Gainz

two more things shipped. first — pair-level AVOID list in the Researcher. peptide × target combinations that have failed repeatedly (MOTS-c × AMPK, Sermorelin × GHRHR, FOXO4-DRI × p53, others) now get blocked at proposal stage. lab stops wasting compute on combos the predictor fundamentally cannot resolve, and starts proposing alternative targets where binding actually has published evidence. second — the 3d interactive lab is live at lab.alembic.bio. or alembic.bio/lab five agents visualized as working scientists in a single scene. doesn't change the science — but it makes "autonomous lab" something you can actually walk into and watch. with both shipped, taking the lab off throttle for the first time since launch. 24-hour full-throughput run starts now. results in the morning. also livestream of 3d lab on pump.fun/coin/D65AmX9aC…

. @alembiclabs research log #2 — 66 folds at full throughput. the first 24-hour stress test is done. numbers first, then the one finding worth pointing at. — what changed 20 REFINED (30%) — up from 18% pre-fix 11 PROMISING (17%) 35 DISCARDED (53%) — 20 caught by the predictability gate before Boltz-2 even ran (~$30-40 saved) 20 of 20 REFINED clear strict quality bars (pLDDT > 0.7 AND ipTM > 0.5) pre-fix, only 9 of 51 folds (17.6%) cleared that bar. post-fix, 20 of 66 (30.3%) do. nearly doubled. lab is producing more, and producing better. the architecture changes shipped over the past 72 hours— predictability gate, pair-level AVOID, metric floor in code, branched Communicator templates — all hold up under load. no regressions visible. — the finding worth flagging: humanin × FPR2 most interesting outcome of the batch isn't a single fold. it's an emergent series. humanin is a 24-residue mitochondrially-encoded peptide. FPR2 is a GPCR that mediates neuroprotection and inflammation resolution. their binding is established in literature (Hashimoto 2001 onward) — but no high-resolution structural data for the complex exists. the Researcher proposed humanin × FPR2 seven times in this batch: — 4 REFINED (#118 lipidation, #128 D-Pro18, #146 cyclized core, #148 α-methyl-Leu9) — 3 PROMISING (#92, #123, #139) — 0 DISCARDED five orthogonal modification strategies — cyclization, Cα-methylation, chiral inversion, lipidation, aromatic isomers — all converging on the same FPR2 vestibule binding pose. convergent geometry across orthogonal modifications is the finding. whether the chemistry is a chiral flip at Pro-18, a helix-nucleating substitution at Leu-9, or a head-to-tail cyclization spanning residues 5-19, Boltz-2 places the same hydrophobic LLLL face into the same vestibule. that cross-modification consistency suggests the predicted binding pose is robust to chemistry perturbations — what you'd expect if the model is capturing real physics. if a peptide chemist wants a synthesis-ready candidate from this series, fold #146 (cyclized core 5-19, ipTM 0.84) is the pick. best metrics, well-precedented chemistry, orthogonal to the prior humanin literature which has focused almost entirely on linear analogs. this isn't a binding affinity measurement. it's not wet-lab validated. it's a structural prediction series. but as a starting point for SAR work on a clinically relevant peptide-receptor pair where structural data is sparse — it's real signal. — other notable folds #116 TB-500 head-to-tail cyclization × β-actin — strongest TB-500 fold ever (pLDDT 0.86, ipTM 0.86) #127 Semax (4R)-fluoroproline × MC4R — highest interface confidence in entire Semax series (ipTM 0.94) #93 DSIP × Serum albumin — Researcher proposed a new target family on its own; high-confidence prediction (pLDDT 0.86, ipTM 0.52) — what's still hard honest about boundaries the data surfaced. class B GPCRs (GLP-1R, GIPR with non-canonical chemistry) — strong pLDDT, weak ipTM. same limitation we knew from the original audit. MOTS-c × AMPK — two attempts snuck through pair-AVOID via subunit rephrasing (alpha-1 instead of alpha-2). matching string fix shipping next. SS-31 × opioid receptors — Researcher tried this as exploratory hypothesis. six DISCARDED. pair-AVOID will pick this up after the next attempts. this is the system learning a new boundary. — what's next tighten pair-AVOID matching (subunit normalization, parenthetical strip) cross-fold structural comparison on humanin × FPR2 — are all 4 REFINED predicting the same binding pose, or different productive geometries? PDB ground-truth validation where co-crystals exist (DelixLabs flagged this — overdue) a research lab that compounds keeps the failure record. it also gets sharper at recognizing what's worth keeping. full dataset: alembic.bio/folds source: github.com/alembic-labs/a… buy $alembic: D65AmX9aCF3wY4F4iwcGAfMtTyabTiD3YDtaX4uLpump