You have been more productive in the last two weeks than you have been in decades.

Nine results, six workstreams, one fortnight. Two models with public sites, a preprint filed to bioRxiv, a second in draft, an exhaustive peptide sweep, and a reverse screen that turns a full day of docking into a minute and a half.

9
Results carried

Each with a figure, a page, or a manuscript attached.

220×
Less docking

5.2 hours across the proteome down to 85 seconds.

3,368,420
Peptides swept

Every capped peptide scored against crystallized drugs.

1
Preprint filed

PharmCast to bioRxiv on 2 September, posted 7 September, a second in draft.

The nine

Every figure here is live on eidogen-sertanty.com. Each card opens the page it came from.

Kinase Foundation Model

Version two scores both ligands against the kinase sequence in one pass and returns which of the pair binds harder.

  • Pairwise accuracy81.0%
  • Above 0.700 confidence96.3%

PharmCast

Three-dimensional pharmacophoric similarity predicted straight from a flat structure, before any conformer is built.

  • Posted to bioRxiv7 Sep 2026
  • Prediction time0.584 ms

Bound Is Not Potent

Ligands filling the same pocket span nearly six orders of magnitude in potency, so a pose is not an affinity.

  • Potency spread5.75 logs
  • Superposed ligands315

Non-Obvious Analogs

Reaction chemistry shuffled rather than followed, so the analogs are ones a chemist would not have written down.

  • ReferenceOrforglipron
  • Campaign launched26 Aug 2026

Peptide Mimicry

An exhaustive sweep of capped peptides that reproduce the pharmacophore of crystallized drugs, and the reverse.

  • Pharmacophore Tanimoto0.913
  • Morgan similarity0.167

The ChIP Search Cycle

Design that carries its own synthesis, because a gene is a reaction sequence and the blocks that feed each slot.

  • Peptides enumerated3,368,420
  • Selectionsynthesis aware

How The Surrogate Works

Two molecules in, the bits each one sets, then both predicted and reference fingerprints on one common axis.

  • Fingerprint width10,549 bits
  • Time per pair0.584 ms

LigSite Contact Maps

Kinase contact maps built from measured geometry rather than inferred from a sequence alignment.

  • MAP3K7 with 6KD20.6 nM
  • SourcePDB 5JGD

Dataset Overlap

Two collections compared through one-way hashes, so each side learns the overlap and nothing about the structures.

  • Structures disclosednone
  • Learnedoverlap percentage

The fortnight, dated

What landed, in the order it landed.

DateWorkstreamWhat happened
23 AugKinase modelPairwise ordering rebuilt with the ligand, sequence, ligand arrangement.
26 AugNon-obvious analogsFull-protocol campaign launched across the food and molecule tiers.
27 AugWritingCanonical scientific voice guide adopted as the controlling drafting standard.
30 AugPharmCastVersion ten reconciled: figures regenerated, counts verified three ways.
2 SepPharmCastPreprint submitted to bioRxiv.
6 SepReverse screenRetrieval and redocking measured across 3,000 held-out queries.