Kcat-Km Version 1
Offline CPU prediction of log10_kcat and log10_km from id,smiles,sequence.
Two fixed ExtraTrees forests use 2,148 features: the original starter features,
sequence summaries, 16 RDKit descriptors with missing flags, and radius-2,
2,048-bit chiral Morgan fingerprints. Descriptor medians and forests are refit
on all 36,970 public training rows only. Public-validation inputs are not used
in fitting. Targets and input strings are preserved; predictions stay logged.
python /app/predict.py --input INPUT --output OUTPUT
INPUT is Parquet with exactly id,smiles,sequence in that order. OUTPUT is CSV
with exactly id,log10_kcat,log10_km, each ID once, and finite bounded values.
All weights/preprocessing assets are local; inference uses two workers and
needs no network, credentials, GPU, or host-side source files. Dependencies are
pinned for Python 3.11.9; the Dockerfile targets Linux x86-64.
The earlier, separately trained local-split candidate measured Mean SCC 0.610338, versus ridge 0.336433. That exact-sequence-disjoint holdout includes close sequence variants. These are model-selection scores, not a new holdout score for the full-training weights or a public leaderboard result.
Local verification evidence, full-refit runtimes and package hashes are kept
outside this package under experiments/full_refit_v1/. Native Windows checks
do not establish Linux Docker compatibility; consult that record for actual
container-test status. Publication and leaderboard submission require separate
user authorization.