CuMMI

When a nanoparticle enters blood, plasma, or any other biofluid, proteins adsorb onto its surface within seconds and form a "protein corona" — the layer that gives the particle its biological identity and determines where it goes. Measuring that corona is slow, expensive, and notoriously sensitive to protocol: isolation method, incubation time, and proteomic instrument all shift the answer. Machine learning has been applied to the problem since 2018, but earlier models were fitted per-dataset — a random forest for silver nanoparticles, regressors for one panel of proteins, a graph network for gold nanoclusters — and none transferred to a nanomaterial or a protein they had not seen.

CuMMI (curriculum-guided multimodal interaction model) attacks that generalization gap directly. Developed by researchers at Westlake University and University College Dublin and posted as a preprint in July 2025, it predicts nanomaterial-protein interactions (NPI) by combining pretrained protein representations with a text encoding of the full experimental context, then training under a curriculum that starts in human plasma and progressively widens to more distant biofluids.

It is paired with the largest curated NPI dataset assembled to date: 1.97 million samples covering 37,392 proteins, distilled from 207 studies after screening more than 2,500 articles. That dataset, rather than a new architecture, is the work's center of gravity — it is what makes held-out-nanomaterial and held-out-protein evaluation possible at all.

#Key Features

• Curriculum over biofluids: Training proceeds through five stages that begin on typical human plasma and progressively add human serum, atypical plasma, non-human blood, and non-blood biofluids before an in-domain refinement pass, rather than pooling everything at once.
• Multimodal protein plus context fusion: Protein sequence and structure embeddings are combined by a learned gate, then fused with a text-encoded description of 37 experimental variables through cross-attention, so the same protein can be scored differently under different materials and protocols.
• Quality-weighted training: Every sample carries a composite score built from five reliability indicators, which scales its contribution to the loss so that sparsely reported literature entries are used without dominating.
• Independence-preserving external validation: Performance is reported on temporal, nanomaterial-held-out, and protein-held-out splits, each isolating a different axis of generalization instead of a single random split.
• Ablation-based explainability: Systematic removal of modalities, single features, and feature pairs identifies nanomaterial core composition and study research purpose as the two most influential tabular variables.

#Technical Details

Protein sequences are embedded with ESM-2 and structures with its companion folding model ESMFold; the 37 tabular features are rendered into structured natural-language prompts and embedded with Linq-Embed-Mistral. All three representations are projected into a shared 1024-dimensional space, sequence and structure are merged by a learned gate, and the resulting protein vector interacts with the context embedding through an 8-head cross-attention block (text as query, protein as key and value) followed by a residual update and an MLP head producing a binary interaction call. Embeddings for unique sequences and context strings are precomputed, so training touches only the fusion stack; class imbalance is handled by up-weighting positives.

The dataset spans four biofluid source domains and six feature groups — nanomaterial properties (14 features), protein source (3), incubation conditions (6), separation parameters (5), proteomic settings (8), and research purpose (1). On the internal human-plasma test set, the multimodal model averages 0.92 across AUROC, AUPRC, recall, precision, and accuracy, versus 0.71 for a protein-only and 0.70 for a context-only ablation. On all three external splits the five-metric mean exceeds 0.75, with AUROC and AUPRC above 0.7; the curriculum itself accounts for much of that, raising temporal-split AUROC from roughly 0.58 under conventional single-domain training to 0.71. Withholding all gold-nanoparticle samples and 2,359 protein accessions from pretraining, then fine-tuning only the prediction head, beats training from scratch at every training-set fraction, with the margin widening as data become scarce.

#Applications

The model targets in vitro nano-bio research: screening which proteins a candidate nanomaterial will enrich before running proteomics, prioritizing surface chemistries for biomarker-capture assays, and designing particles whose corona favors a desired target. Because predictions are conditioned on incubation, separation, and proteomic settings, it can also be used to reason about how much of a reported corona reflects protocol rather than material.

#Impact

CuMMI is a preprint awaiting peer review, and its practical value is bounded by an honest limitation the authors state plainly: the dataset captures the post-interaction endpoint — the relatively stable hard corona — and says nothing about the dynamic, non-equilibrium adsorption and exchange kinetics that govern in vivo behavior. Code for embedding, training, explainability, and fine-tuning is on GitHub under a research-use-only restriction with no license file, the curated dataset is on figshare, and no trained weights have been released. Its more durable contribution may be the corpus itself, which turns a fragmented literature on protein corona measurements into a benchmark where generalization to unseen materials and unseen proteins can actually be scored.