Karolinska Institutet · Dep. of Clinical Neuroscience & Dep. of Neurobiology, Care Sciences and Society

CSF Autophagy & Lysosome
Biomarker Panel Discovery

Pontus Plavén-Sigray and Aris Logothetis

Translational pipeline · 12 datasets · ~3,600 samples · Mouse ↔ Human

The question

Which autophagy/lysosomal proteins can we reliably measure in CSF across both species?

We need candidates that are robustly detected in human and mouse CSF, biologically relevant to autophagy/lysosomal pathways, and actually measurable with targeted assays.

Methods

Pipeline architecture

01Extract & QC

→

02Orthology mapping

→

03Evidence scoring

→

04Autophagy filter

05Peptide feasibility

→

06Module validation

→

07Sensitivity

Weighted evidence scoring replaces binary intersection. EV data contributes a small scoring bonus (5%) but is never required for inclusion. All parameters configurable; sensitivity tested across multiple configurations.

Methods

Input data

🐭 Mouse (9 datasets)

7 CSF datasets — D1–D5, D7, D8 (789–5,961 proteins)
1 brain lysate — D6 (7,645 proteins; plausibility layer)
1 ISF — D9 (789 proteins; orthogonal support)

🧬 Human (3 datasets + references)

D11 Astral — 3,232 proteins × 2,720 samples (primary)
D12 — replication subset of D11 (validation only)
D10 — 990 proteins, independent cohort
R1 — 604 curated autophagy/lysosome genes

EV dataset (SH-SY5Y cell line) contributes a small scoring bonus (5% weight) but is never required for panel inclusion.

Methods

Quality control

Sample filtering

Excluded samples with >80% missing data. Removed lab artefacts (MaxQuant decoys & contaminants).

Blood contamination

235 proteins flagged as likely plasma-derived (albumin, Ig, complement, keratins). Excluded from core panel.

Detectability tiers

Dataset-aware confidence grading:
TIER A robust TIER B moderate
Thresholds scale with dataset size (n=8 → n=2,720).

Methods

Orthology mapping — mouse → human

All mouse gene symbols (D1–D9) mapped to human equivalents via g:Profiler (Ensembl gene trees), so we can compare across species in a single namespace.

89%

Clean 1:1 mappings (8,589 genes) — one mouse gene → one human gene

5%

Ambiguous 1:many (477 genes) — flagged, kept, penalised −0.1 in scoring

6%

No human ortholog found (~604 genes) — dropped from analysis

Paralog families

~30 core autophagy genes had non-trivial mappings due to paralog families (e.g. LC3/GABARAP, ATG4 variants). These were identified upfront by mapping R1 separately and handled explicitly.

Ambiguous cases use an expand policy: all possible human matches kept but flagged — nothing silently lost.

Methods

Evidence scoring

Each protein scored 0–1 from five weighted components. No single source determines inclusion.

Human CSF

30%

Mouse CSF

25%

Autophagy R1

25%

Brain (D6)

10%

EV support

5%

Core panel criteria (hard filters)

Tier A/B in human CSF (D11)
Tier A/B in ≥1 mouse CSF dataset
On the autophagy/lysosome list (R1)
Not flagged as plasma-derived

EV detection is never required for inclusion. It contributes a small scoring bonus only.

Results

From proteome to panel

9,561

unique proteins scored

6,684

detected in ≥1 mouse CSF

3,026

Tier A/B in human Astral

434

overlap with R1 autophagy list

122

pass all 4 core criteria

80

final shortlist (by score)

Results

Top 10 candidates

#	Protein	Score	Feas.	Mouse
1	CTSD	0.975	I	7/7
2	CTSB	0.927	II	7/7
3	CST3	0.927	II	7/7
4	CPQ	0.926	I	7/7
5	PRDX6	0.926	II	7/7
6	PARK7	0.925	I	7/7
7	HEXB	0.922	I	7/7
8	GGH	0.909	II	7/7
9	GAA	0.902	I	7/7
10	LAMP1	0.896	I	7/7

#	Protein	Score	Feas.	Mouse
11	ATP6AP1	0.887	I	7/7
12	NPC2	0.884	I	7/7
13	GM2A	0.884	I	7/7
14	EPDR1	0.874	I	7/7
15	IDS	0.873	I	7/7
16	CTSS	0.872	I	7/7
17	CTSZ	0.870	I	7/7
18	PPT1	0.869	I	6/7
19	MAN2B1	0.869	I	7/7
20	HEXA	0.865	I	7/7

Feas. = peptide feasibility tier for MS (I = ≥2 conserved proteotypic peptides mouse↔human; II = human-only). Mouse = datasets detected in (of 7). All 20 in all 7 mouse CSF datasets except PPT1 (6/7). Dominated by lysosomal enzymes and autophagy effectors.

Results

Detection robustness (of 80 shortlisted)

79

Tier A in human Astral (D11)

78

confirmed in replication (D12)

78

detected in mouse brain (D6)

28

found in all 7 mouse CSF sets

Nearly all candidates are independently replicated, brain-expressed, and cross-species detected.

Results

Peptide feasibility — can we measure these?

99 of 122

TIER I ≥2 conserved proteotypic peptides (mouse ↔ human)

23

TIER II Human peptides available; separate mouse assay needed

0

Tier III (not measurable)

What this means

Tier I proteins can be quantified with a single PRM/MRM assay using identical peptides in both species — directly bridging preclinical and clinical studies.

For antibody-based platforms (Olink, ELISA, suspension bead arrays), different feasibility criteria apply — antibody availability and epitope specificity rather than peptide conservation.

Results

Co-abundance modules — does the panel reflect real biology?

Clustered all 2,323 proteins in the Astral dataset by CSF co-variation across 2,720 individuals — independent of any curated list.

2

modules enriched for
autophagy genes (FDR < 0.05)

21

of 80 shortlisted proteins
in enriched modules

Key finding

~¼ of the panel doesn't just carry an autophagy label — these proteins co-vary together in CSF, suggesting coherent pathway-level biology.

Additionally, 9 hub proteins in enriched modules were not on R1 — data-driven candidates that emerged from co-expression, not curation.

Results

AD model cross-check — App knock-in mice

Cross-referenced the panel against CSF proteomics from App^NL-G-F and App^NL-F knock-in mice vs wild-type (3 pairwise comparisons, ~260 proteins each).

17

of 80 shortlisted proteins
present in AD model data

9

significantly altered
in ≥1 comparison (p < 0.05)

Dominant pattern: lysosomal enzymes are decreased in CSF of amyloid models — consistent with impaired lysosomal secretion in AD.

Gene	Rank	Direction	p-value	Comparison
CTSB	2	↓ ↓	0.017 / 0.002	NL-F, NL-F vs NL-GF
CPQ	5	↓	0.003	NL-F vs WT
HEXB	6	↓	0.028	NL-F vs WT
CTSZ	17	↓ ↓	0.025 / 0.012	NL-F, NL-F vs NL-GF
MAN2B1	15	↓	0.046	NL-F vs WT
CTSA	20	↓	0.014	NL-F vs NL-GF
CTSS	24	↓	0.009	NL-F vs NL-GF
CTSD	1	↓	0.017	NL-F vs NL-GF

↓ = significantly decreased in disease model. CTSB shows inconsistent direction across comparisons (up in NL-GF vs WT).

Sensitivity analysis

How much does EV evidence matter?

Test 1: Up-weight EV to 25%

72 / 80 proteins unchanged. ρ = 0.88. Top 10 stable.

8 lysosomal enzymes with low EV secretion swap out for 8 vesicle trafficking proteins (VPS35, DNM1L, RAB11B, ATP6V1E1) — biologically expected in EVs.

Test 2: Hard EV gate

Require EV detection → 122 drops to 102 candidates. Top 80 loses 6 proteins.

Three of those 6 — CPQ, CTSS, CTSZ — are significantly altered in the AD mouse model. The gate trades disease-validated hits for marginal EV-positive candidates.

Lysosomal enzymes reach CSF via direct secretion, not EV packaging. An EV gate creates a blind spot for exactly the biology we care about. Low EV weight (5–10%) is the right call.

Sensitivity analysis

Stability across all parameter choices

Pipeline tested under 9 configurations varying tier stringency and scoring weights (plus the EV-heavy reweighting above).

0.89

Mean Spearman rank correlation across configs (range 0.79–0.98)

65 / 80

Mean candidates retained across all configurations

Always in the top

These proteins survived every configuration:

CTSD · CTSB · CST3 · CPQ
PARK7 · HEXB · GAA · LAMP1

Summary

What we have

An 80-protein CSF autophagy/lysosomal panel that is robustly detected across both mouse and human, biologically coherent by independent co-expression analysis, stable across analytical choices, and directly amenable to translational targeted assay development — with the majority of candidates measurable using shared peptide sequences across species.

80

panel proteins

65

Tier I assay-ready
(shared peptides)

21

module-validated
(co-expression support)

9

novel module-derived
candidates

Novel candidates

Module-derived candidates — data over curation

Co-abundance analysis clustered all Astral proteins — not just those on the autophagy list. Two modules were enriched for R1 genes. Within those modules, 9 hub proteins were not on R1 but co-vary tightly with known autophagy/lysosomal markers.

These are data-driven hypotheses: proteins the co-expression network says behave like autophagy markers in CSF, despite not being on any curated list. They did not go through the autophagy filter — they emerged independently.

Gene	Module	Connectivity
SELENBP1	824	0.54
LGALS3BP	824	0.53
ADGRL1	824	0.53
GOT1	824	0.52
AEBP1	949	0.51
DCN	949	0.51
F5	949	0.50
COL14A1	949	0.47
CPVL	949	0.47

Next steps

Where do we go from here?

🔬 Assay development

PRM/MRM assays for top Tier I candidates using conserved peptides. Alternatively, assess antibody availability for Olink, ELISA, or suspension bead array assays. Validate in independent CSF cohorts.

💊 ERAP integration

Measure panel in rapamycin-treated CSF samples. Test whether mTOR inhibition produces detectable shifts in autophagy/lysosomal protein panel.

🧪 Novel candidates

Follow up the 9 module-derived proteins. Characterise their relationship to autophagy/lysosomal biology and assess assay feasibility.

CSF Autophagy & LysosomeBiomarker Panel Discovery

Which autophagy/lysosomal proteins can we reliably measure in CSF across both species?

Pipeline architecture

Input data

🐭 Mouse (9 datasets)

🧬 Human (3 datasets + references)

Quality control

Sample filtering

Blood contamination

Detectability tiers

Orthology mapping — mouse → human

Paralog families

Evidence scoring

Core panel criteria (hard filters)

Results

From proteome to panel

Top 10 candidates

Detection robustness (of 80 shortlisted)

Peptide feasibility — can we measure these?

What this means

Co-abundance modules — does the panel reflect real biology?

Key finding

AD model cross-check — App knock-in mice

How much does EV evidence matter?

Test 1: Up-weight EV to 25%

Test 2: Hard EV gate

Stability across all parameter choices

Always in the top

What we have

Module-derived candidates — data over curation

Where do we go from here?

🔬 Assay development

💊 ERAP integration

🧪 Novel candidates

CSF Autophagy & Lysosome
Biomarker Panel Discovery