Joint-embedding world model of protein organization, trained on nanoscale proximity measured in intact tissue by expansion microscopy.
No providers recorded yet. Browse all providers
Two proteins can only interact if they sit in the same place, and in intact tissue which molecules share a nanoscale neighborhood is a physical fact that neither sequence nor a solved structure records. Amyloid-β does not change the sequence of an AMPA receptor; it changes what that receptor sits next to. Models trained on sequence, structure or expression therefore have no channel through which to see the rearrangements a disease makes at this level.
HI-JEPA, developed by Eratos Therapeutics with Manolis Kellis of MIT CSAIL and posted to bioRxiv in August 2026, is trained on that missing quantity. Eratos' ASCEND platform measures pairwise molecular proximity from expansion microscopy, and HI-JEPA learns one embedding per protein, trained so that each protein's vector predicts the vectors of its measured neighbors in latent space. It never reconstructs its input and uses no generated negatives. A set of proteins measured together forms a configuration, which the model treats as a world-model state it can perturb, score and plan over.
The paper is presented as a capability report: it gives the parameter count, the latent dimensions and the training objective, but the architecture, the training procedure and the acquisition protocol are otherwise proprietary and not described, while the evaluation protocol behind every reported number is. It is a preprint and has not been peer reviewed.
HI-JEPA holds 38.9M parameters, with each modality encoded to a 768-dimensional latent and read-out heads projecting to 384 dimensions. Its vocabulary spans 13,447 proteins, split once into 12,468 training and 979 held-out proteins, with no training edge touching a held-out protein across nine relation types and 1,633,725 edges. Measured proximity from ASCEND covers 164 proteins across 37 imaged regions in five studies (cultured neurons, isolated synapses and 5xFAD and wild-type mouse cortex) and contributes 0.04% of training edges; multiplex spatial imaging supplies 43.3% and curated references (interaction, pathway, disease, complex, structure, perturbation and bioactivity) 56.7%.
On high-confidence physical interactions between sequence-dissimilar proteins, with both proteins withheld from training, HI-JEPA reaches AUC 0.908 against 0.514 for ESM-C 6B under partner-count-matched negatives; co-folding with ESMFold2 on a subsample scores 0.538. It identifies proteins from point clouds against 234,048 deposited structures at top-1 0.748, recovers a held-out complex member in the top 100 of 13,447 candidates at recall 0.954, and recovers knockout-disrupted partners at recall@100 0.640, collapsing to 0.028 when ranked against the wrong protein. Ablating measured proximity at training time degrades cross-scale partner recovery from median rank 14 to 68 while navigation and within-scale dynamics hold.
A companion Eratos preprint reports cryptic-pocket detection from single apo structures, reading a per-residue latent that it attributes to the same world model. On 231 held-out CryptoBench apo chains it localizes the cryptic site at top-1 0.848 with residue-level AUC 0.8465, and on WRN helicase, held out of training, it ranks the HRO761 allosteric pocket first. It also recovers 119 sites that OpenDDE misses on 190 held-out targets.
Eratos positions HI-JEPA for target nomination in diseases where sequence and structure offer no starting point, with neurodegeneration a primary focus. Given 5xFAD mouse cortex with no disease label, ranking 1,574 measured assemblies by departure from wild type places amyloid-β bound to AMPA receptor subunits in nine of the top ten. Planning over the same configurations selects GluA2, GluA3 and GluA4 for disruption and predicts that disrupting the PSD-95 scaffold worsens the state, both matching published experiments the model never saw. Neither the code, the weights nor the proprietary corpus is released, so the model is used inside Eratos' own discovery pipeline rather than by outside groups.
HI-JEPA carries the joint-embedding predictive framework into a new kind of biological data: directly measured spatial organization of proteins rather than sequences, images or expression profiles. Its central argument, supported by the ablation, is that cross-scale reasoning comes from the proximity measurement itself rather than from more data, which makes measurement throughput rather than model size the limiting factor. Because the architecture and training procedure are largely undisclosed and nothing is released, the results cannot yet be independently reproduced, and validation on human expansion-microscopy tissue remains future work.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.