Machine Learning
CORTEX-MS
Learned molecular reasoning architecture that maintains and updates competing molecular hypotheses from mass-spectrometry evidence rather than one-shot retrieval.
Research question
Can molecular identification from mass-spectrometry evidence benefit from a learned architecture that maintains and updates competing molecular hypotheses rather than treating identification solely as one-shot classification or retrieval?
Method & architecture
Cortex-MS transforms raw tandem mass spectra (MS/MS) and precursor metadata into structured geometric peak embeddings. These embeddings are routed into an active workspace where competing hypothesis slots evaluate formula consistency, functional fingerprint predictions, and candidate ranking in parallel.
- Hypothesis Competition: Rather than collapsing features in a single forward pass, parallel hypothesis slots compete for evidence support.
- Geometric Peak Manifold: Peak m/z, relative abundance, and collision energy are projected into continuous metric spaces.
- Calibrated Uncertainty: The system preserves candidate margin ambiguities instead of forcing artificial overconfidence.
Results & findings
L4 Saturation Non-Linearity
OBSERVEDSmall-model molecular ML workloads do not automatically saturate an L4 simply by increasing batch size.
Hard Negative Inversion
OBSERVEDThe current contrastive representation has configurations where hard negatives score above true positives.
Host Pipeline Bottleneck
OBSERVEDInput and data-pipeline behavior can become a meaningful constraint relative to accelerator compute capacity.
Transformer Baseline Advantage
UNRESOLVEDWhether the Cortex hypothesis workspace provides an advantage over a properly matched standard Transformer baseline.