Compress, don't memorise
We aim for compact mechanisms that look like the laws of physics, not for embeddings that look like the data.
Science tools
Holobiont
Compressing the laws underneath protein dynamics into physically-invariant signals.
Find the invariant, not the lookup table
Holobiont is a science program built around a specific idea: there are physically-invariant signals underneath protein dynamics that compress better than any amount of memorisation. We pursue those signals, audit them ruthlessly against retrieval-axis leakage, and publish the negatives when an attractive feature turns out to be measurement artifact.
most pipelines
add features until the leaderboard moves and ship the leaderboard, even when the lift is structural artifact.
holobiont
biases the pipeline toward features that look like physics — compact, invariant, auditable — and retires the rest before they reach production.
Three commitments
Audit retrieval axes for leakage
Pre-computed retrieval axes are notoriously easy to leak test labels through. We default to nested cross-validation and report leakage budgets explicitly.
Publish what does not work
Where a feature class hurts performance once leakage is removed, we say so. The product is the mechanism, not the leaderboard.
Where a tempting feature class quietly turns negative
Column statistics — Shannon entropy and amino-acid frequency — look attractive at low coverage. Past a coverage budget around the middle of the chart, they begin to hurt the best estimator. We retired them for production use.
What Holobiont has decided
Each decision is a measured statement, not a marketing one. Where a result hurt, the result and the retirement live on the page.
-
D15.1 Retrieval-axis leakage quantified
A large fraction of an apparently strong feature's lift was attributable to leakage through pre-computed k-NN axes. The methodology was tightened accordingly.
D15.1 83 % of apparent lift was retrieval-axis leakage -
D22 Conservation features retired at coverage
Shannon and frequency column statistics hurt our best estimator beyond a certain coverage budget. We retired them for production use and kept exploring direct-coupling pairs.
D22 Δρ −0.0382 on the best estimator above the coverage cliff -
Today Mechanism-first feature pipeline
The current pipeline is biased toward features that look like physics. Each candidate has an explicit leakage audit and an explicit coverage budget before it ships.
today physics-shaped features only, with explicit leakage audit
Why this is its own program
Protein dynamics is the obvious place to put pressure on the "compress, don't memorise" idea. Holobiont is where we run that pressure — and where we have already retired features that looked attractive but did not survive the audit.
cross-cuts
Holobiont is where the compress-vs-memorise idea gets stress-tested
The mechanism-first audit travels across our research programs — from evaluation discipline through to alignment posture. Where a feature does not survive here, it does not ship anywhere.