CoSiNE

CoSiNE webserver overview

The CoSiNE webserver scores paired-antibody deep mutational scanning (DMS) variants with a neural continuous-time Markov chain. The available variant effect prediction (VEP) workflow returns CoSiNE log-likelihoods, perplexities, correlations with submitted experimental fitness, and optional Thrifty-corrected selection scores.

The interface also includes CoSiNE's unguided and predictor-guided antibody evolution controls. Evolution jobs are currently unavailable because the source project has not published the required RefineGNN binding-predictor checkpoint, so choose Variant effect prediction for a runnable job.

Pricing

A CoSiNE submission currently has a fixed quote of 100 credits. The exact quote is shown before submission. Input size, scoring axes, branch lengths, and other settings do not currently change this quote.

Inputs

Variant effect prediction

The default DMS column names are heavy, light, and fitness. Map different column names in DMS column mapping. CoSiNE derives the wild-type heavy and light consensus site by site unless you provide explicit consensus sequences.

Preparing the DMS table

Column names are case-sensitive, so a header of Heavy does not match the default mapping heavy. Every row needs a non-empty heavy sequence, light sequence, and fitness value, and the fitness value must be numeric. One blank cell or an entry such as ND rejects the whole table rather than skipping that row, so remove or fill such rows before upload. A row with more fields than the header is also rejected, so quote any value that contains a comma.

Within each chain column, all sequences must have the same length. Trim or align every sequence to the same variable-domain span first, and use only the 20 canonical amino-acid letters, because residue codes outside that alphabet cannot be embedded by the model.

Coding DNA for Thrifty correction must be uppercase A, C, G, and T, a single record, and made of complete codons. Lowercase output from some sequence viewers is rejected, so convert it first. If the translation does not match the consensus, check that the DNA covers exactly the residues in the table, with no signal peptide or constant-region codons, and that an explicit consensus matches the same span.

Antibody evolution

Antibody evolution is not currently executable because its required source checkpoint is unpublished.

Settings

Task

Variant effect settings

DMS column mapping

Antibody evolution settings

These settings reproduce the source evolution interface, but evolution jobs remain unavailable until the required binding-predictor checkpoint can be verified.

Outputs

Variant effect prediction

Antibody evolution

When evolution becomes available, its interface returns the native evolution table as cosine_evolution.source.csv, a derived cosine_evolved_sequences.fasta, and cosine_source.log. Planned table fields include the seed and sampled sequences, branch length, guidance type, mutation count, trajectory index, timing, predictor fitness, humanness metrics, and edit distance from the seed. No evolution result is produced while the required source checkpoint is unavailable.

Understanding results

• mutation identifies the amino-acid substitutions scored on that axis. Combined labels join heavy and light mutations with |.
• ll_<t> is the summed CoSiNE log-likelihood for the wild-type-to-variant transition at branch length t. Within the same axis, context, and branch length, a higher value means the model assigns the transition greater likelihood.
• ppl_<t> is exp(-ll / sequence length). Lower values indicate lower per-residue surprise. Because sequence length and context affect the calculation, compare perplexity within a consistent scoring setup.
• corrected_ll_<t>_<bl> is the CoSiNE log-likelihood minus the Thrifty neutral SHM log-likelihood. It is a model-derived selection score, not a measured affinity or binding constant.
• correlation is the selected Spearman or Pearson association between one score column and the submitted fitness values. Its sign depends on the metric: likelihood-based scores and perplexity point in opposite numerical directions.
• p_value and n_samples describe that correlation calculation. They do not provide a universal model confidence score.

Use CoSiNE scores to prioritize and compare variants within a controlled assay and scoring protocol. Confirm candidates with binding, expression, developability, and other relevant experiments.

Limits of the score columns

CoSiNE runs zero-shot, so nothing is fitted to the submitted assay, and a correlation measures how well one setup happens to agree with one table. A correlation needs at least two non-missing samples, and values from a small library move substantially with a few rows. A single p_value does not validate the workflow, because each branch length, axis, and metric produces another correlation to choose from.

The combined axis scores the joint heavy-light transition directly, so its values are not the sum of the heavy and light values. Axes, paired-context choices, and branch lengths also change the scale, so rank variants within one fixed setup rather than comparing raw magnitudes across files.

For a perplexity column, a negative correlation with fitness is the expected direction when higher fitness is better, because lower perplexity marks a more likely transition. For ll_<t> and corrected_ll_<t>_<bl>, the expected direction is positive.

Validating results

Before interpreting scores, confirm that heavy and light sequences are assigned to the correct row, that the chain orientation and domain boundaries are right, and that the fitness values are normalized consistently across replicates. A shifted row or swapped chain can produce a convincing but meaningless correlation.

To test sensitivity, request a small set of branch lengths in one job, such as 0.1,0.2,0.4, and compare the paired and single-chain context settings. The quote does not change with the number of branch lengths, so this costs nothing extra. Decide which setup to report before looking at the correlations, and prefer variant rankings that stay stable across the checks over one optimized coefficient.

A correlation on the library used to choose settings is not evidence of predictive performance. Use held-out variants or an independent library for that, then retest shortlisted variants experimentally.

Sources

• Lu SZ, Vermani A, Sanno K, Lu J, Matsen FA IV, Jagota M, Song YS. "Conditionally Site-Independent Neural Evolution of Antibody Sequences." arXiv (2026). arxiv.org/abs/2602.18982