Antibody sequence editor that proposes insertions, deletions and substitutions as timed jump events, giving lead optimization a tunable edit budget.
No providers recorded yet. Browse all providers
Antibody lead optimization is a budgeting problem. A candidate already works well enough to be worth keeping, and the job is to spend a handful of changes on it — remove a liability, improve expression — without disturbing the paratope. Most generative architectures cannot express that budget. A masked protein language model such as ESM-2 resamples residues at positions the user supplies and returns a sequence of the same length, so where and how many are inputs rather than outputs. Antibody infilling models such as IgLM regenerate a whole span left to right, leaving the number of changes inside it uncontrolled. Discrete diffusion models such as EvoDiff fix the length before decoding.
EditJumps, released by Visium in September 2026, treats editing itself as the generative process. Insertions, deletions and substitutions fire one at a time, at learned per-position rates, in continuous time — a continuous-time Markov jump process over sequences, which the paper identifies as the pure-jump case of generator matching. That observation unifies two prior methods, Edit Flows and EvoFlows, neither of which released code or complete training specifications, and EditJumps is the first open implementation of the framework.
Its practical contribution is generalization: where the original approaches retrain per target family, EditJumps trains one editor on 1.66M Observed Antibody Space homolog pairs and edits unseen leads zero-shot.
VH.VL, it runs on new seed sequences without per-family fine-tuning.The editor is an unfrozen pretrained ESM-2 encoder with five prediction heads —
per-position insertion, deletion and substitution rate heads plus insertion and
substitution token heads — conditioned on time by FiLM modulation. The reported
configuration starts from the stock facebook/esm2_t12_35M_UR50D trunk, totals roughly
35M parameters, and is fine-tuned for 20,000 steps at batch size 16 across three seeds.
Scaling that trunk to 650M moves realized edit scale only from 0.180 to 0.182 — inside the
±0.01 seed variance — at far greater compute cost. Training pairs are unordered OAS
homolog pairs clustered with MMseqs2 at 50% identity and 0.8 coverage.
Evaluation follows EvoFlows' protocol on two antibody targets, the anti-SARS-CoV-2 nanobody Ty1 and the anti-HER2 heavy chain HER2-VH, each with 20 templates and 200 held-out homologs. At budgets of 4.1–4.6 edits per sequence EditJumps reaches greater sequence diversity than evotuning baselines and the alignment-based diffusion model EvoDiff-MSA while matching unnormalized spectrum MMD on Ty1 and landing within 3% on HER2-VH; all learned methods stay separated from the natural homolog floor (0.98 versus 0.60 MMD). Two replication findings accompany those numbers: published edit statistics reconciled only after recovering an unstated rate-scaling hyperparameter, and six of the ten metrics reported in the original work cannot be deterministically recomputed from its text. Expanding the reference alignment from 200 to 800 homologs shifts normalized spectrum MMD by 50% on Ty1 and 26% on HER2-VH, enough to invert method rankings.
The intended use is proposing experimentally testable neighbours of an existing antibody or protein lead: a handful of variants per seed, with the number of changes under the user's control and the changes placed where related natural sequences differ. Generation is unconditional, so it returns variants and edit distances but no fitness score; ranking needs a separate property model, a natural holdout from the target family, or a wet-lab assay. A companion command ranks candidates by ESM-2 pseudo-log-likelihood.
EditJumps matters less as a leaderboard result than as the open reference for a modeling idea that had none. By showing Edit Flows and EvoFlows to be the same pure-jump process and releasing code, configurations and metrics, it makes edit-based generation reproducible and supplies the regularity conditions — non-explosion and clock normalization — that variable-length sequence spaces require. Its evaluation analysis carries further than the model itself: metrics normalized against a reference set flip rankings with the size of that set, a caution for any in-silico protein generation benchmark. The work is a preprint awaiting peer review, the code is MIT-licensed, evaluation is strictly in silico, and the trained editor checkpoints are held in private cloud storage rather than released.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.