bindsight 0.3.5__py3-none-any.whl

bindsight/report/web/templates/evidence.html.j2

@@ -0,0 +1,341 @@
+{% extends "base.html.j2" %}
+{% block title %}Evidence — bindsight{% endblock %}
+
+{% block content %}
+<h1>Evidence</h1>
+<p class="lede">
+  Everything on this page is read from <code>benchmarks/</code> in the
+  repository — the same files the manuscripts cite. Nothing is illustrative,
+  recomputed on the fly, or typed by hand.
+</p>
+
+{% if not study and not designer %}
+  <div class="note note--warn">
+    <span class="note__glyph" aria-hidden="true">!</span>
+    <div class="note__body">
+      <p><strong>No evidence tree in this install.</strong> <code>benchmarks/</code>
+      ships with the repository rather than the wheel.</p>
+      <p class="small"><a href="{{ github }}/tree/main/benchmarks">Read it on GitHub →</a></p>
+    </div>
+  </div>
+{% endif %}
+
+{% if study %}
+<h2 id="discovery">Does it resurface antigens we already trust?</h2>
+<p class="prose">
+  <strong>{{ study.n_scored }} antigen-cohort pairs across {{ study.n_projects }}
+  real TCGA projects.</strong> Each cohort is every patient in the project who
+  contributed both a primary tumour and a solid-tissue normal, contrasted
+  pairwise on the patient. That is a subset of the project — a tumour with no
+  matched normal is not in it — and the subset is chosen by sample availability,
+  never by the antigen under test. No cohort is stratified by any biomarker.
+</p>
+
+<h3>Recall, on the pre-registered denominator</h3>
+{% set cascade = (study.recall_cascade or {}).get('all', {}) %}
+{% set w20 = ((cascade.get('at_k') or {}).get('recall@20') or {}).get('wilson') or {} %}
+<div class="card">
+  <div class="row">
+    <div class="stat">
+      <span class="stat__label">Recall @ 20</span>
+      <span class="stat__value">{{ w20.get('numerator', ABSENT) }}/{{ w20.get('denominator', ABSENT) }}</span>
+      <span class="stat__ci">{{ interval(w20.get('low'), w20.get('high'), digits=2, confidence=w20.get('confidence')) }}</span>
+      <span class="stat__note">
+        Approved-agent antigens only — the tiers the study pre-registered.
+      </span>
+    </div>
+  </div>
+  {% if recall %}
+    <div data-chart="bars" data-spec='{{ tojson(recall) }}' style="margin-top:1rem"></div>
+    <p class="chart__caption">
+      Each bar is a cutoff; the figure beside it is the numerator over the
+      denominator, not a percentage on its own.
+    </p>
+  {% endif %}
+</div>
+
+{% if study.headline %}
+{% set top = study.headline %}
+<h3>The best-ranked antigen, and why it is shown separately</h3>
+<div class="card">
+  <div class="row">
+    <div class="stat">
+      <span class="stat__label">{{ top.symbol }} rank</span>
+      <span class="stat__value">{{ top.rank }}</span>
+      <span class="stat__of">of {{ fmt_int(top.shortlist_size) }} candidates in
+        {{ top.project|replace('TCGA-','') }}</span>
+    </div>
+    <div class="stat">
+      <span class="stat__label">log2 fold change</span>
+      <span class="stat__value">{{ fmt(top.log2fc) }}</span>
+      <span class="stat__ci">
+        {% if top.get('lfc_se') is not none %}± {{ fmt(top.lfc_se) }} (SE){% else %}{{ ABSENT }}{% endif %}
+      </span>
+    </div>
+  </div>
+
+  <div class="note note--info" style="margin-top:1rem">
+    <span class="note__glyph" aria-hidden="true">i</span>
+    <div class="note__body">
+      <p class="small" style="margin:0">
+        <strong>{{ top.symbol }} is not in the recall figure above.</strong> Its
+        regulatory tier is <code>{{ top.get('tier', ABSENT) }}</code>, and the
+        study pre-registered
+        <code>{{ study.primary_tiers|join(', ') or ABSENT }}</code> as the primary
+        denominator. The two numbers answer different questions and are kept
+        apart deliberately: placing them side by side invites a reader to bind
+        the “1” in a recall fraction to a rank of 1 for a different antigen.
+      </p>
+    </div>
+  </div>
+
+  {% if top.get('note') %}
+    <div class="note note--warn" style="margin-top:.75rem">
+      <span class="note__glyph" aria-hidden="true">!</span>
+      <div class="note__body">
+        <p class="small" style="margin:0">
+          <strong>What qualifies this result.</strong> {{ top.note }}
+        </p>
+      </div>
+    </div>
+  {% endif %}
+</div>
+{% endif %}
+
+<h3 id="nulls">Is the ranking doing anything? Two nulls, pointing different ways</h3>
+<p class="prose">
+  Both are reported. Reporting only the flattering one would make the other the
+  finding.
+</p>
+
+<div class="grid grid--2">
+  <section class="card">
+    <h4 class="card__title">
+      <span class="badge badge--warn"><span class="badge__glyph" aria-hidden="true">–</span>negative</span>
+      Against abundance-matched decoys
+    </h4>
+    <div class="stat">
+      <span class="stat__label">Survive Benjamini-Hochberg</span>
+      <span class="stat__value">{{ study.decoy_surviving_correction }}</span>
+      <span class="stat__of">of {{ study.decoy_rows|length }} pairs
+        ({{ study.decoy_nominal }} nominally significant at 0.05)</span>
+    </div>
+    <p class="small" style="margin-top:.75rem">
+      Each antigen is compared with genes matched to it on abundance and
+      dispersion quintile, drawn from the same eligible surfaceome: would a gene
+      that merely <em>looks</em> like this antigen have ranked as well?
+    </p>
+    {% if study.smallest_attainable_decoy_bh is not none %}
+      <p class="small muted">
+        Read with the panel's resolution in mind: each pair's p is bounded below
+        by its own stratum size, so the smallest adjusted value this panel could
+        have produced is <strong>{{ fmt(study.smallest_attainable_decoy_bh, 3) }}</strong>.
+        Under 0.05 — a pair genuinely could have survived, but only one sitting
+        essentially on its floor. A measurement, not a foregone conclusion, and
+        not a sensitive one.
+      </p>
+    {% endif %}
+  </section>
+
+  <section class="card">
+    <h4 class="card__title">
+      <span class="badge badge--ok"><span class="badge__glyph" aria-hidden="true">✓</span>positive</span>
+      Against cohorts carrying no panel antigen, and against a permuted indication
+    </h4>
+    {% set gap = study.indication_gap %}
+    {% set spec = study.specificity_null %}
+    {% if gap.get('point') is not none %}
+      <div class="stat">
+        <span class="stat__label">Within-antigen difference</span>
+        <span class="stat__value">{{ fmt(gap.point, 3) }}</span>
+        <span class="stat__ci">
+          {{ interval(gap.low, gap.high, confidence=gap.get('confidence')) }}
+          over {{ gap.get('n_clusters', ABSENT) }} antigens — excludes zero
+        </span>
+      </div>
+    {% endif %}
+    {% if spec.get('p_value') is not none %}
+      {#
+        A second analysis, not a p-value for the difference above. The difference
+        is a cluster bootstrap over the antigens that appear in the calibration
+        cohorts; this is an exact permutation test over the antigens carrying a
+        single indication, because the permutation assigns one cohort per antigen.
+        The two denominators differ, and stating the p under the difference
+        without saying so invited the reader to read it as its significance.
+      #}
+      <p class="small" style="margin-top:.75rem">
+        <strong>A second test, on its own denominator.</strong> Against every
+        reassignment of cohorts to antigens, the mean standing is
+        <strong>{{ fmt(spec.get('observed'), 3) }}</strong>: exact permutation p =
+        <strong>{{ fmt_p(spec.p_value) }}</strong> over
+        {{ fmt_int(spec.get('n_permutations')) }} enumerated orderings of
+        {{ fmt_int(spec.get('n_antigens')) }} antigens{% if study.specificity_is_at_its_floor %} —
+        the floor for this design, so it is as extreme as an exhaustive
+        enumeration of this panel can show rather than vanishingly small{% endif %}.
+        Restricted to antigens with a single indication, because the permutation
+        assigns one cohort to each. This p is not the significance of the
+        difference above, which is reported with an interval and no p.
+      </p>
+    {% endif %}
+    {% if nulls %}
+      <div data-chart="forest" data-spec='{{ tojson(nulls) }}'></div>
+    {% endif %}
+    <p class="small muted">
+      Each antigen is its own control: the comparison is within antigen, not
+      between two groups of them.
+    </p>
+  </section>
+</div>
+
+<h3>Every pair</h3>
+<div class="table-wrap">
+  <table>
+    <thead>
+      <tr>
+        <th>Antigen</th><th>Cohort</th>
+        <th class="num">log2FC</th><th class="num">padj</th>
+        <th class="num">Rank</th>
+        <th class="num"><abbr title="Where it would have ranked with every gate removed">Counterfactual</abbr></th>
+        <th>Outcome</th>
+      </tr>
+    </thead>
+    <tbody>
+    {% for p in study.pairs %}
+      <tr>
+        <td><strong>{{ p.get('symbol', ABSENT) }}</strong></td>
+        <td>{{ p.get('project', ABSENT)|replace('TCGA-','') }}</td>
+        <td class="num">{{ fmt(p.get('log2fc')) }}</td>
+        <td class="num">{{ fmt_p(p.get('padj')) }}</td>
+        <td class="num">
+          {% if p.get('rank') is not none %}{{ p.rank }} of {{ fmt_int(p.get('shortlist_size')) }}
+          {% else %}{{ ABSENT }}{% endif %}
+        </td>
+        <td class="num">{{ fmt_int(p.get('counterfactual_rank')) }}</td>
+        <td class="small">{{ p.get('outcome_class', ABSENT)|replace('_',' ') }}</td>
+      </tr>
+    {% endfor %}
+    </tbody>
+  </table>
+</div>
+<p class="chart__caption">
+  A rank is printed with the size of the list it sits in. The counterfactual
+  column is what separates “a filter excluded it” from “the ranking placed it low”.
+</p>
+{% endif %}
+
+{% if designer %}
+<h2 id="design">The design half, and the control that withdrew its headline</h2>
+<div class="card">
+  <div class="row">
+    <div class="stat">
+      <span class="stat__label">Designs</span>
+      <span class="stat__value">{{ fmt_int(designer.n_designs) }}</span>
+      <span class="stat__of">from {{ designer.n_trajectories or ABSENT }} backbones</span>
+    </div>
+    <div class="stat">
+      <span class="stat__label">Mean ipTM</span>
+      <span class="stat__value">{{ fmt(designer.mean_iptm) }}</span>
+      <span class="stat__of">validator {{ designer.validator or ABSENT }}</span>
+    </div>
+    <div class="stat">
+      <span class="stat__label">Hardware</span>
+      <span class="stat__value" style="font-size:var(--t-md)">{{ designer.gpu or ABSENT }}</span>
+      <span class="stat__of">bindsight {{ designer.bindsight_version or ABSENT }}</span>
+    </div>
+  </div>
+
+  <div class="withdrawn">
+    <strong>success@0.65 is withdrawn as a measure of design quality.</strong>
+    A paired control folded every design beside a shuffle of its own sequence —
+    same length, same amino-acid composition, only the order destroyed. Under a
+    seeded validator averaging five diffusion draws the shuffles cleared the
+    customary ipTM 0.65 bar <em>more</em> often than the designs did. The ipTM
+    values are real Boltz-2 outputs; what does not follow is that the rate
+    measures the designs. The control's predictions were written down before its
+    data arrived.
+  </div>
+
+  {% if structures %}
+    <h3 class="card__title" style="margin-top:1.5rem">The complexes themselves</h3>
+    <p class="small muted" style="margin-top:0">
+      Boltz-2's predicted binder&ndash;target complex for each design, as committed
+      to this repository. <span class="screen-only">Rotate it.</span><span
+      class="print-only">Each is committed as mmCIF under
+      <code>benchmarks/designer_benchmark/</code>; the viewer is not printed
+      because a WebGL canvas prints blank.</span> The numbers above are summaries
+      of these files, not of anything else.
+    </p>
+    <div class="field viewer-picker" style="max-width:32rem">
+      <label for="binder-pick">Design</label>
+      <select id="binder-pick" data-binder-picker>
+        {% for b in structures %}
+          <option value="{{ b.binder_id }}">
+            {{ b.binder_id }} — ipTM {{ fmt(b.iptm) }}
+          </option>
+        {% endfor %}
+      </select>
+    </div>
+    <div id="viewer" class="viewer" role="img" data-binder-viewer
+         data-structure-url="/api/structure/{id}"
+         aria-label="Three-dimensional view of the predicted binder-target complex">
+      <p class="viewer__fallback small muted" role="alert">
+        The structure viewer needs JavaScript. Every complex is also committed as
+        mmCIF under <code>benchmarks/designer_benchmark/</code>, which is the file
+        this viewer reads.
+      </p>
+    </div>
+    <div class="viewer-legend">
+      <span><span class="swatch" style="background:#9aa5b1"></span>
+        chain&nbsp;<strong>T</strong> — the target</span>
+      <span><span class="swatch" style="background:#0f7d73"></span>
+        chain&nbsp;<strong>B</strong> — the designed binder</span>
+    </div>
+    <p class="chart__caption">
+      The chain letters are in the file, so the two are distinguishable without
+      the colours — on a projector, in greyscale, and to a reader who does not
+      separate teal from grey.
+    </p>
+  {% endif %}
+
+  {% if paired %}
+    <div data-chart="pairedStrip" data-spec='{{ tojson(paired) }}' style="margin-top:1rem"></div>
+    <p class="chart__caption">
+      One line per pair, so the pairing is visible rather than two clouds. A red
+      line is a pair whose shuffle scored higher than the design it came from.
+    </p>
+  {% endif %}
+</div>
+{% endif %}
+
+<h2>Reproducing any of this</h2>
+<p class="prose">
+  Three tiers. The first needs nothing but a clone and runs in seconds; it
+  regenerates the published pages from committed files and should produce no
+  diff at all.
+</p>
+<pre><code>git clone {{ github }}
+cd bindsight
+pip install -e ".[discover,report]"
+python scripts/build_docs_results.py   # rewrites the published pages; nothing should change</code></pre>
+<p class="small muted">
+  Re-running the study needs network and a few CPU-hours; re-running the design
+  needs a GPU. Both chains are documented in the repository.
+</p>
+{% endblock %}
+
+{% block scripts %}
+{% if structures %}
+{#- Vendored, not from a CDN: `bindsight ui` is meant to work on a laptop with
+    no connection once installed, and a page that silently needs one does not.
+    The literal path stays in this template rather than moving into the script:
+    tests/test_web_ui.py sweeps .j2, .py, .css and .js for references to
+    every vendored asset. .js was added to that set when the viewer's script
+    left this template for a file of its own; before that, a path living
+    only in a .js would have made the sweep go quietly blind.
+
+    The viewer itself is bindsight/report/web/static/binder_viewer.js, shared
+    with the published documentation site so the two cannot diverge. -#}
+<script src="/static/vendor/3Dmol-min.js"></script>
+<script src="/static/binder_viewer.js"></script>
+{% endif %}
+{% endblock %}

bindsight/report/web/templates/overview.html.j2

@@ -0,0 +1,140 @@
+{% extends "base.html.j2" %}
+{% block title %}bindsight — RNA-seq to de novo binder design{% endblock %}
+
+{% block content %}
+<h1>bindsight</h1>
+<p class="lede">
+  An open-source pipeline that joins cohort RNA-seq target discovery to de novo
+  protein binder design in one reproducible workflow, with machine-readable
+  provenance from every ranked binder back to the patient samples it came from.
+</p>
+
+<p class="prose">
+  Genomics tooling stops at <em>“here are the interesting genes.”</em>
+  Protein-design tooling starts at <em>“given a target…”</em>. The bridge between
+  them is built ad-hoc, per project, and rarely reproducibly. bindsight ships
+  that bridge as one tool and carries an audit trail across the join.
+</p>
+
+<div class="btn-row">
+  <a class="btn btn--primary" href="/evidence">See what it has shown</a>
+  <a class="btn" href="/try">Run the demo</a>
+  <a class="btn" href="/your-data">Use your own data</a>
+</div>
+
+<h2>What it has, and has not, shown</h2>
+<p class="prose muted">
+  Stated here rather than further in, because the parts that did not work are
+  the reason to trust the parts that did.
+</p>
+
+{% if not have_evidence %}
+  <div class="note note--warn">
+    <span class="note__glyph" aria-hidden="true">!</span>
+    <div class="note__body">
+      <p><strong>The evidence is not available in this install.</strong></p>
+      <p class="small">
+        <code>benchmarks/</code> ships with the repository, not the wheel. Clone
+        the repository to see the study, the controls and the binders, or read
+        them on <a href="{{ github }}/tree/main/benchmarks">GitHub</a>.
+      </p>
+    </div>
+  </div>
+{% else %}
+
+<div class="grid grid--2">
+  <section class="card">
+    <h3 class="card__title">Discovery — one result survived its controls</h3>
+    {% if study %}
+      {% set gap = study.indication_gap %}
+      {% if gap.get('point') is not none %}
+        <div class="stat">
+          <span class="stat__label">Indication-specific gap</span>
+          <span class="stat__value">{{ fmt(gap.point, 3) }}</span>
+          <span class="stat__ci">{{ interval(gap.low, gap.high, confidence=gap.get('confidence')) }}
+            over {{ gap.get('n_clusters', ABSENT) }} antigens — excludes zero</span>
+        </div>
+        <p class="small muted" style="margin-top:.75rem">
+          Known antigens rank higher in the cancer they are actually used in than
+          in cancers that are not theirs. This is a claim about the ordering, not
+          about any individual hit.
+        </p>
+      {% endif %}
+      <p class="small">
+        Against abundance-matched decoys the result is <strong>negative</strong>:
+        {{ study.decoy_nominal }} of {{ study.decoy_rows|length }} pairs are
+        nominally significant and
+        <strong>{{ study.decoy_surviving_correction }}</strong> survive
+        Benjamini-Hochberg.
+      </p>
+    {% endif %}
+    <p><a href="/evidence">The study, its denominators and both nulls →</a></p>
+  </section>
+
+  <section class="card">
+    <h3 class="card__title">Design — the headline was withdrawn</h3>
+    {% if designer %}
+      <div class="stat">
+        <span class="stat__label">Binders designed and folded</span>
+        <span class="stat__value">{{ fmt_int(designer.n_designs) }}</span>
+        <span class="stat__of">
+          from {{ designer.n_trajectories or ABSENT }} backbones on
+          {{ designer.gpu or 'a free GPU' }} — structures committed
+        </span>
+      </div>
+      <div class="withdrawn">
+        <strong>success@0.65 is withdrawn as a measure of design quality.</strong>
+        A paired control folded each design beside a shuffle of its own sequence;
+        the shuffles cleared the bar more often than the designs did. The ipTM
+        values are real Boltz-2 outputs — what does not follow is that the rate
+        measures the designs.
+      </div>
+    {% endif %}
+    <p><a href="/evidence#design">The control, and what it refuted →</a></p>
+  </section>
+</div>
+
+<div class="grid grid--2" style="margin-top:1rem">
+  <section class="card">
+    <h3 class="card__title">Scope</h3>
+    {% if study %}
+      <div class="row">
+        <div class="stat">
+          <span class="stat__label">Cohorts</span>
+          <span class="stat__value">{{ study.n_projects }}</span>
+          <span class="stat__of">whole, unstratified TCGA projects</span>
+        </div>
+        <div class="stat">
+          <span class="stat__label">Antigen-cohort pairs</span>
+          <span class="stat__value">{{ study.n_scored }}</span>
+          <span class="stat__of">panel fixed in git before any cohort ran</span>
+        </div>
+        <div class="stat">
+          <span class="stat__label">Surfaceome</span>
+          <span class="stat__value">{{ fmt_int(study.surfaceome_size) }}</span>
+          <span class="stat__of">eligible accessions</span>
+        </div>
+      </div>
+    {% endif %}
+  </section>
+
+  <section class="card">
+    <h3 class="card__title">What has not been done</h3>
+    <ul class="small">
+      <li>
+        The three-way designer comparison against BindCraft and BoltzGen is
+        <strong>not run</strong>: both need 24–32&nbsp;GB against anything larger
+        than a small domain, so those arms require paid backends.
+      </li>
+      <li>
+        The calibration's own power analysis puts <strong>142 paired designs</strong>
+        at 80% power to detect a 0.05 ipTM difference. The committed control has 20.
+      </li>
+    </ul>
+  </section>
+</div>
+{% endif %}
+
+<h2>Citing this</h2>
+<p class="prose">{{ citation() }} <a href="{{ github }}">{{ github }}</a></p>
+{% endblock %}

bindsight/report/web/templates/runs.html.j2

@@ -0,0 +1,125 @@
+{% extends "base.html.j2" %}
+{% block title %}Runs — bindsight{% endblock %}
+
+{% block content %}
+<h1>Runs</h1>
+<p class="lede">Runs found under <code>{{ root }}</code>.</p>
+
+{% if not runs %}
+  <div class="note note--info">
+    <span class="note__glyph" aria-hidden="true">i</span>
+    <div class="note__body">
+      <p style="margin:0">Nothing here yet.</p>
+      <p class="small" style="margin:.5rem 0 0">
+        Run the demo, or point bindsight at your own cohort.
+      </p>
+      <div class="btn-row" style="margin-bottom:0">
+        <a class="btn btn--primary" href="/try">Run the demo</a>
+        <a class="btn" href="/your-data">Use your own data</a>
+      </div>
+    </div>
+  </div>
+{% else %}
+  <div class="table-wrap">
+    <table>
+      <thead>
+        <tr>
+          <th>Run</th>
+          <th><abbr title="The PROV-O record of what produced this run">Manifest</abbr></th>
+          <th>Candidates</th>
+          <th>Path</th>
+        </tr>
+      </thead>
+      <tbody>
+      {% for r in runs %}
+        <tr>
+          <td><strong>{{ r.name }}</strong></td>
+          <td>
+            {% if r.has_manifest %}
+              <span class="badge badge--ok"><span class="badge__glyph" aria-hidden="true">✓</span>recorded</span>
+            {% else %}
+              <span class="badge badge--warn"><span class="badge__glyph" aria-hidden="true">!</span>absent</span>
+            {% endif %}
+          </td>
+          <td>
+            {% if r.verdict == 'failed' %}
+              <span class="badge badge--err"><span class="badge__glyph" aria-hidden="true">✗</span>a stage failed</span>
+            {% elif r.has_candidates %}
+              <span class="badge badge--ok"><span class="badge__glyph" aria-hidden="true">✓</span>present</span>
+            {% else %}
+              <span class="badge badge--warn"><span class="badge__glyph" aria-hidden="true">!</span>absent</span>
+            {% endif %}
+          </td>
+          <td class="small muted"><code>{{ r.path }}</code></td>
+        </tr>
+      {% endfor %}
+      </tbody>
+    </table>
+  </div>
+
+  {% set broken = runs | selectattr('failures') | list %}
+  {% if broken %}
+    <div class="note note--err">
+      <span class="note__glyph" aria-hidden="true">✗</span>
+      <div class="note__body">
+        <p style="margin:0"><strong>Some runs did not complete.</strong></p>
+        <p class="small">
+          These are not runs that found nothing — a stage crashed, and the empty
+          tables it left behind look identical. The stage and its error:
+        </p>
+        <ul class="small">
+          {% for r in broken %}
+            {% for f in r.failures %}
+              <li><code>{{ r.name }}</code> — <strong>{{ f.name }}</strong>: {{ f.error }}</li>
+            {% endfor %}
+          {% endfor %}
+        </ul>
+      </div>
+    </div>
+  {% endif %}
+  {% set degraded = runs | selectattr('open_targets') | list %}
+  {% if degraded %}
+    <div class="note note--warn">
+      <span class="note__glyph" aria-hidden="true">!</span>
+      <div class="note__body">
+        <p style="margin:0"><strong>Some shortlists were not drawn from the whole cohort.</strong></p>
+        <p class="small">
+          Open Targets is the only genome-wide Ensembl&nbsp;&rarr;&nbsp;UniProt source
+          here. Where it did not answer, the only mapping left is a bundled table of a
+          few well-known genes, and every gene missing from that table was dropped
+          before the surfaceome filter.
+        </p>
+        <ul class="small">
+          {% for r in degraded %}
+            {% set ot = r.open_targets %}
+            <li>
+              <code>{{ r.name }}</code> &mdash;
+              <strong>{{ ot.degraded }} of {{ ot.total }}</strong> candidates mapped from the
+              bundled offline table
+              ({% for b in ot.breakdown %}<code>{{ b.status }}</code>&nbsp;&times;&nbsp;{{ b.count }}{% if not loop.last %}, {% endif %}{% endfor %}).
+              {% if ot.total_outage %}
+                Open Targets answered for none of them, so that table's handful of genes
+                was the entire mapping. <strong>This is not a genome-wide discovery</strong>:
+                a gene's absence from it is no evidence about that gene.
+              {% else %}
+                Open Targets answered for the rest, so the genes it failed on were dropped
+                unless that table happened to list them. <strong>This shortlist is
+                incomplete.</strong>
+              {% endif %}
+            </li>
+          {% endfor %}
+        </ul>
+        <p class="small" style="margin-bottom:0">
+          Re-run with Open Targets reachable for a full discovery.
+        </p>
+      </div>
+    </div>
+  {% endif %}
+  <p class="chart__caption">
+    A run missing its manifest or its candidate table is shown as <em>absent</em>,
+    not as a run that produced nothing. Those are different findings, and
+    reporting the second for the first is how a crashed stage reads as an empty
+    result.
+  </p>
+{% endif %}
+{% endblock %}