use of focal loss, which down-weights the loss contribution from well-predicted labels. Since germline residues are easy to predict, focal loss reduces their influence on the gradient and forces the model to learn NGL patterns. This improved NGL perplexity from the 14-39 range seen in prior models down to 10-12, achieving 15% cumulative probability for known NGL residues compared to less than 2% for earlier models.

AbLang-2 outputs two types of representations:

• Pseudo-likelihoods: Per-sequence pseudo-log-likelihood scores calculated by masking each residue in turn and averaging its log probability. Higher scores indicate a better match to the learned antibody distribution.
• Embeddings: 480-dimensional vectors encoding the sequence context, suitable as features for downstream machine learning tasks.

How to use AbLang-2 online

ProteinIQ provides browser-based access to AbLang-2 with no software installation or command-line usage required. Any registered user can run the model directly from the web interface.

Input

Input	Description
Antibody sequences	FASTA format. For paired mode, join heavy and light chains with a pipe: VH_sequence|VL_sequence. For unpaired mode, provide individual chains of the selected type with descriptive headers.

Settings

Setting	Description
Sequence mode	Paired (VH|VL) expects pipe-separated heavy and light chains. Unpaired processes individual chains separately.
Unpaired chain type	Select whether each unpaired FASTA record is a heavy (VH) or light (VL) chain.
Output type	Pseudo-likelihoods returns per-sequence scores. Sequence embeddings returns 480-dimensional vectors. Both returns both outputs.

Output

Results appear as a spreadsheet with columns depending on the selected output type:

Column	Description
Sequence ID	Identifier from FASTA header
Sequence	Input sequence
Length	Number of residues
Pseudo-likelihood	Score reflecting how well the sequence fits the antibody language model
Embedding dim	Dimensionality of the embedding vector (480)

Use the Full results tab to view or download JSON with complete sequences, unrounded scores, and all 480 embedding values when requested. Scored results also include likelihood_logits, preserving the single-pass mutation-logit matrix, its shape, and the native vocabulary labels for its columns. These matrices include the model's start, stop and chain-separation tokens. The Results table download contains the displayed columns, including sequence previews and embedding dimensions. Embedding-only runs omit the score column.

Interpreting pseudo-likelihoods

The Pseudo-likelihood column contains the native pseudo-log-likelihood score: the mean log probability of the observed residues when each residue is masked in turn. Scores are dimensionless and normally negative; higher, less negative scores indicate a better match to the model. This calculation is slower than the single-pass mutation logits. There is no universal threshold for good or bad scores, but relative comparisons are meaningful:

• Sequences with unusually low scores may contain problematic residues or frameshifts
• Comparing variants of the same antibody can identify mutations that disrupt antibody-like properties
• Tracking scores across humanization or optimization rounds can guide design decisions

Historical results keep their original values. Older scored runs used an average of raw mutation logits, which is a different calculation and should not be compared numerically with the masked pseudo-log-likelihood scores. Older embedding-only tables may also retain a zero in the score column even though no score was computed.

Limitations

AbLang-2 excels at predicting NGL residues but does not guarantee that suggested mutations will improve binding or stability. The model reflects statistical patterns in natural antibodies, not experimental binding data.

Paired mode requires both chains. If only one chain is available, use unpaired mode with the appropriate chain type.