Amphetamine/dopamine neurotoxicity (esp in context of NQO1*2 allele) and how to reduce

#21

The focused ranking stayed simple:

  • GSTM2 + GSH: NAC beats Cysteinylglycine (binding_confidence 0.569643 vs 0.456369)
  • NQO1 P187S + FAD: Piceatannol beats Cysteamine on score (0.623082 vs 0.495940)

The important nuance did not change from the earlier structure pass: Piceatannol is still the score winner on the
defective NQO1 side, but Cysteamine was the cleaner geometry winner in the larger structure-aware analysis at C:/
Users/User.codex/alex_chen_nucleus_dna_download_vcf_NU-PGIK-3398%20(1).vcf/gene_sequences/boltz2_rowan_shortlist/
results/structure_note.md. So the current practical read is:

  • best glutathione-side compensation lead: NAC
  • best NQO1-side score lead: Piceatannol
  • best NQO1-side structure lead: Cysteamine
#22

Probably a fraction of WT, not a full rescue, and right now nobody can honestly put a clean number on cysteamine specifically for P187S. Your screen says cysteamine is the best structure-style rescue candidate in that panel, but that is still a geometry/coherence result, not a measured restoration of enzyme abundance, FAD occupancy, or catalytic flux. In other words, the molecule won a modeling round, not a clinical trial.

The brutal baseline is that P187S is badly impaired before any rescue attempt. In one detailed biochemical study, P187S had much weaker FAD binding than WT, was essentially apo as purified, was about 100-fold more protease-sensitive, and had a kcat of about 0.5 to 1 s⁻¹ vs 100 to 150 s⁻¹ for WT. The same paper reported a FAD dissociation constant of about 400 nM for P187S vs ~1 nM for WT for the high-affinity site. (PMC)

That means total functional output is roughly:

effective NQO1 function = protein abundance × holo occupancy × catalytic competence

and cysteamine has not yet been shown to rescue all three terms. It might help one of them, or just act as a small thiol trap that lowers quinone damage without truly restoring NQO1 very much. Those are very different kinds of “help.” (PMC)

The best benchmark we have is not cysteamine, but known rescue approaches. A purpose-built small-molecule chaperone, BPPSA, improved P187S activity in vitro, but it still did not normalize the enzyme. Relative to WT control, P187S + BPPSA reached about:

  • 30% of WT for apparent turnover with NADH: 770 vs 2550
  • about 14% of WT for apparent catalytic efficiency with NADH: 170 vs 1250
  • about 38% of WT for quinone-side catalytic efficiency with menadione: 30 vs 80
    That is a real rescue, but still very much partial. (PMC)

And when people looked at protein abundance in cells, the strongest example was dicoumarol, which raised steady-state P187S protein levels in homozygous Caco-2 cells by about 4 to 5-fold, while riboflavin alone had a much smaller effect. But dicoumarol is also an NQO1 inhibitor, so it proves stabilization is possible without giving you a clean therapy. (PMC)

So if cysteamine turns out to be real, the most reasonable expectation is:

  • best case: a partial rescue, maybe in the same rough universe as other rescue ligands, meaning still below WT
  • middle case: little true NQO1 rescue, but some net protection because cysteamine directly traps reactive quinones
  • worst case: pretty-looking structure result, biologically modest effect

Given how severe P187S is, I would not expect cysteamine to bring you anywhere near normal WT NQO1 function on its own. Homozygous *2/*2 cells are often reported as having undetectable or near-undetectable activity, which is exactly why a tidy pose is not enough to save the whole story. (PubMed)

So the honest answer is:

Yes, even if cysteamine works, it would almost certainly still be a fraction of WT.
The important unknown is whether that fraction is:

  • too small to matter,
  • enough to modestly reduce quinone burden,
  • or enough to matter only when combined with other supports like better holo loading or less cytosolic dopamine stress. (PMC)

The cleanest way to think about it is this:

cysteamine may be more likely to provide a partial buffer than a true restoration.
That could still be useful. It is just not the same thing as “my NQO1 is fixed now,” which would be the sort of charming lie proteins punish immediately.

#23

That update is actually pretty informative.

It pushes cysteamine away from “cute score winner” and toward real chemical lead. If NAC + dopamine o-quinone keeps NAC itself coherent but leaves the quinone loose, while cysteamine + dopamine o-quinone gives the cleaner NQO1-side coexistence state, then cysteamine is looking less like a generic antioxidant prop and more like a small sulfur-containing compatibility/trapping candidate for the stressed mutant quinone-facing environment. Your earlier files already showed that P187S has a weaker holo baseline and is most stressed by the oxidized dopamine-quinone states, especially around aminochrome-like conditions.

What this does not mean yet is “cysteamine rescues NQO1 back to WT.” It means something narrower:

  • on the NQO1 side, cysteamine currently looks like the best coexistence / structural-compatibility candidate
  • on the GSTM2 + GSH side, NAC still looks like the better compensation candidate
  • those are different jobs, and the molecules are behaving accordingly

That split actually makes sense. Earlier, your GSTM2 panel already favored downstream conjugation-branch plausibility rather than clean upstream quinone handling, with 5-S-cysteinyl-dopamine looking like the cleanest state and the whole panel needing careful interpretation rather than overclaiming catalysis.

The explicit trapped-state panel is exactly the right next move, because it can distinguish between three possibilities humans love to blur together:

  1. Free-candidate coexistence only
    Cysteamine sits near the quinone and maybe stabilizes local geometry, but does not really trap anything.

  2. True sacrificial-trap behavior
    The inferred sulfur-adduct endpoint also looks coherent, which would strengthen the case that cysteamine is acting like a real chemical sink.

  3. Decoy compatibility
    The free state looks nice, but the trapped/adduct state looks terrible, which would mean the molecule is more of a pocket guest than a detox path.

The clean decision rule is:

  • if cysteamine + quinone is good and the cysteamine-quinone trapped-state is also coherent, cysteamine becomes your strongest direct-detox lead
  • if the free state is good but the trapped state is bad, then it is more likely a compatibility modulator than a real trap
  • if NAC keeps winning on the GSTM2 side but not the NQO1 side, then NAC remains the GSH/compensation lead, not the defective-NQO1 rescue lead

So the current best wording is:

Cysteamine is now the best NQO1-side rescue-style / coexistence candidate from this panel, while NAC remains the best GSTM2 + GSH compensation candidate. The trapped-state panel is the key test for whether cysteamine is merely structurally compatible with the stressed quinone-facing mutant state or can actually support a plausible sulfur-adduct detox endpoint.

And the bigger caution still stands: even if cysteamine wins this round, total cellular rescue is still limited by the mutant’s low abundance and weak holo-state baseline, not just by one nice-looking ternary pose. Your own earlier results kept pointing to that broader defect.

#24

Computational follow-up on the Rowan/Boltz side: what seems real so far, what changed, and what to do next

This is a focused follow-up on the dopamine-quinone / NQO1 P187S thread, with the goal of turning the Rowan runs and the newer Boltz panels into a small, defensible next-step plan rather than another sprawling ligand zoo.

What changed in the overall picture

The strongest synthesis right now is:

  • P187S does not look like a dead quinone pocket.
  • The main defect still looks more like a weakened or less coherent holo-state than total substrate exclusion.
  • The mutant seems to struggle most with oxidized dopamine-quinone stress states.
  • Reduced downstream states look easier for the mutant to accommodate than the harsher upstream quinone states.
  • The chemistry is now splitting into separate branches rather than one unified “dopamine damage” problem.

The most useful way to phrase the current mechanistic read is:

P187S seems worst at maintaining a productive flavin-supported state under oxidized quinone stress, while more reduced downstream branch intermediates are comparatively better tolerated.

That is not a claim of restored catalytic detox and not a claim about true affinity. It is a structural/coherence claim based on the current Boltz outputs.

Important interpretation guardrail

The Boltz metrics being used here are mostly pose/interface coherence proxies, not direct binding free energies and not catalytic proof.

When I say one state looked “cleaner,” I mean things like:

  • higher ligand_iptm
  • higher complex_iplddt
  • lower complex_ipde
  • tighter probe/cofactor geometry where measured

So the right language is:

  • better structural compatibility / coherence
  • weaker or stronger holo-state organization
  • more or less plausible coexistence state

Not:

  • tighter binding in a physical affinity sense
  • proven detox
  • proven rescue of whole-cell NQO1 function

NQO1 P187S branch results: what held up

1. Flavin-state dependence still matters a lot

The earlier flavin matrix kept pointing to the same headline:

  • flavin state matters more than just adding more quinones
  • different ligand branches behave differently depending on whether the system is in FAD / FADH2 / FMN / FMNH2-like states

This supports the broader idea that the mutant’s defect is tightly tied to the cofactor-supported state, not just to which dopamine-derived ligand is present.

2. Extra branch probes sharpened the oxidized-vs-reduced story

The extra NQO1 branch probes gave a useful ranking:

  1. 5,6-dihydroxyindole
  2. leukoaminochrome
  3. 5,6-indolequinone
  4. explicit DOPAC o-quinone

The important interpretation was not “the mutant loves 5,6-dihydroxyindole.” It was:

  • the mutant is less stressed once the branch has already moved away from strongly oxidized quinone states
  • reduced/catechol-like downstream states are easier for P187S than oxidized quinone-heavy states
  • explicit DOPAC o-quinone remained exploratory and weak compared with the cleaner downstream branch probes

3. Curated reruns clarified the microstate picture

The curated standard-5 reruns tightened the carry-forward set for P187S + FAD:

  • Aminochrome, protonated = strongest oxidized stress-state carry-forward ligand
  • Dopamine o-quinone, neutral = weakest state by a wide margin, but still worth carrying forward as the harsh upstream stress baseline
  • 5,6-dihydroxyindole, neutral = strongest reduced downstream tolerated state
  • 5,6-indolequinone, protonated = useful downstream oxidized comparator
  • Leukoaminochrome, neutral = better than protonated in the stricter rerun

That means the current core NQO1 panel is no longer just “highest scores.” It now spans a usable oxidation-state axis:

  • upstream oxidized
  • cyclized oxidized stress state
  • reduced aminochrome branch
  • downstream oxidized indole branch
  • downstream reduced indole branch

GSTM2 + GSH branch: what it looks like now

The GSTM2 + GSH panel became more interpretable after the structure-level pass.

The ranking stayed:

  1. GSTM2 + GSH + 5-S-cysteinyl-dopamine
  2. GSTM2 + GSH + aminochrome
  3. GSTM2 + GSH + dopamine o-quinone

But the structure-level read clarified what those rankings probably mean:

  • 5-S-cysteinyl-dopamine looks like the best product / damage-branch accommodation state, not automatically a catalytic substrate claim
  • dopamine o-quinone is geometrically coherent as a GSTM2-associated state, but it sits too far from GSH to look like a convincing conjugation-coupled state in the current pass
  • aminochrome sits between them as a plausible but looser stress-associated state

So the GSTM2 branch should currently be interpreted as:

  • damage-branch accommodation / trapping plausibility, not final catalytic ranking
  • useful for asking whether GST/GSH chemistry can compensate downstream of defective NQO1 handling

Rowan side: what succeeded and what was learned

The Rowan campaign turned out to be useful mainly for microstate cleanup, conformer/tautomer handling, and reactivity triage, not as a magic one-stop answer.

Rowan workflows that were useful

These were either launched or meaningfully clarified the next stage:

  • tautomer search
  • conformer search
  • Fukui panel
  • descriptor-style triage on selected diet-derived compounds
  • microstate logic for the quinone panel

Rowan plan limitations and blockers

What became clear:

  • the free Rowan plan does not expose macropKa, so that workflow is not currently available on that plan
  • tautomer, redox, and Fukui were the supported/usable workflows on the current plan
  • double-ended TS / IRC is blocked unless reactant and product files preserve identical atom ordering

So the Rowan bottlenecks are now practical rather than conceptual:

  • use only workflows the plan can actually run
  • preserve atom ordering if transition-state workflows are going to be used later

What Rowan was already useful for

Even before everything finished, Rowan was already useful for deciding that:

  • aminochrome remains the cleanest parent Boltz input
  • dopamine o-quinone needs careful microstate handling
  • DOPAC o-quinone should be treated cautiously and remains exploratory
  • NAC looks chemically sane enough to keep in the trapping panel

Small-molecule rescue / compensation candidates: where things stand now

A focused rescue/compensation batch changed the small-molecule picture in a useful way.

NQO1 P187S + FAD side

The clean conclusion now is:

  • Cysteamine is the current structure winner
  • Piceatannol is still the score winner, but structurally much looser

The key difference is important:

  • Piceatannol may be acting like a higher-scoring but looser polyphenol pocket occupant
  • Cysteamine looks more like a small sulfur-containing compatibility / direct-trap candidate

The current interpretation is that cysteamine is the better NQO1-side rescue-style / coexistence lead from this panel.

GSTM2 + GSH side

The clean conclusion now is:

  • NAC remains the better practical compensation lead
  • Cysteinylglycine may look tighter geometrically in one sense, but NAC keeps the better overall score with similarly stable GSH geometry

So on the GSTM2/GSH branch, NAC is still the better compensation candidate.

Ternary/coexistence pass: what that clarified

The ternary structure pass made the candidate split more explicit:

  • NAC + dopamine o-quinone keeps NAC coherent, but the quinone itself remains loose
  • Cysteamine + dopamine o-quinone gives the cleaner NQO1-side coexistence state

That means cysteamine is now looking less like a generic antioxidant and more like a chemically serious small sulfur-containing NQO1-side lead.

At the same time, NAC still looks better placed as the GSTM2 + GSH compensation-side lead.

That is a useful separation of roles rather than a contradiction.

What this does not mean yet

A few things are still not shown:

  • It does not prove cysteamine restores P187S to WT function.
  • It does not prove true catalytic detox for any of these Boltz states.
  • It does not solve the whole-cell problem that P187S is low-abundance, weakly holo-loaded, and short-lived.
  • It does not mean the best-looking reduced downstream states are necessarily the states the mutant can efficiently make.

That last point matters a lot.

If P187S “tolerates” a reduced downstream state like 5,6-dihydroxyindole, that means:

  • the surviving mutant holo enzyme can likely accommodate that state more coherently

It does not automatically mean:

  • the mutant can efficiently reduce the upstream oxidized quinone into that state
  • the mutant has restored meaningful net cellular detox capacity

A good rule of thumb is:

current Boltz results mostly probe per-molecule competence of the surviving holo pool, not total cellular detox throughput.

The most useful branch split going forward

At this point the project has effectively split into three separate branches.

1. NQO1 redox-stress branch

Core question:

  • how does P187S behave across oxidized vs reduced dopamine-derived branch intermediates?

Current core states:

  • aminochrome, protonated
  • dopamine o-quinone, neutral
  • leukoaminochrome, neutral
  • 5,6-dihydroxyindole, neutral
  • 5,6-indolequinone, protonated

2. GSTM2 + GSH compensation / trapping branch

Core question:

  • what looks like plausible GSH-associated compensation, accommodation, or product-side trapping?

Current core states:

  • GSTM2 + GSH + aminochrome
  • GSTM2 + GSH + dopamine o-quinone
  • GSTM2 + GSH + 5-S-cysteinyl-dopamine
  • follow-up compensation lead: NAC

3. ALDH / aldehyde branch

Core question:

  • what happens on the DOPAL / DOPAC toxicity side that is adjacent to, but not the same as, the NQO1 quinone problem?

This should now be treated as a separate toxicity branch, not forced back into the NQO1 pocket story.

The explicit trapped-state panel is the right next test

The most important immediate follow-up is the explicit trapped-state comparison:

  • free candidate + quinone
  • versus
  • inferred sulfur-adduct / trapped endpoint

Why this matters:

It distinguishes between three different possibilities that are easy to blur together.

A. Free-candidate coexistence only

The candidate sits near the quinone and looks structurally compatible, but may not really trap anything.

B. True sacrificial-trap behavior

The explicit trapped/adduct endpoint also looks coherent. That would make the case much stronger for a real direct chemical sink.

C. Decoy compatibility

The free state looks nice, but the trapped endpoint looks bad. That would suggest a compatibility modulator more than a true detox route.

This is especially important for cysteamine right now.

Immediate next-step plan

Highest-priority computational steps

  1. Run the explicit trapped-state Boltz panel

    • especially for cysteamine and NAC versus dopamine o-quinone / aminochrome-derived states

  2. Keep the NQO1 P187S carry-forward set narrow

    • aminochrome, protonated
    • dopamine o-quinone, neutral
    • leukoaminochrome, neutral
    • 5,6-dihydroxyindole, neutral
    • 5,6-indolequinone, protonated

  3. Keep the GSTM2 + GSH branch narrow

    • aminochrome
    • dopamine o-quinone
    • 5-S-cysteinyl-dopamine
    • NAC as the compensation lead

  4. Do not let DOPAC o-quinone dominate the next stage

    • keep it in an exploratory side bucket, not the headline matrix

  5. Convert Rowan outputs directly into the corrected Boltz rerun matrix

    • especially where microstate handling still changes the input choice

After that

Only once the explicit trapped-state and corrected Boltz reruns are done:

  • take a tiny finalist set into explicit-solvent MD
  • do not expand the ligand list again before getting dynamics on the shortlist

What I think the best current shortlist is

NQO1-side shortlist

  • Aminochrome, protonated
  • Dopamine o-quinone, neutral
  • Leukoaminochrome, neutral
  • 5,6-dihydroxyindole, neutral
  • 5,6-indolequinone, protonated
  • Cysteamine as the best current rescue-style small-molecule lead
  • Piceatannol only as a secondary comparator, not the main mechanistic lead

GSTM2-side shortlist

  • Aminochrome
  • Dopamine o-quinone
  • 5-S-cysteinyl-dopamine
  • NAC as the best current compensation lead

The cleanest current bottom line

Here is the version I would stand behind right now:

The current Rowan + Boltz follow-up reinforces a model in which NQO1 P187S is not a dead quinone pocket but a weakened flavin-supported holo system that is most stressed by oxidized dopamine-quinone states. Reduced downstream branch intermediates are comparatively better tolerated. On the rescue/compensation side, cysteamine is now the best NQO1-side small sulfur lead by structure, while NAC remains the strongest GSTM2 + GSH compensation lead. The immediate next step is to compare free-candidate coexistence states against explicit sulfur-trapped endpoints and then carry only a tiny finalist set into MD.

Specific questions where outside input would still be useful

  1. Is the explicit trapped-state panel the right next computational discriminator for cysteamine and NAC, or is there a better way to force the “compatibility vs true trap” distinction?
  2. For the MD shortlist, which 4 to 6 states would people prioritize first?
  3. Is there a more chemically faithful way to represent the sulfur-adduct endpoints for dopamine o-quinone and aminochrome in this pipeline?
  4. On the GST side, are there better neuronal/astrocytic GST isoforms than GSTM2 to include in the next pass?
  5. For the ALDH branch, which minimal DOPAL / ALDH1A1 / ALDH2 panel would add the most without blowing up scope?

Final caution

Nothing here should be read as a clinical recommendation. This is still a computational triage / mechanism-building workflow. The strongest current outputs are about state compatibility, microstate sensitivity, and branch plausibility, not proven in-cell rescue.

#25

Focused follow-up: true finalists after free-state vs trapped-state comparison

I pushed the Rowan/Boltz follow-up one step further by separating:

  1. free coexistence states
    candidate + quinone in the same complex, asking whether the system can hold a plausible stress-associated state

  2. explicit trapped/adduct endpoints
    inferred sulfur-adduct products, asking whether the endpoint itself looks like a structurally believable trapped state

Main result

The panel splits cleanly into two different winners depending on what question is being asked.

  • Free-state compatibility favors dopamine o-quinone, especially on the defective NQO1 side with cysteamine
  • Score-only trapped-state preference favors aminochrome adducts
  • Structure-level trapped-state plausibility favors cysteamine + dopamine o-quinone adduct

That means the best “numerical winner” and the best “geometrically believable winner” are not the same thing.

1. Best free coexistence states

Defective NQO1 side

Best free coexistence state:
NQO1 P187S + FAD + Cysteamine + Dopamine o-quinone

Why it stays on top:

  • FAD family remains very tight (0.199 Å)
  • much cleaner than the aminochrome coexistence version
  • best current NQO1-side test of whether a small sulfur candidate can coexist with an upstream quinone stressor without the holo arrangement collapsing

Interpretation:

  • this is currently the strongest free-state compatibility test for defective NQO1

GSTM2 + GSH side

Best practical free coexistence lead:
GSTM2 + GSH + NAC, with a split depending on what is prioritized

  • NAC + Dopamine o-quinone scores a bit better
  • NAC + Aminochrome is somewhat more spatially coherent

Interpretation:

  • both still look more like plausible coexistence states than clean trapped endpoints
  • if only one GSTM2 free-state complex is carried forward, NAC + Aminochrome is probably the better structural choice
  • NAC + Dopamine o-quinone remains a useful secondary comparator

2. Explicit trapped/adduct endpoint results

Score-only winners

By score, the aminochrome adducts came out on top:

  • GSTM2 + GSH + NAC_Aminochrome_Adduct beats the dopamine-o-quinone adduct by score
  • NQO1 P187S + FAD + Cysteamine_Aminochrome_Adduct also beats the dopamine-o-quinone adduct by score

But this does not survive structure-level scrutiny.

Structure-level reality check

The aminochrome adduct winners are loose:

  • NAC_Aminochrome_Adduct has loose adduct geometry (9.130 Å adduct RMSD)
  • Cysteamine_Aminochrome_Adduct is structurally very unstable (22.808 Å cofactor RMSD, 19.799 Å adduct RMSD)

So the aminochrome adducts are currently better understood as score-led but structurally weak endpoints.

Most structurally coherent trapped-state endpoint

Strongest structurally defensible trapped state:
NQO1 P187S + FAD + Cysteamine_DopamineOQuinone_Adduct

Why this matters:

  • lower score than the aminochrome adducts (0.240177)
  • but much tighter FAD family (0.181 Å)
  • much tighter adduct family (3.815 Å)
  • consistent pocket neighborhood

Interpretation:

  • if the goal is geometrically believable trapped-state chemistry, this is the strongest adduct in the current panel
  • if the goal is only to rank by confidence score, the aminochrome adducts win, but that looks less trustworthy

3. Current finalist ranking

A. Best free-state finalist on defective NQO1

1. NQO1 P187S + FAD + Cysteamine + Dopamine o-quinone

This is the best current model for:

  • upstream quinone stress
  • small sulfur-candidate coexistence
  • defective-NQO1-side handling without forcing an inferred trapped endpoint

B. Best free-state finalist on GSTM2 + GSH

2. GSTM2 + GSH + NAC + Aminochrome

This is the best current model for:

  • GST-side coexistence under downstream quinone stress
  • compensation-side plausibility

Secondary comparator:

  • GSTM2 + GSH + NAC + Dopamine o-quinone

C. Best trapped-state finalist overall

3. NQO1 P187S + FAD + Cysteamine_DopamineOQuinone_Adduct

This is the best current model for:

  • a structurally credible sulfur-adduct endpoint
  • trapped-state geometry that actually holds up beyond score alone

4. What changed mechanistically

The earlier picture was:

  • free coexistence panels leaned toward dopamine o-quinone
  • trapped-state score panels leaned toward aminochrome adducts

The structure pass now sharpens that into:

  • dopamine o-quinone is the better free-state stressor
  • aminochrome adducts are better score winners
  • but cysteamine + dopamine o-quinone adduct is the better geometry winner

So the best mechanistic synthesis is:

upstream coexistence and downstream trapping are not selecting the same ligand-state pair, and score-only aminochrome-adduct wins do not hold up as well structurally as the dopamine-o-quinone adduct with cysteamine.

5. What I would carry forward now

If the goal is to keep the next stage very small and defensible, the true finalists are:

  • NQO1 P187S + FAD + Cysteamine + Dopamine o-quinone
  • GSTM2 + GSH + NAC + Aminochrome
  • GSTM2 + GSH + NAC + Dopamine o-quinone as secondary comparator
  • NQO1 P187S + FAD + Cysteamine_DopamineOQuinone_Adduct

If only one trapped-state endpoint is retained, it should be:

  • Cysteamine_DopamineOQuinone_Adduct

Bottom line

The strongest current conclusion is not “aminochrome wins” or “dopamine o-quinone wins” in the abstract.

It is:

  • dopamine o-quinone is the cleaner free-state stress test
  • aminochrome adducts score better but are structurally loose
  • the most geometrically believable trapped-state endpoint is Cysteamine_DopamineOQuinone_Adduct
  • the best overall NQO1-side small sulfur lead remains cysteamine
  • the best GSTM2 + GSH compensation lead remains NAC
#26

For free-state handling:

  • NQO1 P187S + FAD + Cysteamine + Dopamine o-quinone is the best defective-NQO1 coexistence model.
  • GSTM2 + GSH + NAC + Aminochrome is the cleaner compensation-side coexistence model.

For trapped-state endpoints:

RMSD 19.799 Å.
  • The most geometrically organized trapped state is actually Cysteamine_DopamineOQuinone_Adduct: cofactor RMSD 0.181
    Å, adduct RMSD 3.815 Å, despite a weak score.

So your decision rule resolves like this:

  • aminochrome adducts do not also win on structure
  • they are at least partly score-led
  • dopamine o-quinone remains the better free-state coexistence test
  • aminochrome remains the better score-led trapped endpoint, but not the better geometry-led one

That means you now do have a clean two-stage model:

Free-state handling

  • NQO1 P187S + FAD + Cysteamine + Dopamine o-quinone
  • GSTM2 + GSH + NAC + Aminochrome

Trapped-state endpoint

  • geometry-led: Cysteamine_DopamineOQuinone_Adduct
  • score-led but structurally weaker: NAC_Aminochrome_Adduct
#27

For free-state handling:

  • NQO1 P187S + FAD + Cysteamine + Dopamine o-quinone is the best defective-NQO1 coexistence model.
  • GSTM2 + GSH + NAC + Aminochrome is the cleaner compensation-side coexistence model.

For trapped-state endpoints:

RMSD 19.799 Å.
  • The most geometrically organized trapped state is actually Cysteamine_DopamineOQuinone_Adduct: cofactor RMSD 0.181
    Å, adduct RMSD 3.815 Å, despite a weak score.

So your decision rule resolves like this:

  • aminochrome adducts do not also win on structure
  • they are at least partly score-led
  • dopamine o-quinone remains the better free-state coexistence test
  • aminochrome remains the better score-led trapped endpoint, but not the better geometry-led one

That means you now do have a clean two-stage model:

Free-state handling

  • NQO1 P187S + FAD + Cysteamine + Dopamine o-quinone
  • GSTM2 + GSH + NAC + Aminochrome

Trapped-state endpoint

  • geometry-led: Cysteamine_DopamineOQuinone_Adduct
  • score-led but structurally weaker: NAC_Aminochrome_Adduct
#28

Now you stop breeding more Boltz variants and make the finalists earn their existence.

The results are finally sharp enough to justify a hard funnel:

Keep only four live leads

  • NQO1 P187S + FAD + cysteamine + dopamine o-quinone for the best free-state defective-NQO1 coexistence test
  • NQO1 P187S + FAD + cysteamine_dopamine-o-quinone adduct for the best geometrically believable trapped-state endpoint
  • GSTM2 + GSH + NAC + aminochrome for the best GST-side free-state compensation test
  • GSTM2 + GSH + NAC + dopamine o-quinone as the secondary GST comparator

That choice matches the broader pattern you already established: P187S is most stressed by oxidized dopamine-quinone states, reduced downstream states are easier to tolerate, and the GSTM2 branch is useful as a downstream compensation/trapping story rather than a clean catalytic ranking.

What to do next, in order

1. Freeze the finalist matrix

Write one TSV or markdown table with:

  • system
  • ligand/adduct
  • question being tested
  • why it survived
  • one sentence on caveat

Humans absolutely will lose the thread otherwise.

Use these labels:

  • free-state handling
  • trapped-state endpoint
  • secondary comparator
  • dropped as score-only / dropped as structurally loose

2. Move the finalists into short explicit-solvent MD

This has been the right next move for a while. Your own earlier synthesis already said the best follow-up was short explicit-solvent MD on a very small finalist set, not more broad screening.

Do:

  • 3 × 50 ns as triage for each finalist
  • then extend only survivors to 3 × 150–200 ns

Primary readouts:

  • FAD RMSD / pose-family spread
  • ligand or adduct RMSD
  • minimum sulfur-to-quinone/adduct-relevant distances
  • contact occupancy around the same pocket neighborhood that kept recurring
  • water intrusion / pocket opening
  • whether the “nice” Boltz state survives past minimization and early equilibration

Do not do one heroic 500 ns vanity run. That is how people turn GPU hours into mythology.

3. Fix the chemistry of the trapped states before overclaiming them

Right now the aminochrome adducts are score winners but structurally loose, while the cysteamine + dopamine-o-quinone adduct is the better geometry winner. That means the next chemistry task is not more ranking. It is checking whether the explicit adduct graphs are chemically sane. Use a small-molecule workflow for:

  • geometry optimization
  • charge/protonation sanity
  • atom-mapped reactant/product pairs
  • if possible later, TS/IRC or at least reaction plausibility for the sulfur-addition step

Because if the adduct graph is bogus, Boltz will happily rate nonsense with a straight face.

4. Treat the branches as separate stories

You now have three distinct branches:

  • NQO1 oxidized-quinone handling
  • GSTM2 + GSH compensation / trapping
  • ALDH/DOPAL adjacent toxicity

Do not collapse them back into one big “dopamine damage” blob. That was already a confound in the original thread, and the ALDH/DOPAL branch remains important but adjacent, not the clean next NQO1 target.

5. Define the success criteria before the MD

Otherwise you’ll end up staring at trajectories until one of them reveals Jesus.

For a finalist to “survive,” require something like:

  • no early collapse of FAD-supported geometry
  • ligand/adduct family stays within a sane RMSD band
  • recurring contact network persists in at least 2 of 3 replicas
  • trapped-state endpoint is more coherent than a random loose guest pose
  • free-state winner and trapped-state winner remain meaningfully distinct if they are really different mechanistic states

6. Start the experimental bridge now, not later

Your original post already laid out the real answer: eventually you need a cell system that can tell you whether genotype-specific vulnerability shows up at realistic load, with readouts like 5-Cys-DA, aminochrome, DA-GSH conjugates, DOPAC/HVA, GSH depletion, and the quinone shunting ratio.

So while the MD runs:

  • write the Tier 1 LUHMES / NQO1-KO or knock-in experiment plan

  • define the exact intervention arms:

    • amphetamine only
    • amphetamine + NAC
    • amphetamine + cysteamine
    • maybe MPH comparator

  • keep the outputs tied to your computational winners instead of inventing a whole supplement carnival

The actual decision tree from here

If cysteamine + dopamine-o-quinone free state and cysteamine_dopamine-o-quinone adduct both survive MD, cysteamine becomes your strongest direct NQO1-side lead.

If the free state survives but the adduct does not, then cysteamine is more likely a compatibility/stabilization candidate than a true sulfur-trap endpoint.

If NAC + aminochrome survives better than NAC + dopamine-o-quinone, that supports the GST compensation story being more credible at the aminochrome/downstream stage than at the earliest quinone stage.

If none of the trapped states survive MD, then the current adduct story is mostly Boltz theater and you should pivot to small-molecule reaction modeling before going further.

My blunt recommendation

Your next move is:

one final finalist table → short MD on 4 systems → chemistry sanity check on the adduct graphs → experimental design note tied to cysteamine/NAC.

That is the point where this stops being a decorative computational thread and starts becoming a real mechanism program.

#29

For your top-4 MD pipeline, I would not use a “turnkey MD server.” I’d use a cheap Linux GPU VM and run your own
OpenMM + CHARMM/GAFF-prepped stack.

Best current options:

For your exact workload, I’d pick:

  1. RunPod Pod on RTX 4090 or A5000/L4 if available.
  2. Vast.ai RTX 4090 if you want to minimize cost and are willing to babysit the host selection.
  3. TensorDock RTX 4090 if you want a simpler fallback.
  4. Lambda A10/A6000 if you care more about predictable infra than price.

What I’d actually submit:

  • one Linux CUDA VM, not serverless
  • Ubuntu base
  • OpenMM
  • your prepped force-field files
  • run the 4 finalist systems serially first, not in parallel

Why:

  • these are single-system MD jobs, not elastic inference jobs
  • OpenMM wants a normal GPU box with persistent disk and checkpoints
  • your systems should fit on a 24 GB GPU if you keep them to standard solvated single-GPU runs

My practical recommendation:

  • use RunPod first
  • target RTX 4090 24 GB
  • if price is the only thing that matters, check Vast.ai second
  • do not pay for A100/H100 unless you already know you need the extra VRAM or want much longer production runs

The cheap first pass would be:

  • 4 systems
  • minimization + equilibration + short pilot MD
  • checkpoint everything
  • only then decide whether any one system deserves a longer run
#30

My read:

  • LatchBio:
  • Deep Origin:
    • stronger if you want a managed discovery platform with built-in docking / ABFE / RBFE / protein-ligand tooling
    • less attractive if you want cheap, transparent, self-directed MD
    • current site shows docking and simulation capabilities, including ABFE/RBFE, but code-based simulation access is
      still framed as waitlist/beta rather than a simple public “rent this GPU and run OpenMM” flow:
    • I did not find a clear public per-GPU price table like LatchBio’s

So the blunt recommendation is:

  • if you mean “I want to run my own OpenMM/MD jobs cheaply”: LatchBio over Deep Origin
  • if you mean “I want a more opinionated drug-discovery platform with built-in simulation products”: Deep Origin
  • if you mean “cheapest overall”: still neither. A raw GPU VM on RunPod or Vast.ai is likely cheaper and gives you
    more control
#31

how much foresight do i need for parametrization, how easy is it gto get wrong

• You need more foresight than for docking, but less than for full QM method development.

The real risk is not “the software fails.” It is “the software gives you something that runs, but the chemistry is
wrong enough to mislead you.”

What you need to think about up front:

  • exact protonation state
  • tautomer/redox state
  • whether the ligand is covalent or noncovalent
  • whether a cofactor is standard, modified, or partially reduced
  • whether you need one molecule parameterized or several coupled states
  • whether the adduct/product you are modeling is actually the right constitutional isomer

How easy it is to get wrong:

  • easy, if you rush
  • especially for:
    • FAD/FM N/FADH2
    • quinones
    • semiquinone/radical-like species
    • sulfur adducts
    • anything with multiple plausible microstates

Typical failure modes:

  • wrong charge state
  • wrong tautomer
  • wrong atom typing on unusual heterocycles
  • high-penalty CGenFF parameters accepted without review
  • broken geometry because the starting structure was bad
  • treating a reactive intermediate like a stable ligand
  • forgetting that one “named compound” may actually mean multiple chemistry states at pH 7.4

What is low-risk:

  • standard amino acids
  • standard nucleotides
  • ordinary small neutral drug-like ligands with clear protonation

What is medium/high-risk in your project:

  • dopamine o-quinone
  • aminochrome
  • leukoaminochrome
  • sulfur adduct endpoints
  • reduced flavins

So the correct mindset is:

  • parameterization is not hard because it is computationally heavy
  • it is hard because it is chemically easy to lie to yourself

Practical rule:

  1. lock the exact structure first
  2. lock the microstate second
  3. inspect parameter penalties/typing
  4. only then run MD

If you want, I can write you a short pre-MD parameterization checklist specifically for your top finalists so you know
what must be verified before spending GPU money.

#32

+# Pre-MD Parameterization Checklist For Top Quinone Finalists
2 +
3 +This is the gate before spending remote GPU time.
4 +
5 +The goal is not merely to make the system run. The goal is to avoid running a chemically wrong system that prod
uces believable but misleading MD.
6 +
7 +## Current finalists
8 +
9 +Free-state handling:
10 ± NQO1 P187S + FAD + Cysteamine + Dopamine o-quinone
11 ± GSTM2 + GSH + NAC + Aminochrome
12 ± GSTM2 + GSH + NAC + Dopamine o-quinone as a secondary comparator
13 +
14 +Trapped-state endpoint:
15 ± NQO1 P187S + FAD + Cysteamine_DopamineOQuinone_Adduct
16 +
17 +## Rule 1: Freeze the exact chemistry first
18 +
19 +Before any parameterization:
20 ± decide the exact constitutional structure
21 ± decide the exact protonation state
22 ± decide the exact tautomer
23 ± decide the exact redox state
24 ± decide whether the state is free coexistence or explicit adduct endpoint
25 +
26 +If any of those are still ambiguous, do not parameterize yet.
27 +
28 +## Rule 2: Treat these systems as high-risk chemistry
29 +
30 +These are not routine neutral ligands.
31 +
32 +High-risk pieces in this project:
33 ± FAD
34 ± dopamine o-quinone
35 ± aminochrome
36 ± sulfur adduct products such as Cysteamine_DopamineOQuinone_Adduct
37 +
38 +Why they are high-risk:
39 ± multiple plausible microstates
40 ± reactive quinone chemistry
41 ± unusual heteroaromatic atom typing
42 ± cofactor-like ligands with nontrivial charges
43 +
44 +## System-by-system checklist
45 +
46 +## 1. NQO1 P187S + FAD + Cysteamine + Dopamine o-quinone
47 +
48 +Must verify:
49 ± FAD is the oxidized dinucleotide state, not reduced or partially reduced
50 ± dopamine o-quinone is in the Rowan-selected primary microstate
51 ± cysteamine protonation is fixed explicitly
52 ± all three nonstandard components use mutually compatible charge conventions
53 ± the starting pose does not contain impossible overlaps between FAD, quinone, and cysteamine
54 +
55 +Parameterization danger points:
56 ± wrong quinone amine protonation
57 ± inconsistent total charge across cofactors/ligands
58 ± accepting flavin parameters without reviewing penalties
59 +
60 +## 2. GSTM2 + GSH + NAC + Aminochrome
61 +
62 +Must verify:
63 ± GSH is parameterized in the intended thiol state, not oxidized/disulfide-like
64 ± NAC protonation is fixed explicitly
65 ± aminochrome is the neutral parent state used in the Rowan-guided Boltz workflow
66 ± free-state modeling is kept separate from explicit conjugate/product modeling
67 +
68 +Parameterization danger points:
69 ± mixing up GSH trap chemistry with an already-conjugated endpoint
70 ± using the wrong aminochrome microstate
71 ± letting the tool infer inconsistent sulfur charges across GSH and NAC
72 +
73 +## 3. GSTM2 + GSH + NAC + Dopamine o-quinone
74 +
75 +Must verify:
76 ± dopamine o-quinone uses the Rowan-selected protonated ammonium state
77 ± GSH and NAC are both modeled as free coexistence partners, not pre-reacted products
78 ± the initial structure does not place all sulfur/electrophile centers in impossible contact
79 +
80 +Parameterization danger points:
81 ± neutralizing dopamine o-quinone incorrectly
82 ± building a coexistence system that is actually closer to a covalent product than a free state
83 +
84 +## 4. NQO1 P187S + FAD + Cysteamine_DopamineOQuinone_Adduct
85 +
86 +Must verify:
87 ± the adduct constitutional graph is the one you actually intend to test
88 ± total charge is chemically sensible
89 ± the adduct is parameterized as a single molecule, not as separate fragments
90 ± FAD remains the oxidized state unless you intentionally change it
91 +
92 +Parameterization danger points:
93 ± wrong sulfur-addition site on the quinone
94 ± adduct ring/aromaticity assignment errors
95 ± broken atom typing around the newly formed C-S bond
96 ± hidden mismatch between the adduct you scored in Boltz and the adduct you parameterize for MD
97 +
98 +## Parameter review checklist
99 +
100 +For every nonstandard molecule:
101 ± inspect the formal charge
102 ± inspect protonation manually
103 ± inspect aromaticity/ring assignment
104 ± inspect atom types
105 ± inspect any reported parameter penalties or missing terms
106 ± inspect the minimized standalone molecule geometry before building the full solvated system
107 +
108 +If a tool reports high penalties for key torsions, ring atoms, or sulfur-connected centers, do not ignore that.
109 +
110 +## Minimum quality bar before MD
111 +
112 +Do not launch production MD until all of this is true:
113 ± all nonstandard molecules have a locked structure and microstate
114 ± no missing parameters remain
115 ± no obviously extreme penalties remain unreviewed
116 ± a vacuum or implicit-solvent minimization of the isolated ligand/adduct looks sane
117 ± the assembled complex minimizes without exploding
118 ± a short restrained pilot run keeps the cofactor/ligand geometry physically plausible
119 +
120 +## Cheap pilot before real GPU spend
121 +
122 +For each finalist:
123 +1. parameterize the nonstandard species
124 +2. minimize the isolated species
125 +3. build the complex
126 +4. run restrained minimization
127 +5. run very short equilibration
128 +6. check whether the ligand/cofactor stays in a sane geometry
129 +
130 +Only after that should you run a longer OpenMM production job.
131 +
132 +## Practical decision rule
133 +
136 +If the chemistry is clear but the parameter penalties are ugly, treat the system as provisional.
137 +
138 +If both the chemistry and parameterization look clean, that is the system worth sending to the remote GPU first
.
139 +
140 +## Recommended first remote order
141 +
145 +4. GSTM2 + GSH + NAC + Dopamine o-quinone
146 +
147 +This order favors the systems that are both mechanistically important and less likely to fail immediately for a
voidable chemistry reasons.

#33

• Moderate, not high.

My blunt estimate:

  • chance it gives you a genuinely new mechanistic insight: 40-60%
  • chance it gives you a clean, publishable or decision-grade answer by itself: 10-20%
  • chance it mostly confirms what you already suspect, with better confidence and a few
    eliminations: 60-80%

Why I think it is still worth doing:

  • you already squeezed a lot out of Boltz alone
  • the shortlist is now narrow and chemically sharper
  • MD is the first step that can tell you whether the finalists are actually stable over
    time, or whether the Boltz poses are decorative
  • that is exactly the unresolved question now

What new insight it could realistically add:

  • whether Cysteamine + dopamine o-quinone stays coherent in the defective NQO1 pocket
  • whether NAC + aminochrome is really a plausible GST-side coexistence state
  • whether the adduct endpoint is more stable than the free-state complex
  • whether the whole story is actually a cofactor-collapse problem rather than a ligand-
    choice problem

What it probably will not do:

  • prove therapeutic efficacy
  • prove in vivo detox
  • settle radical chemistry cleanly
  • give you a magic unexpected rescue molecule out of nowhere

So the real value is not “discover a miracle.” It is:

  • kill false positives
  • rank the finalists more honestly
  • decide whether this is worth deeper chemistry or experimental follow-up