Neural Proteomics Fields (NPF)

• Neural Proteomics Fields (NPF) are proteomics frameworks using neural models to convert protein data into structured, predictive representations.
• They encompass diverse formulations from alignment-free sequence modeling to spatio-temporal fields and graph-constrained dynamics for applications such as biomarker discovery and neuropeptide function.
• NPF bridges deep learning with domain-specific constraints, enabling both high predictive performance and mechanistic interpretation across various biological contexts.

Neural Proteomics Fields (NPF) denotes a family of proteomics frameworks in which neural or machine-learning models are used to convert protein-related data into structured representations, predictive mappings, or explicit fields over biological, spatial, or temporal domains. Across the literature represented here, the term is not restricted to a single formalism. It includes neural representation learning on reverse-phase protein array data for biomarker discovery and mutation association in acute myeloid leukemia (Liang et al., 2017), alignment-free sequence-based discovery of neuropeptide precursors and other high-level protein functions (Ofer, 2016), spatio-temporal fields of misfolded-protein burden on brain networks (Garbarino et al., 2019), low-rank spatial embeddings of compartment-specific proteomes (Felizzi et al., 2012), implicit neural fields for super-resolved sequencing-based spatial proteomics (Zhao et al., 24 Aug 2025), physics-informed protein identification from nanopore time series (Dutt et al., 2023), and a broader umbrella for deep learning across protein sequence, structure, function, and interaction tasks (Luo et al., 2024).

1. Conceptual scope and major formulations

Taken together, these works present NPF as a family of formulations rather than a single architecture. The common thread is the treatment of proteomic information as a learnable object: a latent feature space, a continuous spatial field, a graph-constrained dynamical system, or a morphology-conditioned predictor. In one line of work, NPF is explicitly defined as the systematic use of neural representation learning on proteomic measurements to discover biomarkers, associate them with genetic lesions, and organize them into interpretable signaling pathways for clinical decision support (Liang et al., 2017). In another, it is operationalized as proteome-scale discovery and annotation of neuronal modulators and other brain-relevant protein functions directly from primary sequence (Ofer, 2016). Elsewhere, it is a mesoscopic vector-valued field of misfolded protein concentrations evolving over a brain network (Garbarino et al., 2019), or an implicit neural representation that reconstructs protein expression in continuous tissue space from sparse sequencing spots and histology (Zhao et al., 24 Aug 2025).

Formulation of NPF	Primary data	Representative goal
Neural representation learning on proteomic measurements	RPPA, clinical-genomic annotations	biomarker discovery, mutation classification
Alignment-free sequence-based proteomics	protein primary sequence	neuropeptide precursor discovery, function annotation, RT prediction
Spatial and spatio-temporal protein fields	tissue coordinates, brain networks, compartmental proteomes	super-resolution, propagation modeling, compartment organization
Single-molecule and electrochemical NPF	nanopore currents, proteinoid electrochemistry	protein identification, interfacing protein-based excitable materials

This diversity is not merely terminological. It marks a shift from viewing proteomics as a collection of isolated measurements toward viewing it as a field problem: proteins are modeled as nodes on interaction graphs, coordinates in tissue, evolving burdens on anatomical networks, or structured signals in time series. A recurrent implication is that NPF links predictive modeling to mechanistic interpretation, although the balance between those aims varies substantially across works.

A separate source of ambiguity is the acronym itself. The acronym “NPF” also collides with “Neural Process Family”; the supplied note for that item explicitly states that the source material does not concern proteomics (Jha et al., 2022). In proteomics usage, NPF refers to neural learning on protein data rather than to neural-process meta-learning.

2. Molecular and sample-level representation learning

At the sample level, one canonical NPF formulation is the analysis of proteomic measurements for mutation association and biomarker discovery in cancer. In acute myeloid leukemia, stacked autoencoders were trained on 231 antibody-based reverse-phase protein array measurements from a biologically restricted cohort of 62 newly diagnosed, treatment-naïve patients with normal cytogenetics and FLT3-ITD as the sole mutation status variable (Liang et al., 2017). The model used unsupervised pretraining followed by supervised fine-tuning, and feature ranking was based on the sum of absolute input-to-first-hidden-layer weights,

$\mathrm{Importance}_i = \sum_j |w_{(i \to j)}|.$

This reduced the panel from 231 proteins to 20 “critical” proteins, cutting the data points from 14,322 to 1,240. On the 20-protein panel, the reported deep-learning performance was Accuracy $=97\%$, Sensitivity $=90\%$, and Specificity $=100\%$, compared with $81\%$ accuracy for the deep network on all 231 proteins and $87\%$ accuracy for a conventional neural network on the reduced panel. The selected proteins were interpreted as mapping to receptor tyrosine kinase signaling and downstream cascades associated with FLT3-ITD, including RAS/MAPK, PI3K/AKT/mTOR, and STAT, as well as apoptotic and stress-response regulators. In this formulation, NPF is both a predictive classifier and a pathway-generating device.

A second formulation works directly on primary sequence. In the NeuroPID and ProFET framework, NPF becomes alignment-free proteome-scale learning of high-level protein function from engineered sequence-derived features (Ofer, 2016). The feature space spans amino acid composition, k-mer frequencies, reduced alphabets, physicochemical descriptors such as GRAVY and pI, information-theoretic measures such as Shannon entropy, cleavage-associated motifs, disorder proxies, CTD transforms, and propensity-scale statistics. Feature selection is performed with RF-RFECV, and classification is carried out with SVMs, Random Forests, Gradient Boosting Trees, Logistic Regression, and Extra Trees. On the NeuroPID V1 dataset, cross-validation yielded Accuracy $0.928$–$0.930$, Precision $0.918$–$0.938$, Recall $=97\%$0–$=97\%$1, MCC $=97\%$2, and AUC $=97\%$3–$=97\%$4. On broader ProFET benchmarks, $=97\%$5 of 17 tasks achieved Accuracy and F1 $=97\%$6, and $=97\%$7 exceeded $=97\%$8. Here NPF is not a latent field over tissue or time; it is a learned functional field over proteome space, built from whole-sequence descriptors and applied at proteome scale.

A related sequence-to-observable formulation is peptide retention time prediction. DeepRT uses deep CNN and RNN feature learning directly from peptide sequence, followed by PCA retaining $=97\%$9 cumulative variance and downstream SVR, Random Forest, and Gradient Boosting regression, combined by bagging (Ma et al., 2017). The CNN branch uses one-hot encoding with padding, four convolutional layers, leaky ReLU with negative slope $=90\%$0, and dropout ratios $=90\%$1 or $=90\%$2; the RNN branch uses LSTM layers with embedding length $=90\%$3 and projection length $=90\%$4. On two public datasets, the reported Pearson correlations were $=90\%$5 and $=90\%$6, with RMSE and $=90\%$7 smaller than ELUDE and GPTime. This work exemplifies a core NPF principle: experimentally relevant proteomics observables can be predicted from neural sequence representations without hand-crafted expert features.

3. Spatial, compartmental, and brain-network fields

One of the most literal realizations of NPF is the treatment of protein abundance as a continuous or graph-defined field. In neurodegenerative disease modeling, misfolded protein burden is represented as a vector-valued trajectory

$=90\%$8

over brain regions, with $=90\%$9 macro-nodes connected by a structural connectome (Garbarino et al., 2019). The group trajectory is modeled by a Gaussian Process constrained by a non-linear Accumulation–Clearance–Propagation system, and subject-specific observations are aligned by a monotonic shift

$=100\%$0

In the Alzheimer’s disease application, the graph was reduced to 11 macro-regions and fit to 1091 ADNI participants with 2380 AV45-PET scans. Forward-integrated predictions on held-out follow-up data showed lower RMSE for the ACP model than for a diffusion baseline in most regions, including frontal $=100\%$1 versus $=100\%$2, temporal $=100\%$3 versus $=100\%$4, and cingulate $=100\%$5 versus $=100\%$6. In this setting, NPF is a mechanistically constrained spatio-temporal disease field rather than a generic predictor.

A distinct spatial formulation appears in compartmental proteomics. Proteins from neurite and soma proteomes are embedded on the sphere $=100\%$7 by maximizing

$=100\%$8

where the edge weights combine STRING reliability with neurite-versus-soma enrichment derived from LC-MS/MS (Felizzi et al., 2012). The update rule

$=100\%$9

is iterated to equilibrium, yielding a low-rank spatial organization of the proteome with $81\%$0. Euclidean distances and adjacency-filtered distances define per-protein scores such as $81\%$1, while perturbation scores $81\%$2 quantify reorganization around driver proteins such as Cdc42 or Rac1 after node removal. The reported functional scores correlated with observed neurite dynamics upon knockdown, with ITSN1 at $81\%$3, Arhgap17 at $81\%$4, and Vav3 at $81\%$5. This formulation interprets the proteome as a low-dimensional spatial field whose geometry reflects compartment-specific signaling organization.

The most explicit neural-field formulation is the super-resolution of sequencing-based spatial proteomics. Here NPF learns, for each tissue section, a continuous function $81\%$6 from normalized 2D coordinates and histology patches to protein abundance (Zhao et al., 24 Aug 2025). The Spatial Modeling Module uses Fourier positional encodings

$81\%$7

with $81\%$8, followed by $81\%$9 cascaded MLP blocks with residual connections and latent dimension $87\%$0. The Morphology Modeling Module combines a frozen UNI ViT-Large backbone pre-trained on 100M pathology images via DINOv2 with a tissue-specific CNN pyramid and reciprocal deformable cross-attention. Training minimizes MSE on log-transformed protein values, one dedicated network per tissue. On the Pseudo-Visium SP benchmark, NPF achieved mean MSE $87\%$1 and mean PCC $87\%$2 across 12 samples. On real 10X Visium spatial proteomics, it reported MSE $87\%$3, PCC $87\%$4 on Human Tonsil and MSE $87\%$5, PCC $87\%$6 on Human Tonsil Add-on Antibodies. This is NPF as an implicit continuous tissue field, conditioned jointly on coordinates and morphology.

4. Single-molecule sensing and protein-based excitable media

NPF also extends to single-molecule sensing, where the learned object is not a bulk proteome but the morphology-sensitive dynamics of individual translocation events. In solid-state nanopore protein identification, ultrathin silicon nitride nanopores of thickness $87\%$7 nm and diameter $87\%$8–$87\%$9 nm were used to sense four similarly sized proteins: bovine hemoglobin, human serum albumin, bovine serum albumin, and concanavalin A (Dutt et al., 2023). Event waveforms were acquired at 100 kHz bandwidth with 200 ksps and at 10 MHz bandwidth with 40 Msps, with adaptive detection at thresholds $0.928$0. Features included normalized blockade, dwell time, segmented intra-event medians $0.928$1, $0.928$2, $0.928$3, area $0.928$4, skewness, and kurtosis, all normalized to open-pore conductance $0.928$5. Random Forest outperformed Rotation Forest, and Scheme 3—ten segments plus $0.928$6 area, skew, and kurtosis—gave the best results. Without clustering, the best F-values were $0.928$7 at 100 kHz and $0.928$8 at 10 MHz. With K-means clustering, the best 10 MHz result reached F-value $0.928$9, Sensitivity $0.930$0, and Specificity $0.930$1. In this context, NPF is a physics-informed mapping from raw nanopore time series to protein embeddings and classifications.

A more unconventional branch treats proteinaceous materials themselves as neural-like substrates. Proteinoids, formed by heating amino acids to $0.930$2–$0.930$3C under inert atmosphere and rehydrating the resulting polymers into microspheres, were studied as “protoneurons” that generate electrical spikes and can be assembled into proto-nano-brains (Mougkogiannis et al., 2023). Bulk activity was recorded with paired iridium-coated stainless steel sub-dermal needle electrodes at 1 sample/s, using a current threshold of $0.930$4 and a minimum peak distance of 5 s. Across compositions, the mean number of spikes per recording was $0.930$5, with a range from 8 to 900, and mean inter-spike intervals ranged from $0.930$6 s to $0.930$7 s. Differential pulse voltammetry was proposed as an electrochemical interface, using a Zimmer Peacock Anapot EIS with $0.930$8 V, $0.930$9 V, step size 1 mV, base scan rate $0.918$0 mV·s$0.918$1, pulse amplitude $0.918$2 mV, pulse width $0.918$3 ms, and 100 s equilibration. Distinct DPV peak patterns were reported across 12 proteinoid compositions, and peak height was described as proportional to microsphere count. This work extends NPF from analytics to protein-based excitable matter, although its interpretation is more provisional than in the better-standardized machine-learning studies.

These single-molecule and electrochemical formulations broaden the meaning of “field.” In nanopores, the field is an event-wise embedding of molecular shape, charge, and interaction dynamics. In proteinoid systems, it is a network-like electrochemical state space of spiking microspheres. Both cases move NPF beyond static abundance matrices toward dynamical readouts of protein behavior.

5. Methods, objectives, and evaluation regimes

Across applications, NPF is methodologically heterogeneous but computationally recognizable. The broader review of deep learning in proteomics organizes the space by modality—sequence, structure, networks/graphs, and multi-source fusion—and by task type—classification, regression, generation, and link prediction (Luo et al., 2024). Sequence-centric models include CNNs, RNNs/LSTMs, Transformers, GANs, and message-passing models; structure-centric models include AlphaFold-style architectures, residual CNNs, MSA-aware Transformers, and single-sequence predictors; network-centric models include GCNs, GATs, hierarchical graph transformers, and autoencoders. Standard objectives include cross-entropy

$0.918$4

binary cross-entropy for multi-label function annotation,

$0.918$5

mean squared error

$0.918$6

and graph message-passing layers of the form

$0.918$7

Evaluation is correspondingly task-specific. The AML biomarker study used Accuracy, Sensitivity, and Specificity under iterative 10-case holdout validation (Liang et al., 2017). Nanopore protein identification reported Precision, Recall, F1-score, and Specificity under 10-fold cross-validation plus held-out testing (Dutt et al., 2023). DeepRT used Pearson correlation, RMSE, and $0.918$8 (Ma et al., 2017). Sequence-function studies emphasized Accuracy, Precision, Recall, F1, MCC, ROC, and AUC, with additional use of Fmax and related measures in broader function annotation contexts (Ofer, 2016, Luo et al., 2024). Spatial super-resolution used MSE and Pearson Correlation Coefficient (Zhao et al., 24 Aug 2025). Brain-network NPF used RMSE and posterior uncertainty summaries (Garbarino et al., 2019). The metric layer is therefore integral to the definition of any particular NPF instantiation; there is no universal benchmark spanning all usages.

Implementation patterns recur despite this diversity. One pattern is staged learning: unsupervised pretraining followed by supervised fine-tuning in stacked autoencoders (Liang et al., 2017), or deep feature extraction followed by PCA and conventional regression in DeepRT (Ma et al., 2017). A second is explicit normalization against nuisance variation, such as conductance normalization by $0.918$9 in nanopore sensing (Dutt et al., 2023) or SuperCurve normalization in RPPA proteomics (Liang et al., 2017). A third is structure-aware modeling: graph topology in brain propagation and interactome embedding, morphology in spatial proteomics, and cleavage or disorder motifs in sequence-based neuropeptide discovery. These patterns show that NPF is not simply “deep learning applied to proteins”; it typically couples learned representations to domain-specific constraints.

6. Limitations, misconceptions, and future directions

A persistent limitation across NPF studies is restricted generalizability. The AML autoencoder study used only 62 cases, with no independent held-out test set or external cohort, no reported class balance proportion, and no explicit batch correction beyond SuperCurve (Liang et al., 2017). The neurite-versus-soma embedding study emphasized qualitative association with knockdown phenotypes but did not report AUCs or p-values (Felizzi et al., 2012). The Alzheimer’s disease propagation model relied on a symmetric, thresholded average healthy connectome, constant group-level kinetics, and PET measurements affected by partial volume effects, off-target binding, and atrophy confounds (Garbarino et al., 2019). Per-tissue training in seq-SP super-resolution directly addresses inter-tissue variability but requires retraining for each tissue and can be compute-intensive (Zhao et al., 24 Aug 2025).

Another recurrent issue is interpretability. Some NPF frameworks make interpretability central by design: feature ranking by connection weights in AML, pathway assignment of critical proteins, RF-RFECV in sequence-function modeling, explicit graph kinetics in ACP, and geometric low-rank organization in neurite proteomics [(Liang et al., 2017); (Ofer, 2016); (Garbarino et al., 2019); (Felizzi et al., 2012)]. Others achieve strong empirical performance with less transparent internal structure, especially when using large morphology backbones or event classifiers (Zhao et al., 24 Aug 2025, Dutt et al., 2023). A plausible implication is that NPF is pulled between mechanistic interpretability and predictive optimization, with different subfields settling that tradeoff differently.

The most obvious controversy concerns unconventional or weakly standardized settings. In the proteinoid microsphere study, sampling at 1 Hz cannot substantiate ms-scale spike-shape claims, some tabulated frequency values in mHz do not match $0.938$0, the $0.938$1 V potential window was not discussed in relation to water stability, and reproducibility, linearity coefficients, and LOD/LOQ were not reported (Mougkogiannis et al., 2023). In nanopore sensing, the gains from high bandwidth and clustering are strong, but data rates at 40 Msps are operationally demanding and models remain sensitive to event heterogeneity and voltage-induced secondary populations (Dutt et al., 2023). In sequence-function learning, negative sets may contain unknown positives, and handcrafted features, even when extensive, remain an approximation to the full determinants of function (Ofer, 2016).

Future directions are correspondingly convergent. Several works explicitly point toward multi-omics fusion, graph neural networks, variational autoencoders, Transformers, and richer interpretability tooling such as SHAP values, integrated gradients, and layer-wise relevance propagation (Liang et al., 2017). Nanopore-based NPF points toward end-to-end waveform encoders, metric learning, domain adaptation across pores and instruments, cluster-aware experts, and benchmark standardization (Dutt et al., 2023). Sequence-centric NPF points toward protein LLMs, structural prediction integration, PTM-aware modeling, and positive–unlabeled learning (Ofer, 2016). Spatial NPF points toward finer uncertainty modeling, distribution-aware losses, and broader cross-tissue validation (Zhao et al., 24 Aug 2025). The broader proteomics review emphasizes multimodal integration, self-supervised and transfer learning, interpretability, efficiency, and ethical handling of biomedical data as the major determinants of future progress (Luo et al., 2024).

In aggregate, NPF is best understood as a research program rather than a fixed model class. Its defining move is to treat proteomic data—whether sequence, abundance matrix, interactome, tissue coordinate, nanopore trace, or protein-based excitable material—as an object from which neural methods can learn structured fields. Those fields may be latent, spatial, temporal, morphological, or electrochemical, but they share the same ambition: to turn complex protein data into predictive and, where possible, mechanistically interpretable representations.