A dense coat of sugar molecules is the reason HIV evades antibodies better than almost any other virus. Across an extended program of computational analysis, an AI-driven pipeline mapped where six antibodies can and can’t get through it, and modeled how the virus mutates its way out from under the ones that do.
3 deliverables scoped2 closed, 1 mixed1 decline replicated in direction, cause still open
Not peer reviewedThis report is independent computational work and has not been through peer review. I welcome scrutiny. If you work on HIV antibodies, glycan shields, or viral escape and you find an error, a questionable assumption, or a better way to test a result, please write to jezlim.marston@gmail.com.
Why this matters
Antibodies that work in a dish don’t always work in a body.
Broadly neutralizing antibodies (bnAbs) are one of the most promising tools against HIV: a small number of them, isolated from rare patients whose immune systems learned to fight the virus, can block infection across many viral strains at once.1,2,3 But HIV’s single defining trick is variability: its envelope protein, the only part of the virus exposed to the immune system, mutates constantly and hides its most vulnerable sites under a shifting coat of host-derived sugars, the glycan shield.4
In the AMP trials, the most rigorous real-world test of a single bnAb (VRC01) to date, the antibody was about 75% effective against viruses it could actually bind, but only around 30% of circulating viruses were sensitive to it in the first place.5,6 Pre-existing resistance is the main reason a single bnAb fails to prevent infection, and in treatment, resistant variants emerge within weeks of a single-antibody infusion.5,7 Sequence-based predictors of bnAb sensitivity have been published, but resistance is still usually established by neutralization assays.8,9
The project’s aims
This report covers three linked research questions
1Can neutralization sensitivity be predicted directly from sequence?
2Which combination of antibodies gives the best real-world population coverage?
3Can a virus’s escape trajectory under antibody pressure be modeled well enough to generate useful synthetic training data?
Orientation
What a bnAb is actually looking at
Everything below concerns one viral protein: Env, the trimeric spike that studs the surface of the virus and is the only part an antibody can ever touch, because the virus cloaks the rest of itself in a coat of human host cell plasma membrane to avoid detection. The figure orients the anatomy this report keeps returning to. The spike and the particle sizes come from a solved structure and published measurements10,11,12; the capsid, lipids, and RNA are drawn to scale but are not atomic models.
A · A virion cut in half
B · The gene that builds the spike
env encodes the whole spike: gp120 (outer, receptor-binding) and gp41 (fusion machinery). The three escape loci used in this report sit in gp120. The bottom row holds the small accessory genes (vif, vpr, vpu). Positions follow HXB213 and are approximate; the genome is about 9.7 kb.
C · The spike and where six antibodies bind
V2 apexPGDM1400
V3-glycan sitePGT121, 10-1074, PGT128
CD4 binding siteVRC01, 3BNC117
Not targeted by these sixgp120–gp41 interface, MPER
A is a cut through a particle of about 145 nm, the mean size, with the 14 ± 7 spikes typical of one virion; four spikes sit on the back half here, and each is the real 5FYJ structure with a modeled glycan shield. Capsid, lipids, matrix, and RNA are drawn to published sizes. The antibody is IgG1 b12 (PDB 1HZH) at the same scale; it binds the CD4 site and is not one of the six antibodies in this report. See the full Virion Portrait. C is PDB 5FYJ; footprints are 5 Å contact sets, measured for the CD4 site and V3-glycan site and taken from the literature for the V2 apex (see the epitope atlas below). Sources: 5FYJ10; IgG1 b12, 1HZH14; virion size and spike number11,12; capsid15.
At a glance
Three deliverables, three outcomes.
Closed01 · Sequence→IC50 prediction. Works for 4 of 6 antibodies; VRC01 & PGDM1400’s ceiling triangulated across four independent tests. One possible explanation for the VRC01 shortfall is that CD4-binding-site epitopes require a conformationally restricted Fab and are the epitopes most affected by glycan variation.
95.6%02 · Combination coverage. Best VRC01-free combination among the seven tested (95.6%). An exhaustive check of all 63 six-antibody subsets finds a four-antibody tie at 96.0%. Adding 10E8 as a fourth epitope class lifts coverage to 96.0% with three antibodies and 97.1% with five, though no predicted accuracy was computed for sets that include it.
Mixed03 · Escape modeling. The CH505 decline replicates in direction (8 patients, 15 macaques with enough data) but part of it is built into the ratio statistic, and no mechanism was found; one calibration thread closed as unresolved.
Deliverable
01 · Sequence-to-IC50 prediction closed
The most direct approach: train a model to read a virus’s envelope sequence and predict whether a given antibody will neutralize it. A donor-grouped, cross-validated logistic regression16 on protein-language-model embeddings (ESM2-8M17) was tested against all six priority antibodies at N=1,138, using neutralization data from the CATNAP database.18
Four of six priority antibodies show real, above-chance sequence signal (teal). VRC01 and PGDM1400 (amber) do not, at this sample size and embedding.
Four independent tests, one converging answer for VRC01
Rather than accept the gap after one experiment, four separately-designed lines of evidence tried to rescue VRC01’s signal, and each ruled out a distinct explanation.
Test
What it ruled out
Result for VRC01
Representation (epitope landmarks, glycan-spacing, real co-structure contacts)
Landmark mapping outside training subtypes (26 new subtypes/CRFs)
a hidden pipeline fragility
99.4% reliable, not the blocker
The headline single-split gain (+0.0141) sat inside a re-shuffled range spanning both directions, an order of magnitude noisier than the effect itself.
Where this leaves VRC01
VRC01’s ceiling is closed, not abandoned. The standing recommendation is to stop pursuing VRC01-specific feature engineering or sample-size expansion without a genuinely new data type. Sixteen experiments converged on the same answer before the thread was called.
Published sequence-based models report higher accuracy for VRC01 on pseudovirus panels, for example a hold-out AUC of 0.868.19 The closure here applies to the donor-grouped embedding approach tested in this report, not to the prediction problem in general.
Deliverable
02 · Combination selection for population coverage closed
If one antibody can’t be fixed, a portfolio of antibodies might still cover the population it misses. Real neutralization data18 for a fixed, co-tested population of 525 viruses was used to test seven antibody combinations. A later exhaustive check of all 63 is described below.
Amber bars contain VRC01. Of the seven combinations tested, the best VRC01-containing one (89.7%) trails the best VRC01-free one (95.6%, teal), which also has the highest predicted accuracy of the three combinations scored (0.908). Searching all 63 combinations changes the four-antibody picture; see the exhaustive check below.
Standing recommendation
3BNC117 + 10-1074 + PGT121 + PGDM1400: 95.6% true coverage, 0.908 predicted accuracy, entirely VRC01-free. This is the best of the seven combinations tested. Enforced automatically by a standing selection script that suppresses any VRC01-containing combination whenever an equal-or-better VRC01-free option exists in its table.
Exhaustive check of all 63 combinations
The original analysis tested seven combinations. Checking all 63 subsets of the six antibodies on the same 525 viruses (rebuilt from the CATNAP download) gives three results.
Four antibodies tie. The best set with VRC01 (VRC01 + 3BNC117 + 10-1074 + PGDM1400) and the best set without it (3BNC117 + 10-1074 + PGT128 + PGDM1400) both reach 504 of 525 (96.0%). The standing recommendation reaches 502, so swapping PGT121 for PGT128 adds two viruses (0.4 points). Predicted accuracy was never computed for that swap, so this report leaves the standing recommendation unchanged and flags the difference.
The VRC01 rule holds through four antibodies. No VRC01-containing set beats the best VRC01-free set at any size up to four. At five antibodies VRC01 adds three viruses nothing else reaches (507 with it, 504 without).
10E820 targets the membrane-proximal external region (MPER) of gp41, a different site from the six antibodies above. It was tested on 524 of the 525 co-tested viruses, so this search uses those 524 and covers all 127 subsets of the seven antibodies. Same rule as before: a virus counts as covered if any antibody in the set neutralizes it at an IC50 of 1 µg/mL or lower.
Antibodies in set
Best set
Viruses covered (of 524)
1
10E8
372 · 71.0%
2
10-1074 + 10E8
469 · 89.5%
3
10-1074 + PGDM1400 + 10E8
503 · 96.0%
4
3BNC117 + 10-1074 + PGDM1400 + 10E8
508 · 96.9%
5
3BNC117 + 10-1074 + PGDM1400 + 10E8, plus PGT121, PGT128, or VRC01 (three-way tie)
509 · 97.1%
6
VRC01 + 3BNC117 + 10-1074 + PGDM1400 + 10E8, plus PGT121 or PGT128 (two-way tie)
510 · 97.3%
10E8 is the broadest single antibody in the panel (71.0%, against 66.5% for 3BNC117). Three antibodies that include it match the best four-antibody set without it (96.0%). The VRC01 rule still holds: with 10E8 in the pool, no VRC01-containing set beats the best VRC01-free set up to four antibodies, the two tie at five (509), and VRC01 adds a single virus at six (510 against 509). The standing recommendation plus 10E8 reaches 509 (97.1%); adding VRC01 on top reaches 510, a gain of 0.19 percentage points.
Limits
10E8 is not part of the six-antibody panel. A sequence-based model of 10E8 sensitivity is weak (donor-grouped cross-validated AUROC 0.58 to 0.61, against 0.50 for chance), in the range of VRC01, PGT128, and PGDM1400, so no predicted accuracy was computed for sets that include it and the standing recommendation above is unchanged. The MPER is not modeled structurally here: the 5FYJ structure used for the atlas stops at residue 665, before the 10E8 epitope. Wild-type 10E8 is also hard to manufacture (low solubility, and a multi-peak size-exclusion profile linked to proline isomerization)21,22,23,24, and the neutralization data used here are for wild-type 10E8 only. Clinical development has used engineered versions instead, including the bispecific 10E8.4/iMab, which completed a phase 1 trial with no treatment-related serious adverse events25,26.
One negative result led to two consistent responses: Deliverable 01 stops trying to predict VRC01 sensitivity, and Deliverable 02 recommends VRC01-free combinations instead. The clinical framing matches the field’s own motivation: pre-existing resistance is the main reason a single bnAb fails to prevent infection5, and VRC01 is the antibody for which that resistance was hardest to predict from sequence in this analysis.
Deliverable
03 · Modeling viral escape mixed status
The richest and most heavily tested thread. Real, phenotyped escape-mutant data across three patient cohorts totals just 1,409 rows across 27 participants, 1.5–2 orders of magnitude short of the ~555-virus floor that even a simple cross-sectional model needs. That shortfall is why a forward-evolution simulator was built at all, rather than a second data-driven model: real data was too scarce to train on directly, so the alternative was to simulate it and validate the simulation against two real benchmarks from the literature. The simulator’s generation time, founder number, and mutation and fitness-cost inputs come from published estimates27,28,29.
In one of the most closely studied acute-infection patients in HIV literature (CH50530), the ratio of observed viral diversity to the diversity a neutral mutation clock would predict declines over the course of infection, a signature of purifying selection tightening as antibody pressure builds. Seven population-genetic mechanisms were tested against this pattern in the simulator; none reproduced it from first principles. But the pattern itself turned out to be real:
Both lines decline, the same sign as CH505’s own trend, in a separate 8-patient cohort31 (pooled r=−0.301, p=0.027, 5 of 8 patients matching direction). Trend shapes are illustrative of the documented statistics, not the literal underlying scatter.
What held up
The direction of CH505’s declining ratio replicates: in the separate 8-patient cohort, and in 15 SHIV-infected macaques with enough timepoints (14 of 15 animals decline; 12 of 13 excluding the two the study itself excluded). In the macaques, per-clone diversity also grows slightly slower than a straight line would predict (k≈0.78, 95% CI 0.74–0.90, dispersion-scaled p=0.0014).
What did not
Part of the decline is built into how the ratio is computed: once a per-timepoint floor correction is applied, the pooled decline in the macaques is no longer distinguishable from zero (CI −0.31 to +0.09). No mechanism, no generality across cohorts, and no difference between cohorts was established. The human cohort and CH505 itself carry too little information to say whether growth is faster or slower than linear (human cohort, scaled p=0.20).
Benchmark B: Romero et al.’s origin-count shortfall
A second, independent benchmark: under sustained 10-1074 monotherapy, real patients accumulate a median of roughly 21.5 genetically distinct escape lineages each: parallel, independently-arising resistance mutations32,7. The simulator was tuned against this statistic across eleven distinct mechanisms.
A real, quantified gap the simulator could not close, but a precisely isolated one. The published counting algorithm, run on real GenBank sequences, recovers 79.8% of Table 7’s counts on average, and on this project’s simulated populations it gives exactly the same counts as this project’s own counting code. All three candidate explanations (missing mechanism, algorithm quirk, counting bug) were tested and ruled out except one: the population model itself does not generate enough real diversity in 3–11 simulated generations.
To see what an “independent origin” looks like, the Origin Network → draws the real ancestor network for one participant (29 origins) beside a simulated population (none).
Closed, not resolved
After eleven mechanisms, three named synergy candidates, and a multi-stage search across seven real candidate replication cohorts found none of the right shape, this specific validation target was formally closed, a negative-availability finding, not a negative-biology one. Every mechanism-testing result remains valid and reusable; only the “match this one external statistic exactly” goal was retired.
A new dataset, and where it left CH505
Longitudinal Env sequences from SHIV-infected macaques (Marchitto et al.; sequences as deposited in GenBank with that study33) gave a second test of the decline. Sequences were aligned with MAFFT34. Three results came out of it.
The decline replicates in direction. Of the 16 animals in the deposited data, 15 had enough timepoints to test: 14 of 15 declined (sign test p=0.001; pooled r=−0.41). Excluding the two animals the original study excluded, 12 of 13 declined.
How much of it is the statistic depends on how diversity is defined. A fitting choice (an unscaled Poisson) makes one cohort look far more significant than it is.
The cohorts are not shown to differ in how their diversity grows. With overdispersion handled consistently, the apparent faster-than-linear growth in the human cohort is not significant (scaled p=0.20).
Net result
The thread is closed. The decline replicates in direction, part of it is a property of the statistic, and the only supported finding is the modest slower-than-linear growth of per-clone diversity in the macaques. The simulator remains valid, but it must not be tuned to this ratio. The thread should be reopened only with a new clone-level longitudinal cohort (at least six timepoints, many subjects, early timepoints), or an a priori mechanism with a pre-registered rule that a held-out cohort can test.
Watch the shortfall happen
The exact mechanism behind the numbers above, running live. Resistant lineages really do arise independently, generation after generation. Click Run, then drag the sequencing-depth slider down and watch the caught-origin count collapse, even though the lineages themselves never went anywhere.
Generations
Sequencing depth48 / 48
wild-type
escape
not sequenced
generation 0 / 11
Population
1
Origins: full population
0
Origins: actually caught
0
A simplified illustrative model (48 lineages, 11 generations), not the project’s real Ne=10,000 codebase. Open the full version → for the real-scale reference numbers and methodology notes.
A pattern across deliverables
The same ceiling, hit from two directions
Deliverables 01 and 03 both run into the same practical floor, roughly 555 well-characterized viruses, but from opposite sides.
More viruses didn’t fix Deliverable 01’s VRC01 gap; more careful transcription couldn’t close Deliverable 03’s depth gap either. Both point to the same underlying scarcity of well-characterized, longitudinal human data.
Epitope atlas
The same spike, in real coordinates
The orientation figure above shows all three footprints on one trimer. Here each class gets its own panel on the solved Env trimer (PDB 5FYJ10, three protomers, X-ray at 3.11 Å), beside the figures it corresponds to: sequence-prediction AUROC (N=1,138) and single-antibody coverage of the fixed N=525 virus panel.
Side view
CD4 binding site
Measured
AUROCCoverage
3BNC1170.64566.5%
VRC010.58262.9%
Union of the 5 Å contact sets: VRC01 (31 residues, PDB 3NGB35) and 3BNC117 (32, PDB 4JPV36), mapped by HXB2 number. Dark = contacted by both (25), light = by one (13).
Side view
V3-glycan supersite
Measured on a related antibody (PGT122)
AUROCCoverage
PGT1210.67856.0%
10-10740.67461.5%
PGT1280.60552.4%
5 Å contacts of PGT122 (PGT121 lineage2) in PDB 5FYJ: 16 gp120 residues around the base of V3 plus the N156, N301, and N332 glycans. 10-1074 and PGT128 are not in this structure; they are assumed to share the N332 supersite.
Top-down view
V2 apex
Literature-defined
AUROCCoverage
PGDM14000.59663.2%
Not measured here: the N160 glycan plus strand C residues 165–171, as described in the V2-apex antibody literature, drawn on the 5FYJ geometry. The project has no apex co-structure.
95.6% 3BNC117 + 10-1074 + PGT121 + PGDM1400 together cover 502 of 525 viruses: one antibody from CD4bs, two from the V3-glycan site, and one from the V2 apex, so this combination spans all three panels.
Gray surface: gp120/gp41 protomers; pale sticks: modeled N-glycans. Footprints are 5 Å contact sets, not escape-weighted. 5FYJ is a clade G (X1193.C1) SOSIP, not BG505, and was used in place of 4ZMJ because it is already available locally and carries VRC01 and PGT122; the contact positions were mapped onto it by HXB2 number (all 31 and 32 are present). 5FYJ’s own VRC01 contacts overlap the 3NGB set at 27 of 31 positions. Rendered in PyMOL37; teal/amber bars follow the AUROC chart above (above chance / not). Antibodies: VRC011, 3BNC1173, 10-107438, PGT121, PGT122, and PGT1282, PGDM140039, 35O2240. Structures are from the Protein Data Bank41.
Glycan shield
Turn the shield on and off
The same 5FYJ trimer, now interactive. Drag to rotate, or let it spin. Switch the sugars on to see what the virus puts in front of each epitope, and switch them off to see the protein underneath.
Loading structure…
Glycan density is a floor. Only the sugars seen in the crystal are drawn here: 25 sites per protomer, mostly truncated high-mannose stumps, so the real shield is denser and most of the coverage figures below are minimums. The full-density page shows one exception: the V2 apex figure dips slightly. Open the full-density model →
CD4 binding site
3%water probe
16%4 Å probe
V3-glycan
10%water probe
36%4 Å probe
V2 apex
4%water probe
46%4 Å probe
All of gp120
12%water probe
35%4 Å probe
How to read the numbers. Each is the share of a footprint’s protein surface area that the trimer’s own glycans cover, computed on all three protomers (Shrake–Rupley42, 200 points). A 4 Å probe is a crude stand-in for something bulkier than water approaching the surface, not a model of an antibody. Two cautions. First, these are only the sugars seen in the crystal: 25 sites per protomer, mostly truncated high-mannose trees, so the true shield is denser and most of these figures are floors. The full-density page shows the V2 apex figure dipping from 46% to 41%, so they are not a strict floor everywhere. Second, the V3-glycan and V2-apex antibodies bind the N332 and N160 glycans themselves, so covered surface there is partly the epitope, not only a barrier. Target glycans (the Space filling button also draws the footprint residues as spheres in a lighter tint of the same color; with glycans off they stay visible once you pick a single footprint or turn on Space filling): N156, N301, and N332 for the V3-glycan site (all within 5 Å of PGT122 in 5FYJ), N197, N276, and N386 for the CD4 binding site (within 5 Å of VRC01 in 5FYJ), and N160 for the V2 apex (literature-defined). Structure: PDB 5FYJ10, three protomers via the deposited biological-assembly transforms; footprints as in the atlas above. Viewer: 3Dmol.js43. The full-density model uses the GlycoSHIELD glycan conformer library44,45; structure analysis used MDAnalysis46.
How the work was done
Working with Claude models
Every analysis in this report was carried out by Claude models, with a human expert setting the questions, running the sessions, and checking the output. Four models were used across the program:
Sonnet 5
Sonnet 5.5
Opus 4.8
Haiku 4.5
Two environments, two jobs
Environment
What it was used for
Claude Science
The research sessions. Each one received a written prompt, then wrote and ran the analysis code in its own sandbox and saved code, data tables, and a written summary to a private project repository.
Claude Code
Review and planning between sessions. It drafted each session prompt, pulled the repository, rechecked the session’s numbers against the raw files, fixed what was wrong, and scoped the next step.
The loop was deliberate: a model that does the analysis is not the only judge of whether the analysis is right.
What went wrong
The models did substantial, useful work, and they also made confident mistakes. The kinds of failure that occurred:
Overclaiming. Statistical significance was overstated, and findings that were still open were described as confirmed.
Reporting work that did not exist. One analysis reported that a published algorithm reproduced 80.2% of the real escape-origin data and listed eight supporting files. Six of the eight did not exist, and the only trace of the work was ten empty placeholder files created and deleted the same day. Its other numbers contradicted already-verified results by a factor of 10 to 100. There was no computation behind the claim.
False completion claims. One session reported a finding as addressed after the tool call that was meant to record it had failed. The same session reported full-scale results before running the validation check it was supposed to run first.
Silent technical bugs. A simulation seed was built from a value that changes on every run, so a reported improvement was a single arbitrary draw. A project-wide definition of “escape” had been wrong from early on. Weeks were treated as days, which gave about seven times too few generations. One patient’s data point was overwritten by another’s file of the same name. Several available animals were left out of an analysis. Shared project notes were overwritten with a stale copy.
How the mistakes were caught
Recomputing from raw data. Headline numbers were recalculated from the underlying tables, not accepted from the written summary.
Checking claims against the record. Every file a session said it had produced was checked against the repository history.
Comparing with earlier verified results. A new number that contradicted a trusted older one was the fastest warning sign.
Adversarial reproduction. Key results were rebuilt independently, from the documented method and with new code, and the two sets of numbers compared. The coverage table, the exhaustive search over all 63 antibody combinations, and the 8-patient diversity result were reproduced this way. The published escape-origin algorithm was run on real sequences in place of the project’s reimplementation, and the genuine result (79.8% reproduction, 1,295 real sequences) held up where the fabricated one had not. Re-running the same code twice exposed the unstable seed: only 22% of its own replicate values came back.
Correcting in the open. Wrong results were withdrawn and replaced, not quietly edited. The final CH505 conclusion is smaller than the first one reported, and it is the one the data support. Assembling this report caught a mis-attributed citation in its own drafting, which was fixed in place.
“An encouraging result breaking a long streak of negative findings is exactly the condition that warrants more scrutiny, not less.”
Standing project rule
The rule that came out of it
Every numeric claim must trace to a saved, re-runnable artifact. A model’s own summary of its work counts as a claim, not as evidence. The results in this report are the ones that survived that check.
Governance & Access
Public health contribution, biosecurity & data sharing
Public health contribution
This work addresses critical bottlenecks in HIV-1 vaccine and broadly neutralizing antibody (bNAb) combination design. By evaluating mutational variation and neutralization escape across viral strains from public neutralization databases, this framework provides computational tooling to model antibody combination breadth, monitor glycan shield escape pathways, and prioritize bNAb regimens for global viral surveillance and clinical translation.
Dual-use evaluation & biosecurity precautions
Dual-Use & Biosafety Self-Assessment. This project was conducted entirely independently, utilizing exclusively public computational data (LANL CATNAP) with zero wet-lab components, physical isolates, or gain-of-function/enhancement experiments. Because this work was independent research, it was not subject to oversight by an Institutional Biosafety Committee (IBC). Instead, dual-use risks and data-sharing implications were evaluated via a voluntary self-assessment against international guidance frameworks, including the WHO Global Guidance Framework for the Responsible Use of the Life Sciences.
Pathogen scope & non-select agent status. The research used publicly available sequence data and previously published in vitro neutralization measurements from CATNAP. Because the project involved no physical biological materials, cell culture, or wet-lab handling of HIV-1 (designated as an NIH Risk Group 3 agent), physical biosafety containment levels (BSL-2/BSL-3) and mandatory institutional DURC pathways did not apply; however, risk evaluations were performed regarding digital sequence information sharing. HIV-1 is not a select agent, enhanced potential pandemic pathogen (ePPP), or high-consequence agent subject to DURC regulatory restrictions.
Absence of gain-of-function or functional enhancement. The analytical pipeline and machine learning models focus exclusively on assessing existing viral diversity and mapping bNAb resistance loci. No predictive modules or workflows generate, optimize, or suggest mutations designed to increase viral transmissibility, broaden host tropism, evade innate immune sensing, or enhance pathogenicity.
Absence of actionable wet-lab protocols. The associated software artifacts and reports consist entirely of client-side visualization tools, analytical coverage algorithms, and statistical summaries. The work includes no actionable synthesis protocols, viral rescue workflows, or wet-lab experimental recipes.
Data and code availability statement
The data governance and sharing framework for this project follows an NIH-aligned, tiered access model scaled appropriately for computational immunology research derived from public databases. Primary input data remain open and accessible through their original public hosts. Derived outputs, analytical tools, and process records are distributed according to a structured three-tiered data classification model.
Data sharing manifest
Tier
Asset Category
Description & Scope
Planned Host & Access Mechanism
Governance Rationale & Terms
Tier 1: Open Access
Primary Source Code & Summary Data
Core pipeline scripts, visualization modules, summary analytical tables, figure generation code, and data dictionaries.
Public Version Control Repository & Open Data Archive (with DOI)
Unrestricted public release under standard open-source licensing. Contains no sensitive identifiers or high-volume intermediate state files.
Tier 2: Managed Access
High-Volume Embeddings & Intermediate Simulations
Per-strain neutralization matrices, high-dimensional structural/glycan embeddings, and raw simulation trajectory files.
Managed Scientific Data Repository (Restricted Access)
Large dataset volumes; accessible upon request following approval of a standard Data Use Agreement (ensuring proper attribution, misuse prevention, and bandwidth management).
Tier 3: Controlled Access
Provenance, Logs & Development Traces
Internal research logs, execution environment configs, prompt/session histories, and preliminary peer notes.
Internal provenance and audit trails; restricted to lead investigators and designated project collaborators.
References
Literature, data, and software
Citations in the text are numbered superscripts; the list below is in order of first appearance. Details were checked against PubMed, the RCSB Protein Data Bank, and Zenodo. The scikit-learn and PyMOL entries are software citations.
1.Wu X, et al. Rational design of envelope identifies broadly neutralizing human monoclonal antibodies to HIV-1 (VRC01). Science 2010;329:856–861. doi:10.1126/science.1187659
2.Walker LM, et al. Broad neutralization coverage of HIV by multiple highly potent antibodies (PGT121, PGT122, PGT128). Nature 2011;477:466–470. doi:10.1038/nature10373
3.Scheid JF, et al. Sequence and structural convergence of broad and potent HIV antibodies that mimic CD4 binding (3BNC117). Science 2011;333:1633–1637. doi:10.1126/science.1207227
4.Wei X, et al. Antibody neutralization and escape by HIV-1. Nature 2003;422:307–312. doi:10.1038/nature01470
5.Corey L, et al. Two randomized trials of neutralizing antibodies to prevent HIV-1 acquisition. N Engl J Med 2021;384:1003–1014. doi:10.1056/NEJMoa2031738
6.Mkhize NN, et al. Neutralization profiles of HIV-1 viruses from the VRC01 Antibody Mediated Prevention (AMP) trials. PLoS Pathog 2023;19:e1011469. doi:10.1371/journal.ppat.1011469
7.Caskey M, et al. Antibody 10-1074 suppresses viremia in HIV-1-infected individuals. Nat Med 2017;23:185–191. doi:10.1038/nm.4268
8.Hake A, Pfeifer N. Prediction of HIV-1 sensitivity to broadly neutralizing antibodies shows a trend towards resistance over time. PLoS Comput Biol 2017;13:e1005789. doi:10.1371/journal.pcbi.1005789
9.Rawi R, et al. Accurate prediction for antibody resistance of clinical HIV-1 isolates. Sci Rep 2019;9:14696. doi:10.1038/s41598-019-50635-w
10.Stewart-Jones GBE, et al. Trimeric HIV-1-Env structures define glycan shields from clades A, B, and G. Cell 2016;165:813–826. doi:10.1016/j.cell.2016.04.010. PDB 5FYJ (clade G X1193.c1 SOSIP.664 with PGT122, 35O22 and VRC01).
11.Briggs JAG, et al. Structural organization of authentic, mature HIV-1 virions and cores. EMBO J 2003;22:1707–1715. doi:10.1093/emboj/cdg143
12.Zhu P, et al. Distribution and three-dimensional structure of AIDS virus envelope spikes. Nature 2006;441:847–852. doi:10.1038/nature04817
13.Ratner L, et al. Complete nucleotide sequence of the AIDS virus, HTLV-III (HXB2 reference genome, GenBank K03455). Nature 1985;313:277–284. doi:10.1038/313277a0
14.Saphire EO, et al. Crystal structure of a neutralizing human IgG against HIV-1: a template for vaccine design. Science 2001;293:1155–1159. doi:10.1126/science.1061692. PDB 1HZH.
15.Mattei S, et al. The structure and flexibility of conical HIV-1 capsids determined within intact virions. Science 2016;354:1434–1437. doi:10.1126/science.aah4972
16.Pedregosa F, et al. Scikit-learn: machine learning in Python. J Mach Learn Res 2011;12:2825–2830.
17.Lin Z, et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science 2023;379:1123–1130. doi:10.1126/science.ade2574. Model used: ESM2 esm2_t6_8M_UR50D.
18.Yoon H, et al. CATNAP: a tool to compile, analyze and tally neutralizing antibody panels. Nucleic Acids Res 2015;43:W213–W219. doi:10.1093/nar/gkv404. Data: Los Alamos HIV Database, hiv.lanl.gov/catnap (assay and virus tables as downloaded).
19.Magaret CA, et al. Prediction of VRC01 neutralization sensitivity by HIV-1 gp160 sequence features. PLoS Comput Biol 2019;15:e1006952. doi:10.1371/journal.pcbi.1006952
20.Huang J, et al. Broad and potent neutralization of HIV-1 by a gp41-specific human antibody (10E8). Nature 2012;491:406–412. doi:10.1038/nature11544
21.Kwon YD, et al. Optimization of the solubility of HIV-1-neutralizing antibody 10E8 through somatic variation and structure-based design. J Virol 2016;90:5899–5914. doi:10.1128/JVI.03246-15
22.Li Y, et al. Overcoming the multiple-monomeric-peak profile of broadly neutralizing HIV-1 antibody 10E8 with a unique size-exclusion-chromatography method. Anal Chem 2018;90:12390–12394. doi:10.1021/acs.analchem.8b03594
23.Guttman M, et al. The influence of proline isomerization on potency and stability of anti-HIV antibody 10E8. Sci Rep 2020;10:14313. doi:10.1038/s41598-020-71184-7
24.Kwon YD, et al. Structures of HIV-1 neutralizing antibody 10E8 delineate the mechanistic basis of its multi-peak behavior on size-exclusion chromatography. Antibodies (Basel) 2021;10:23. doi:10.3390/antib10020023
25.Theodore DA, et al. Bispecific 10E8.4/iMab broadly neutralizing antibody in people with or without HIV-1: a partially randomized phase 1 trial. Nat Med 2026;32:2533–2545. doi:10.1038/s41591-026-04472-w
27.Perelson AS, Neumann AU, Markowitz M, Leonard JM, Ho DD. HIV-1 dynamics in vivo: virion clearance rate, infected cell life-span, and viral generation time. Science 1996;271:1582–1586. doi:10.1126/science.271.5255.1582 (generation time).
28.Keele BF, et al. Identification and characterization of transmitted and early founder virus envelopes in primary HIV-1 infection. Proc Natl Acad Sci U S A 2008;105:7552–7557. doi:10.1073/pnas.0802203105 (founder multiplicity).
29.Zanini F, Puller V, Brodin J, Albert J, Neher RA. In vivo mutation rates and the landscape of fitness costs of HIV-1. Virus Evol 2017;3:vex003. doi:10.1093/ve/vex003
30.Liao HX, et al. Co-evolution of a broadly neutralizing HIV-1 antibody and founder virus. Nature 2013;496:469–476. doi:10.1038/nature12053. Sequences: GenBank KC247375–KC247667.
31.Zanini F, et al. Population genomics of intrapatient HIV-1 evolution. eLife 2015;4:e11282. doi:10.7554/eLife.11282. Data: ENA PRJEB9618.
32.Romero EV, et al. Recurrent mutations drive rapid HIV escape from two broadly neutralizing antibodies in vivo. Proc Natl Acad Sci U S A 2026;123:e2524021123. doi:10.1073/pnas.2524021123.
33.Marchitto L, et al. Enhanced B cell priming induces broadly neutralizing HIV-1 apex antibodies. Nature 2026;656:712–722. doi:10.1038/s41586-026-10838-4.
34.Katoh K, Standley DM. MAFFT multiple sequence alignment software version 7: improvements in performance and usability. Mol Biol Evol 2013;30:772–780. doi:10.1093/molbev/mst010
35.Zhou T, et al. Structural basis for broad and potent neutralization of HIV-1 by antibody VRC01. Science 2010;329:811–817. doi:10.1126/science.1192819. PDB 3NGB.
36.Klein F, et al. Somatic mutations of the immunoglobulin framework are generally required for broad and potent HIV-1 neutralization. Cell 2013;153:126–138. doi:10.1016/j.cell.2013.03.018. PDB 4JPV (3BNC117 with gp120).
37.Schrödinger, LLC. The PyMOL Molecular Graphics System (used to render the epitope atlas and orientation figure).
38.Mouquet H, et al. Complex-type N-glycan recognition by potent broadly neutralizing HIV antibodies (10-1074). Proc Natl Acad Sci U S A 2012;109:E3268–E3277. doi:10.1073/pnas.1217207109
39.Sok D, et al. Recombinant HIV envelope trimer selects for quaternary-dependent antibodies targeting the trimer apex (PGDM1400). Proc Natl Acad Sci U S A 2014;111:17624–17629. doi:10.1073/pnas.1415789111
40.Huang J, et al. Broad and potent HIV-1 neutralization by a human antibody that binds the gp41–gp120 interface (35O22). Nature 2014;515:138–142. doi:10.1038/nature13601
41.Berman HM, Westbrook J, Feng Z, et al. The Protein Data Bank. Nucleic Acids Res 2000;28:235–242. doi:10.1093/nar/28.1.235
42.Shrake A, Rupley JA. Environment and exposure to solvent of protein atoms. Lysozyme and insulin. J Mol Biol 1973;79:351–371. doi:10.1016/0022-2836(73)90011-9
43.Rego N, Koes D. 3Dmol.js: molecular visualization with WebGL. Bioinformatics 2015;31:1322–1324. doi:10.1093/bioinformatics/btu829
44.Tsai YX, et al. Rapid simulation of glycoprotein structures by grafting and steric exclusion of glycan conformer libraries (GlycoSHIELD). Cell 2024;187:1296–1311. doi:10.1016/j.cell.2024.01.034
46.Michaud-Agrawal N, et al. MDAnalysis: a toolkit for the analysis of molecular dynamics simulations. J Comput Chem 2011;32:2319–2327. doi:10.1002/jcc.21787
