MS-based proteomics

The HPA currently includes two independent bulk MS-based proteomics datasets, generously made available by the respective research groups and integrated separately into the resource using HPA filtering criteria and tissue ontology mapping. These datasets were generated by the respective research groups and integrated by HPA to improve data accessibility and facilitate exploration and comparison across independent proteomics studies.

Based on two publications:

MS data overview

ProteinTalks-DB data: The dataset was generated by profiling human tissues with DIA mass spectrometry. It includes healthy adult tissues as well as fetal, tumor and adjacent non-tumor samples, of which only healthy adult tissues are included in the HPA Tissue Resource. Autopsy samples were collected from the body donation centers of Dalian Medical University and Shanghai Jiao Tong University, with a post-mortem interval within 10 hours. The project was approved by the Ethics Committee of Dalian Medical University. Written informed consent was obtained from the legally authorized representatives of all donors. Anatomical accuracy was ensured by macrodissection by two experienced anatomists following a standardized protocol. For tissues that could not be clearly partitioned anatomically, this was complemented by histopathological review of H&E-stained FFPE sections by two pathologists.

Tissue subtypes and anatomical locations were pooled into main tissue types matching the HPA tissue ontology. For example, frontal, insular, occipital, parietal and temporal lobes were pooled as cerebral cortex, and ascending, transverse, descending and sigmoid colon as colon. Tissue types without a counterpart in the HPA ontology, such as ear, eye, nose, bone, cartilage, tendon and heart valves, were not included. Tissue-specific markers were additionally used to select samples consistent with the histological composition and sampling strategy of the corresponding normal-tissue RNA dataset. The resulting dataset integrated into HPA Tissue Resource covers 43 tissue types from 239 samples. Complete list of tissues is available here. The complete original dataset is available through the ProteinTalks database, HPA inclusion TRUE/FALSE is available in the source details.

The cancer subset of the data includes paired tumor and adjacent non-tumor tissues. Each sample was checked for tumor content and cancer-type identity using three measures: 1- a malignancy score based on proliferation markers, 2- a tumor-likeness score comparing each sample with the tumor and non-tumor profiles of its cancer type, 3- cancer-type markers. Inclusion in the HPA Cancer Resource required samples to meet all three quality-control criteria. Rare histological subtypes and cancer types outside the scope of the resource were excluded. To preserve the paired design, a tumor–non-tumor pair was included only when both samples met the inclusion criteria. The resulting dataset covers 766 tumor / adjacent non-tumor pairs (1,532 samples) in 25 cancer types. A complete list of cancer types represented is available here. The complete original dataset is available through the ProteinTalks database, HPA inclusion TRUE/FALSE is available in the source details.

MannLab data: The bulk MS dataset comprises 20 tissues profiled by MS-based proteomics from a single healthy female donor obtained from the MunIch cardiovaScular StudIes biObaNk (MISSION). Tissue collection and handling were approved by the Ethics Committee of the Technical University of Munich and conducted in accordance with the Declaration of Helsinki. Donor information was handled in a fully anonymized manner. The 20 tissues profiled by bulk MS are cerebral cortex, lung, salivary gland, stomach, duodenum, jejunum, ileum, colon, liver, pancreas, kidney, ovary, fallopian tube, coronary artery, heart muscle, skeletal muscle, skin, spleen, lymph node, and bone marrow. Each tissue was analyzed in three biological replicates. The same tissue specimens were used for the cell type-resolved Deep Visual Proteomics (DVP) analysis presented in the Single Cell Resource, enabling direct integration of bulk and cell type-resolved measurements.

Sample preparation and mass spectrometry

ProteinTalks-DB: Approximately 1 mg of each FFPE sample was punched out and dewaxed with heptane. It was then rehydrated with a gradient of 100%, 90% and 75% ethanol and acid-hydrolysed with 0.1% formic acid, followed by alkaline hydrolysis in Tris-HCl (pH 10) at 95 °C for 30 min. Proteins were reduced and alkylated in 6 M urea/2 M thiourea lysis buffer with TCEP and IAA. Lysis was then performed with PCT on a Barocycler NEP2320-Enhanced (Pressure BioSciences). Proteins were digested with LysC and trypsin under PCT, and the peptides were desalted on C18 spin columns. Peptides were separated on a nanoElute UHPLC system (Bruker Daltonics) with a 95 min gradient at 0.3 µl/min on an in-house analytical column (120 Å, 1.9 µm, 150 mm × 75 µm). They were analysed on a timsTOF Pro mass spectrometer (Bruker Daltonics) in diaPASEF mode, using 64 windows over 400–1,200 m/z. A project-specific spectral library was built from fractionated samples acquired in ddaPASEF mode and processed with FragPipe (v18, MSFragger v3.5). Tumor and adjacent non-tumor samples were prepared and analysed with the same protocol as the healthy tissue samples.

MannLab: All tissues were processed as formalin-fixed paraffin-embedded (FFPE) samples and sectioned at 3 µm thickness. Sample preparation was performed using Solid-Phase Extraction Capture (SPEC): bulk tissue pieces were scraped from FFPE sections, deparaffinized, lysed, and proteins digested into peptides with LysC/trypsin on a strong anion exchange (SAX) tip. SPEC immobilises peptides in nanoliter volumes, enabling fast, robust and ultra-sensitive sample preparation with minimal losses. Peptides were separated on an Evosep One system coupled to an Orbitrap Astral Zoom mass spectrometer (Thermo Fisher Scientific) and analysed in data-independent acquisition (DIA) mode. Separation was performed on an Aurora Rapid C18 column (5 cm × 75 μm, 1.7 μm, IonOpticks) at 60 °C using the Whisper Zoom 80 SPD gradient. MS1 scans were acquired at 240,000 resolution over 380–980 m/z. Fragment ion spectra were acquired across 200 equidistant 3 Th DIA isolation windows covering the precursor range, with 6 ms maximum injection time and 25% HCD collision energy.

Data processing and normalization

ProteinTalks-DB healthy data: Raw files were processed with DIA-NN (v2.2.0) using the FragPipe-generated spectral library, searched against the UniProtKB/Swiss-Prot human proteome (UP000005640, 20,377 canonical sequences, downloaded 7 May 2020). Match-between-runs (MBR) was disabled. Only proteotypic peptides and protein groups without indistinguishable members were retained. Stringent FDR thresholds were applied, controlled with an entrapment strategy (Q.Value < 0.001, PG.Q.Value < 0.01, Global.Q.Value < 0.001, Global.PG.Q.Value < 0.001). Protein intensities were obtained from the ProteinTalks database, and no additional normalization was applied. Technical replicates were first merged into a single sample-level value by taking the median intensity. For each tissue type, the samples, including sub-tissues pooled into the same tissue type, were then combined into a single tissue-level value by taking the median intensity. A protein value was set to missing for a tissue type if it was detected in fewer than 25% of the tissue samples. Protein detection was defined as a non-missing intensity value.

ProteinTalks-DB cancer data: Raw files were processed with DIA-NN as described above. In the original study, proteins missing in more than 50% of samples within a cancer type were removed and the remaining missing values were imputed. For the Cancer Resource, protein intensities were log2-transformed and UniProt identifiers were mapped to Ensembl gene identifiers (Ensembl v109). As more proteins were detected in tumor samples than in adjacent non-tumor tissue, with approximately 7,500 versus 4,600 proteins detected per sample on average, protein intensities were normalized using a set of stable reference proteins. Among proteins detected in at least 90% of samples, the 10% showing the lowest variability across samples were selected, resulting in 367 reference proteins, primarily housekeeping proteins such as proteasome subunits and translation factors. Each sample was then adjusted to align the median intensity of these reference proteins across samples, correcting for differences in overall intensity while preserving the relative protein abundance distribution within each sample. Within each cancer type, proteins were retained if detected in at least 25% of tumor or 25% of adjacent non-tumor samples, allowing proteins predominantly detected in either group to be included. Missing values were not imputed. For each cancer type and sample type (tumor or adjacent non-tumor), a median was calculated across all samples, treating non-detected proteins as below the detection limit. A protein was considered detected in tumor (or non-tumor) tissue if it was detected in at least 50% of those samples, otherwise no value is given. Protein detection was defined as a non-missing intensity value.

Differential expression between tumor and adjacent non-tumor tissue: Tumor and paired adjacent non-tumor tissue were compared within each cancer type. How each protein was assessed depended on how often it was detected in each group. For every protein, the number of tumor and non-tumor samples in which it was detected is reported, including for proteins that were not tested for differential expression. Proteins detected in at least 10% of tumor and at least 10% of non-tumor samples were compared with proDA (Ahlmann-Eltze & Anders 2019), with patient as a blocking factor (design ~ patient + sample type). Instead of imputing missing values, proDA models them through a detection curve for each sample, which describes how likely a protein is to be missed at a given intensity. In a sample that detects fewer proteins, a missing value is more likely to mean that the protein was below the detection limit than that it was absent, so the model treats those values as informative. This matters here because more proteins were detected in tumor samples than in adjacent tissue. P-values were adjusted with the Benjamini–Hochberg method within each cancer type, over the tested proteins. Proteins with an adjusted p-value < 0.05 and |log2 fold change| > 0.5 were classified as up or down in tumor.

Proteins detected in fewer than 10% of the tumor or non-tumor samples were not tested with proDA, as the estimate would then depend largely on the detection limit rather than on measured values. Instead, proteins were classified as mainly detected in tumor or mainly detected in non-tumor if two criteria were met: (1) the detection rate differed between the two groups by at least 50 percentage points and at least six samples; (2) at least 80% of the samples in the group with lower detection frequency had a detection limit below the median protein abundance observed in the group with higher detection frequency. For each sample, the detection limit was defined as the 10th percentile of its measured log2 intensities. This detection control is necessary because greater proteome coverage was observed in tumor samples, meaning that a protein may appear tumor-specific simply because it falls below the detection limit in non-tumor samples. No fold change or p-value is reported for proteins mainly detected in tumor or non-tumor. Proteins not meeting the criteria are classified as not tested, and only their detection counts are provided. Glioblastoma (8 pairs) and clear cell ovarian carcinoma (10 pairs) have relatively few samples, and their results should therefore be interpreted with caution.

MannLab: Raw files were processed with DIA-NN (v1.9.2) in library-free mode using a deep-learning-predicted spectral library generated from the reviewed human reference proteome (UniProt UP000005640, downloaded March 2024). Individual tissue replicates were searched together in one search with MBR and peptidoform analysis enabled. The resulting report files were combined, and protein intensities were normalized using directLFQ (version 0.2.19) with standard settings. No imputation of missing values was applied. For each tissue, the three biological replicates were combined into a single tissue-level proteome by taking the median protein intensity across replicates. A protein value was set to missing for a tissue if more than half of its replicates lacked a measurement for that protein. Protein detection was defined as a non-missing intensity value.

Classification of MS-based protein data

The bulk MS data were used to classify protein-coding genes based on their tissue-specific protein expression patterns along two dimensions: specificity category and distribution category. The same classification framework is applied to transcriptomics data elsewhere in the HPA, enabling direct comparisons between RNA and protein expression patterns. The classification was performed independently for the two MS datasets, reflecting differences in tissue coverage. The integrated ProteinTalks-DB dataset includes normal tissue samples from 43 distinct tissues, grouped into 34 main tissue types, while the MannLab dataset covers 20 tissues, grouped into 16 main tissue types. For both datasets, tissues were grouped according to the same strategy used for bulk RNA-seq classification, ensuring consistency across data modalities. For the Cancer resource, the ProteinTalks-DB cancer data were classified across the 25 cancer types in the same way, using the tumor medians. Fold changes between cancer types were calculated on linear intensities. Genes detected in only one cancer type were classified as cancer enriched.

Explanation of the specificity category

The MS-based proteomics data was used to classify all genes according to their tissue-specific into two different schemas: specificity category and distribution category.

Category Description
Enriched Intensity level in a particular tissue type is at least four times any other tissue/cancer type
Group enriched Intensity level in a group (of 2-5 tissues) is at least four times any other tissue/cancer type
Enhanced Enhanced: Intensity level in one or several cancer types is at least four times higher than the mean intensity across all tissue/cancer types, provided the lowest-expressing enhanced tissue/cancer type is at least twofold higher than the average of the next three tissues/cancer types (a gap ratio of ≥ 2)
Low specificity Detected in at least one tissue type but not elevated in any tissue type
Not detected Missing intensity value in all tissue types


An additional category "elevated", containing all genes in the first three categories (tissue/cancer enriched, group enriched and tissue/cancer enhanced), has been used for some parts of the analysis. TS/CS-score (Tissue Specificity/Cancer Specificity score) is calculated for “elevated” tissues/cancer. TS/CS-score is calculated as the fold change from the tissue/cancer with highest RNA to the tissue/cancer with second highest RNA.

Explanation of the distribution category

Category Description
Detected in single Detected in a single tissue/cancer
Detected in some Detected in more than one but less than one third of tissues/cancer
Detected in many Detected in at least a third but not all tissues/cancer
Detected in all Detected in all tissues/cancer types
Not detected Missing intensity value in all tissue/cancer types