Protein residue-contact prediction from amino-acid sequence alone. We train models from scratch on MarinFold's curated AF2 and ESMFold2 contact maps, without MSAs or pretrained protein language model features as inputs.

The goal is to find architectures that improve contact R-precision and compare them with MarinFold's autoregressive LLM approach. Predictions can be passed to Helico for structure generation.

• Baseline: a 40M-parameter bidirectional sequence encoder with a symmetric pair network using convolutions and triangle multiplication.
• Architecture search: Codex proposes model code changes; isolated Iris jobs train candidates at batch priority. Experiments use matched training exposure, two-seed confirmation, a bounded compute allocation, and a persistent results ledger.
• Evaluation: MarinFold's frozen experimental splits and unmodified scoring code. Contacts are native-amino-acid ConFind degree >= 0.001 at sequence separation >= 6. Only the 97 validation proteins guide search; the 217 test and 19 de novo proteins were held out until final model selection.

The first training and evaluation campaign is complete. R-precision:

¹ Decontaminated exp232 m2-p06, step 145199. Sparsa remains substantially worse; model size, training compute, and readout differ. Eight architecture-search trials completed; none met the promotion rule. See results and checkpoint and search results. See also the longer scaling results and current pair-trunk experiments.

uv sync
uv run pytest -q
uv run sparsa predict --checkpoint /path/to/checkpoint.pt \
  --sequence ACDEFGHIKLMNPQRSTVWY --out outputs/example --top-l 1

Prediction writes a dense score matrix and a zero-based, positive-only Helico contact list. Architecture-search checkpoints require their accompanying source snapshot. Training requires Iris and access to the curated CoreWeave S3 data.

Architecture search and controls · Training and evaluation guide · Benchmark provenance · Discrete-diffusion screen · Long diffusion campaign