ALICE

A gene-therapy vector has to succeed at several things simultaneously: it must package efficiently during production, bind the receptor that carries it across the blood-brain barrier, and reach the right cells once inside. Engineering campaigns usually optimize one of those at a time, and a generative model trained to imitate strong receptor binders learns exactly that one trait — the authors show that a peptide language model tuned on high-binding capsid inserts produces sequences whose production fitness collapses.

ALICE (AI-driven Ligand-Informed Capsid Engineering) separates generation from optimization to address this. A RoBERTa language model is pretrained on 2.89 million UniProt peptide sequences, then transferred through a SeqGAN stage onto 72,753 AAV9 capsid insert variants so that it writes plausible capsid peptides. A separate module, function-guided evolution (FE), then takes the generator's top candidates and evolves the sequences — not the model weights — using a contrastive objective computed against two small predefined reference sets: 76 sequences with high viability and strong Ly6a/Ly6c1 binding, and 53 with weak binding and low viability. Because the references and the property predictors are fixed and only the population under evolution changes, a new design objective is expressed as new references rather than a retrained generator.

The work comes from Xuhua Wang's group at Zhejiang University, with collaborators at Nantong University. The first preprint appeared in October 2024, a second version added ALICE-X in February 2026, and the peer-reviewed article followed in Cell Press Blue in June 2026. The design target is deliberately narrow — the 7-mer insert between residues 588 and 589 of the AAV9 VP1 capsid protein — unlike AAVDiffusion, which denoises whole capsid sequences.

#Key Features

• Function-guided evolution: A heuristic search (genetic algorithm, estimation of distribution, simulated annealing, or a hybrid) scores each evolving sequence by alignment-based similarity to a positive reference set and distance from a negative one, fusing several functions into a single insert.
• Works from tiny labeled sets: The multi-function objective is defined by 129 reference sequences, with the large capsid libraries used only to fit the property predictors.
• Curriculum learning in ALICE-X: A two-phase upgrade first integrates functions against the references, then adds novelty and anti-convergence terms so the search explores sequence space the wet-lab data never covered.
• Frozen property oracles: Gradient-boosting regressors, chosen over six alternatives including SVR, XGBoost and random forests, score production fitness and receptor binding at every generation.
• Wet-lab validated designs: Top candidates were packaged and assayed in mice, closing the loop from in silico design to measured transduction.

#Technical Details

An ablation isolates each stage: SeqGAN alone failed to improve production fitness, BERT-SeqGAN improved fitness at the cost of Ly6a/Ly6c1 binding, and RoBERTa-SeqGAN improved binding while losing fitness — only the full RoBERTa + SeqGAN + ranking-filtration + FE stack scored highest. The evaluation models were trained on 119,851 sequences for production fitness, 89,169 for Ly6a binding and 89,040 for Ly6c1 binding, reaching Pearson correlations of 0.96, 0.81 and 0.81. The whole system holds roughly 51 million parameters and trains in 7.25 GPU hours on a single RTX 3090. Of eight ALICE designs tested in BALB/cJ mice, AAV.ALICE-N2 showed a ~2-fold titer increase and 372-fold higher brain transduction than AAV9, with AAV.ALICE-N6 at 64-fold. Retargeting to human transferrin receptor 1 required ALICE-X plus a new evaluation model fit on 6,387 sequences; the resulting AAV.ALICE-H3 carries a 'YTK' motif absent from the training data, with ~1.5-fold better viability than AAV9 and ~251-fold greater brain-cell selectivity in human TFRC knock-in mice than in wild-type mice.

#Applications

The immediate use is designing AAV capsid variants for central nervous system gene delivery without running a full in vivo selection round for each objective. Because the target is encoded in a reference set of a few dozen sequences, groups holding modest in-house screening data can pose objectives the published predictors were never trained for. The released code spans pretraining, semantic tuning, evolution and ranking, so the pipeline can be re-pointed at another receptor given a predictor for it.

#Impact

ALICE addresses the multi-function fusion problem that limits generative capsid design, and its wet-lab results are unusually concrete for the area: designs that independently rediscovered the sequence neighborhoods of previously reported CNS-tropic capsids, plus a human-receptor-targeting variant carrying a motif with no precedent in the training library. The limitations are equally concrete. Only the 588/589 7-mer insert of one serotype is designed, so the capsid scaffold is fixed; moving to a new receptor required training a new property predictor even though the generative core carried over; and the reported gains are measured against wild-type AAV9 rather than engineered clinical benchmarks. Code, the pretrained checkpoint and the training data are publicly downloadable; the repository carries an MIT license file, while its own documentation restricts use to non-commercial academic purposes.