Sofia Bousseta

This thesis demonstrates that single-molecule translocation events recorded on the Pipette- Elastomer Interfacial Nanopore (PEIN) platform can be turned, by an unsupervised analyt- ical workflow, into a quantitative molecular fingerprint of complex neuronal supernatants. The PEIN platform forms a contact pore between a pulled glass nanopipette and a PDMS membrane. Two Neurobasal-medium supernatants from parallel neuronal cultures were com- pared under identical recording conditions: both contained the same supplements, and the “new” aliquot (DIV4) additionally contained doxycycline. The biologically meaningful con- trast between them is their neuronal conditioning time — DIV4 (4 days) versus approxi- mately three months and ten days in vitro — with both aliquots frozen on the same day. The expectation is that the old aliquot contains a mixture of consumed and degraded sup- plement components together with molecules secreted by mature neurons over time, while the new aliquot remains enriched in the freshly added, intact supplement components.

A total of 12,411 events at ±600 mV were each reduced to a 20-dimensional feature vector (10 physical features plus 10 audited waveform features), projected with PCA and then UMAP, and grouped with HDBSCAN. The pipeline never sees the supernatant or voltage label until after clustering. Despite this, the 28 clusters recovered cleanly separate the two supernatants: seven are∼100% new-dominant, thirteen are∼100% old-dominant, and only eight are mixed. The clustering result alone — before any biological interpretation — already establishes that the two media produce reproducibly different translocation-event populations.

The strongest single signature is cluster 6 (3985 events, 99.8% new,∼34% of the total dataset), a compact, short-dwell, conductive population whose physical signature is consis- tent with a small folded protein and whose biological interpretation — intact BDNF / GDNF and other fresh growth factor supplements — matches the conditioning-time contrast. Anditional long-dwell aggregating population (cluster1) appears almost exclusively at −600 mV. A decision tree assigns candidate molecules to each cluster: intact growth factors and small supplement components dominate the new-dominant clusters; degraded protein, amino-acid pool, baseline artefacts and adsorbing / aggregating populations dominate the old-dominant ones — exactly the fingerprints expected from the consumed-versus-fresh interpretation of the two media. A within-recording analysis identifies a methodological caveat with direct implications for the field: the apparent internal structure of cluster 6 reflects pipette-resistance drift during a single recording rather than multiple molecular species. The pipette-normalised feature ∆Irel is robust to this drift; absolute amplitudes are not. This motivates a clear reporting recommendation for any future PEIN study, including the on-cell measurements toward which this work is one stage on the route.

The full pipeline, the 28-cluster catalogue, the candidate-molecule decision tree, and all figures are released as a self-contained reproducible package.