NanoPro-3M: Nanomaterial-Protein Dataset

• NanoPro-3M is a comprehensive nanomaterial–protein interaction dataset with over 3.2M samples and 37K proteins sourced from 2,500+ articles.
• The resource integrates multimodal data by combining structured descriptors with protein embeddings for accurate affinity prediction via the NanoProFormer architecture.
• Empirical evaluations show that multimodal fusion outperforms single-modality approaches, enabling robust zero-shot inference and fine-tuning for downstream nano-bio applications.

Searching arXiv for the specified paper to ground the article. {"query":"arXiv (Yu et al., 18 Jul 2025) NanoPro-3M A million-scale dataset and generalizable foundation model for nanomaterial-protein interactions","max_results":5} NanoPro-3M is a million-scale nanomaterial–protein interaction resource introduced with NanoProFormer in the paper "A million-scale dataset and generalizable foundation model for nanomaterial-protein interactions" (Yu et al., 18 Jul 2025). It is described as the largest nanomaterial-protein interaction dataset to date, comprising over 3.2 million samples and 37,000 unique proteins sourced from UniProt, with data aggregated from more than 2,500 peer-reviewed articles identified through Web of Science, PubMed, and Scopus searches through Sept 20 2024. The dataset and model target affinity prediction for nanomaterial–protein pairs, with outputs formulated either as continuous relative protein abundance (RPA) or as binary interaction labels. The work positions multimodal representation learning as a route to stronger generalization, including missing features and unseen nanomaterials or proteins, and frames the resource as a foundation for reducing experimental reliance in nano-bio interaction studies (Yu et al., 18 Jul 2025).

1. Dataset scope and compositional structure

NanoPro-3M aggregates more than 3,200,000 nanomaterial–protein interaction pairs and more than 37,000 unique proteins. The source corpus consists of more than 2,500 peer-reviewed articles, and the dataset construction includes duplicate removal by DOI as well as exclusion of short or low-citation papers (Yu et al., 18 Jul 2025). This scale is central to the project’s stated objective: broad coverage of nanomaterial chemistries, assay conditions, and proteomic outputs.

For the filled-only dataset, the nanomaterial core-type distribution is explicitly reported as follows.

Core type	Share
Metal oxide-based	30.8 %
Polymer-based	21.4 %
Metal-based	19.2 %
Lipid-based	17.8 %
Core-shell	7.1 %
Carbon-based	1.5 %
Others	1.2 %

This distribution indicates that the resource is dominated by metal oxide-based, polymer-based, metal-based, and lipid-based systems, while carbon-based and miscellaneous categories are comparatively sparse. A plausible implication is that model behavior will be most directly shaped by the statistical regularities present in these more heavily represented classes, although the paper’s emphasis is on generalized prediction rather than per-class reweighting (Yu et al., 18 Jul 2025).

2. Modalities, preprocessing, and missing-data strategy

NanoPro-3M combines structured experimental descriptors with learned representations of proteins and nanomaterials. The structured tabular component contains 29 features: 14 nanomaterial physicochemical properties, 9 incubation conditions, 5 separation parameters, and 1 proteomic-setting feature. Reported nanomaterial descriptors include core composition, surface modification, zeta potential in water or dispersion medium, hydrodynamic diameter (DLS size), polydispersity index (PdI), concentration, and shape. Incubation descriptors include protein source, concentration, medium, temperature, time, flow speed, static or flow condition, and culture versus in vivo. Separation parameters include method, centrifugation speed, time, repetitions, and temperature, while the proteomic-setting feature is the depth of LC–MS analysis (Yu et al., 18 Jul 2025).

Protein sequences are represented by 2,560-dimensional embeddings derived from ESM2_t36_3B_UR50D. Nanomaterial-side tabular information is also recast into “holistic” tabular text embeddings, represented as 4,096-dimensional vectors using Linq-Embed-Mistral with structured prompts that inject numeric feature values. This design makes the nanomaterial branch a learned semantic projection of structured metadata rather than a direct feed-forward encoding of raw scalar inputs (Yu et al., 18 Jul 2025).

The preprocessing pipeline includes LLM-assisted extraction of tables plus manual verification, semantic alignment of nomenclature across more than 100 categories, and unit standardization to mg/L for concentrations, nm for sizes, and mV for zeta potentials. The paper gives the example of core “Au and Fe₃O₄” being unified to “Au@Fe₃O₄.” RPA estimation is based on normalization of spectral counts, iBAQ, intensity, peptide number, and emPAI, with a Box–Cox transform used for regression. Train/validation/test partitioning is reported as 8:1:1 stratified by binned RPA, with raw and filled counterparts kept in the same split (Yu et al., 18 Jul 2025).

Missingness is addressed by both fill policies and imputation. “Local fill” sets within-study missing proteins to zero RPA, yielding 1.19 M samples. “Global fill” sets RPA to 0 for proteins commonly observed in more than 10 % of samples across at least 3 studies but unreported in a given context, yielding 190 k samples. For numeric tabular features, the reported imputation strategies are weighted mean and mode within grouped nanomaterial categories, LLM-based retrieval augmented generation for contextual inference, and example-based LLM inference together with a reasoning LLM for remaining gaps. The resulting datasets are a merged set of 2.76 M samples and a filled-only set of 1.38 M samples (Yu et al., 18 Jul 2025).

One recurrent concern in large integrated experimental resources is whether aggressive filling or harmonization can distort the target signal. The paper addresses this concern empirically rather than rhetorically, by separately evaluating filled and raw inputs; those results are discussed below.

3. Formal prediction objective

The prediction problem is defined in terms of a nanomaterial representation $x_n \in \mathbb{R}^{d_n}$, described as the projected tabular embedding, a protein embedding $x_p \in \mathbb{R}^{d_p}$, and a true binding affinity $y$, which may be either continuous RPA or a binary class label. The prediction model is

$\hat{y} = f_\theta(x_n, x_p).$

For regression, NanoPro-3M uses mean squared error:

$$L_{\mathrm{reg}}(\theta) = \frac{1}{N}\sum_{i=1}^{N}\left(f_\theta\!\left(x_n^{(i)},x_p^{(i)}\right)-y^{(i)}\right)^2.$$

For classification, the formulation is weighted cross-entropy on binary labels $y \in \{0,1\}$, with positive-class weight $w_+ = 2\cdot(N_-/N_+)$:

$$L_{\mathrm{cls}}(\theta) = -\frac{1}{N}\sum_{i=1}^{N} w_{y^{(i)}}\Bigl[ y^{(i)}\log \hat{y}^{(i)} + \left(1-y^{(i)}\right)\log\left(1-\hat{y}^{(i)}\right) \Bigr].$$

The paper’s binary classification threshold is RPA $> 0.001 \%$ (Yu et al., 18 Jul 2025). The coexistence of regression and classification objectives reflects two distinct but related usage modes: quantitative affinity estimation and coarse-grained interaction detection. This suggests that NanoPro-3M is intended not merely as a ranking corpus but as a substrate for both thresholded decision-making and continuous-response modeling.

4. NanoProFormer architecture and optimization

NanoProFormer is a multimodal encoder with two parallel branches. In the protein branch, $$x_p \rightarrow \mathrm{ESM2} \rightarrow h_p \in \mathbb{R}^{2560} \rightarrow$$ linear projection $x_p \in \mathbb{R}^{d_p}$0. In the tabular branch, the 4,096-dimensional text embedding $x_p \in \mathbb{R}^{d_p}$1 is projected to $x_p \in \mathbb{R}^{d_p}$2. The fused representation is then built with stacked Multi-Head Cross-Attention (MHCA) layers (Yu et al., 18 Jul 2025).

The cross-modal attention is specified as

$x_p \in \mathbb{R}^{d_p}$3

with

$x_p \in \mathbb{R}^{d_p}$4

and the same operation is repeated with roles swapped for symmetric fusion. The final fused vector $x_p \in \mathbb{R}^{d_p}$5 is passed to an MLP prediction head. Architecturally, this means the model does not simply concatenate modalities; it explicitly models directed nanomaterial-to-protein and protein-to-nanomaterial interactions before prediction.

The parameterization combines very large pretrained backbones with a comparatively small task-specific fusion stack. The reported counts are approximately 3 B parameters for ESM2, approximately 7 B parameters for Linq-Embed-Mistral, and approximately 25 M parameters for the projection and fusion modules. Training uses AdamW with learning rate $x_p \in \mathbb{R}^{d_p}$6, weight decay $x_p \in \mathbb{R}^{d_p}$7, batch size 64, 1,000 warmup steps, and a cosine decay schedule. Training proceeds for 30 epochs on the merged dataset of 2.76 M samples, with early stopping on validation $x_p \in \mathbb{R}^{d_p}$8 or AUC (Yu et al., 18 Jul 2025).

The deployment modes are zero-shot inference and fine-tuning. In zero-shot mode, new $x_p \in \mathbb{R}^{d_p}$9 and $y$0 are encoded via frozen pretrained encoders, then processed through fusion and the head to produce $y$1. In fine-tuning mode, the encoders and fusion module are frozen and only the MLP prediction head is trained on task-specific data. The reported fine-tuning split is 70 % train, 15 % validation, and 15 % test, with typical fine-tuning durations of 10–20 epochs and learning rate $y$2 (Yu et al., 18 Jul 2025). This setup places most adaptation burden on the terminal predictor while preserving the pretrained multimodal representation.

5. Empirical performance, robustness, and ablation

On the merged test set, the paper reports clear separation between multimodal and single-modality baselines for both binary classification and continuous regression (Yu et al., 18 Jul 2025).

Modality	Classification (Accuracy / F1 / AUC)	Regression ($y$3 / RMSE / MAE)
Tabular only	0.80 / 0.82 / 0.92	0.75 / $y$4 / $y$5
Protein only	0.87 / 0.83 / 0.94	0.82 / $y$6 / $y$7
Multimodal	0.90 / 0.84 / 0.96	0.88 / $y$8 / $y$9

These results support the paper’s central claim that multimodal modeling significantly outperforms single-modality approaches. The relative gap is especially pronounced for regression, where the multimodal model reaches $\hat{y} = f_\theta(x_n, x_p).$0 versus $\hat{y} = f_\theta(x_n, x_p).$1 for protein only and $\hat{y} = f_\theta(x_n, x_p).$2 for tabular only. A plausible interpretation is that quantitative affinity prediction depends more strongly on cross-modal complementarity than thresholded classification does.

Robustness to missing data is evaluated by training on the merged dataset and testing on filled versus raw inputs. The classification accuracy is 0.90 on filled inputs and 0.89 on raw inputs; regression performance is $\hat{y} = f_\theta(x_n, x_p).$3 on both filled and raw inputs (Yu et al., 18 Jul 2025). This directly addresses the concern that the fill procedures might artificially inflate benchmark performance. The reported metrics indicate near-identical performance under both input conditions.

The ablation analysis further characterizes what the model uses. At the modality level, the paper summarizes the pattern as multimodal $\hat{y} = f_\theta(x_n, x_p).$4 protein-only $\hat{y} = f_\theta(x_n, x_p).$5 table-only. At the feature level, the top individual impacts, measured by $\hat{y} = f_\theta(x_n, x_p).$6 or $\hat{y} = f_\theta(x_n, x_p).$7 upon feature removal, are ranked as follows: core composition, zeta potential, PdI in water, separation method, proteomic depth, nanomaterial concentration, and incubation protein source. For feature interactions, the paper reports that redundancy dominates in classification, while synergy dominates in regression (Yu et al., 18 Jul 2025). The reported importance ordering also grounds the paper’s statement that the model identifies key determinants of corona formation.

6. Downstream tasks and broader significance

NanoProFormer is presented not only as an affinity predictor but also as a transferable representation for downstream nano-bio tasks. In the paper’s fine-tuning studies, the reported downstream outcomes are: antibody binding prediction with fine-tuned AUC increased by 33 % on a small dataset, cell receptor enrichment with fine-tuned $\hat{y} = f_\theta(x_n, x_p).$8 increased by 0.68, disease biomarker capture with AUC up to 0.97 post-fine-tuning, and deep proteomic profiling with fine-tuned $\hat{y} = f_\theta(x_n, x_p).$9 from $$L_{\mathrm{reg}}(\theta) = \frac{1}{N}\sum_{i=1}^{N}\left(f_\theta\!\left(x_n^{(i)},x_p^{(i)}\right)-y^{(i)}\right)^2.$$0 to $$L_{\mathrm{reg}}(\theta) = \frac{1}{N}\sum_{i=1}^{N}\left(f_\theta\!\left(x_n^{(i)},x_p^{(i)}\right)-y^{(i)}\right)^2.$$1 (Yu et al., 18 Jul 2025). The article also states that downstream applicability is demonstrated through both zero-shot inference and fine-tuning.

The stated practical implications are reducing experimental burden and accelerating in vitro assays. Specifically, the paper argues that reliable in silico affinity screening lowers the need for extensive proteomic assays, while zero-shot predictions can guide nanomaterial design and protocol choices, including surface chemistry, incubation settings, and separation protocols, before bench work (Yu et al., 18 Jul 2025). In this framing, the model is not replacing proteomics; it is reorganizing when and where expensive empirical measurements are most necessary.

The broader impact is formulated in terms of “corona-aware” nanomaterial design for diagnostics, therapeutics, and environmental safety, and in terms of benchmarking. The paper states that the resource provides a static benchmark analogous to ImageNet for nano-bio interactions (Yu et al., 18 Jul 2025). This suggests a shift from narrowly scoped, study-specific predictors toward reusable pretrained infrastructure for nanomaterial–protein interaction endpoints. Within that perspective, NanoPro-3M functions simultaneously as a curated data asset, a multimodal learning benchmark, and a transfer substrate for related prediction tasks.