Progress update #5: Refinement and redesign

Previous post showed Track A advancing with defined redesign areas and Track B's GO with 18 cross-category sites.

Objective: To stress-test both tracks, fix what breaks and redesign what carries residual assumption risk before any experimental capital is committed.

Work completed:

(i) A formal response to five expert critiques of Track A, each answered with new computation.

(ii) A blank-slate re-investigation of Track B: zero annotation dependence, experimental structures, a real fifth-category complex, a negative control.

(iii) One threshold-calibration failure corrected in the first-pass atlas.

Execution and Evidence

Track A (Critiques):

• Critique 1 — "true synergy versus additivity." Median safety CI ≈ 0.966 vs therapeutic 0.802. Loewe CI < 0.80 means only 80% of the predicted additive dose is needed (dose-reduction beyond additivity), not "both drugs working at once." Verdict: GO pending checkerboard confirmation.

• Critique 2 — linker activation, resistance, differential warhead PK. CP essentiality is Tier A-experimental for all nine Tier 1 targets; non-essential-CP indications (Gram-negatives, most fungi, all viruses) deprioritized. PBPK Monte Carlo overturned the portfolio-level GREEN: the TME holds the synergy window with 96.3% probability (median ratio 1.22 at 6 h), but neutral compartments fail (spleen/macrophage 0.37; brain, liver, lung, plasma ~0.30). Solid tumours GO; M. tuberculosis MODIFY pending a fixed-ratio macrophage assay; GBM NO-GO for systemic dosing without a BBB-rescue strategy. Resolvable engineering, not platform failure.

• Critique 3 — selectivity panel breadth. SEA expanded against ChEMBL 35 (2.3M compounds; 14,853 targets) with Vina confirmation. Five RED off-targets need P0 counter-screens: 1st warhead — DYRK1A (~6×), CLK1 (~14×), EGFR (~20×); 2nd warhead — Transferrin R1 (~5×; most critical, with Fe³⁺-competition assay and serum iron/ferritin monitoring), HDAC6 (~19×). CYP3A4 YELLOW (~71×); PLP-enzyme promiscuity and hERG GREEN. DYRK1A, EGFR and HDAC6 are validated oncology mechanisms, so possible polypharmacology in TNBC/AML is to be characterized, not assumed harmful.

• Critique 4 — synthetic lethality versus essentiality. P(SL) was decomposed and a three-tier nomenclature enforced: DUAL ESSENTIAL + SYNTHETICALLY LETHAL (≥ 0.80; 8 targets), STRONG ESSENTIALITY (0.60–0.79; 24), INDIVIDUAL ESSENTIALITY ONLY (< 0.60; 18). All 12 viral targets (P(SL) 0.38–0.45) were reframed as host-directed antiviral therapy: mechanism-valid but not effectively targeted by this prodrug, so excluded from the priority tier.

• Critique 5 — broad-spectrum credibility. Organism-specific penalties replaced any uniform-potency claim (Gram-negative 15–25×, CNS 15×). 32/62 REALISTIC (AUC/IC₅₀ ≥ 2.0), 26/62 BORDERLINE REALISTIC, 4/62 CHALLENGING and excluded; GBM 0.80 (needs BBB delivery), P. aeruginosa 0.60, rabies 0.62, A. baumannii 0.55. 9 Tier 1 targets pass every gate: M. tuberculosis, TNBC, ovarian, RCC, HCC, AML/MDS, Plasmodium, ER+ breast, colorectal.

Track B: the blank-slate redesign

• Why redesign. The first-pass atlas carried three assumption layers: curated UniProt annotation defined the active site, MAFFT anchored the alignment and all structures were predicted monomers, with no experimental bacterial structures, viral-context complex, or negative control. A threshold-calibration failure was also caught: the lenient threshold called >90% of sites "present" in every category and was replaced with a stricter, disclosed one.

• The re-run, geometry-only. Sixteen fresh structures from RCSB and AlphaFold: five experimental bacterial SufS, three fungal and three parasitic models, three human states, the ten-chain host ISC assembly 6NZU (viral category) and T4 lysozyme 2LZM (negative control). Detector reimplemented from scratch (LIGSITE-style grid + Fpocket-style clustering; Biopython alignment replacing MAFFT; outlier-pruned Kabsch; PROPKA3 and FreeSASA).

• Findings. 438 raw pockets, 203 peptide-viable (46%); only 13 (3%) active-site-adjacent. Core RMSDs 0.28–1.80 Å across 22.5–88.0% sequence identity. 13 cross-family site slots; exactly one ('Site Slot 3', 38 member pockets) reaches all five categories with dual-method equivalence (geometric fingerprint AND sequence-guided centroid + Jaccard).

What changed: Track A has testable redesign levers in place of open questions. Track B has one candidate that survived 2 independent tests across all five categories.

Current interpretation: Both tracks are better defined and confidence has moved from "candidate" to "prioritised for experiment", so what remains is the boundary between computation and experiment.

Next step: The refined results indicate a redesign-ready prodrug and a conserved groove, leading to the current conclusion in the final post.