Sequence is the only input, fixed at 66 aa; sampling settings pinned

README.md

@@ -25,6 +25,11 @@ sections below document the ones that exist.
 Predict how an Intrinsically Disordered Peptide (IDP) behaves as its
 concentration rises, directly from sequence.
 
+The only input is the sequence, and it must be **exactly 66 amino acids** — the
+entire CondenSeq library the model was trained on is 66-mers (all 14,578 of them),
+so other lengths are off-distribution. Sampling is fixed at 512 concentration
+levels and 100 ODE steps, seed 6, matching the offline runs.
+
 Enter a sequence and this section runs the three stages of the CELL-FM condensate
 pipeline end to end:
 
@@ -57,9 +62,10 @@ probability 0.8, and the raw per-image calls as a downloadable CSV.
 
 ### Presets
 
-The dropdown is loaded with NUP98's FG-repeat region — wild type plus its full
-mutation panel (F→S, F→Y, charge and spacer variants). Comparing a mutant's AUC
-against WT is the quickest way to see the model respond to sequence grammar.
+The dropdown is loaded with the NUP98 IDP — wild type (the default) plus its full
+mutation panel, 92 sequences of 66 aa (F→S, F→Y, charge and spacer variants).
+Comparing a mutant's AUC against WT is the quickest way to see the model respond
+to sequence grammar.
 
 ---
 
@@ -71,10 +77,10 @@ fallback (verified bit-identical to the Triton path: same AUC, same AAC, same
 256/256 per-image calls) — but a CPU run of the defaults would take hours and
 time out. Use ZeroGPU or a GPU hardware tier.
 
-The default settings — 512 images at 100 ODE steps — take **~85 s on an A40**
+The fixed settings — 512 images at 100 ODE steps — take **~85 s on an A40**
 (measured: 11.5 s per batch of 64, plus ~9 s of one-off model loading). Runtime
-scales linearly in both settings; lower either in **Sampling settings** for a
-faster, noisier curve.
+scales linearly in both, so a section that needs to be faster should lower them in
+`pipeline.py` (`NUM_IMAGES`, `NUM_STEPS`).
 
 On ZeroGPU the per-call budget is set by `ZEROGPU_DURATION` (default 300 s), which
 covers the defaults with room to spare.

app.py

@@ -58,8 +58,6 @@ INK = "#0b0b0b"
 MUTED = "#52514e"
 SURFACE = "#fcfcfb"
 
-MAX_SEQUENCE_LEN = 2048
-
 
 def load_presets():
     """NUP98 WT plus its mutation panel, as demo starting points."""
@@ -166,19 +164,13 @@ def results_table(result: pipeline.Result) -> pd.DataFrame:
 
 
 @GPU_DECORATOR
-def analyse(sequence, num_images, num_steps, batch_size, seed, progress=gr.Progress()):
-    sequence = pipeline.clean_sequence(sequence)
-    if len(sequence) > MAX_SEQUENCE_LEN:
-        raise gr.Error(f"Sequence is {len(sequence)} aa; the model takes at most {MAX_SEQUENCE_LEN}.")
-
-    result = pipeline.run(
-        sequence,
-        num_images=int(num_images),
-        num_steps=int(num_steps),
-        batch_size=int(batch_size),
-        seed=int(seed),
-        progress=progress,
-    )
+def analyse(sequence, progress=gr.Progress()):
+    try:
+        sequence = pipeline.clean_sequence(sequence)
+    except ValueError as exc:
+        raise gr.Error(str(exc)) from exc
+
+    result = pipeline.run(sequence, progress=progress)
 
     table = results_table(result)
     csv_path = os.path.join("/tmp", "pred_vs_intensity.csv")
@@ -216,20 +208,17 @@ def build_condensate_titration_section():
     with gr.Row():
         with gr.Column(scale=2):
             preset = gr.Dropdown(
-                choices=list(PRESETS), value="NUP98 WT", label="Preset",
-                info="NUP98 FG-repeat WT and its mutation panel. Editing the box below overrides this.",
+                choices=list(PRESETS), value="NUP98 WT", label="Load a sequence",
+                info="NUP98 IDP and its mutation panel; fills the box below, which you can edit.",
             )
             sequence = gr.Textbox(
-                value=DEFAULT_SEQUENCE, lines=4, label="Protein sequence",
-                info="Single-letter amino acids; FASTA headers are stripped.",
+                value=DEFAULT_SEQUENCE, lines=4,
+                label=f"Protein sequence ({pipeline.SEQUENCE_LENGTH} aa)",
+                info=(
+                    f"Exactly {pipeline.SEQUENCE_LENGTH} single-letter amino acids — the whole "
+                    "CondenSeq library is 66-mers. FASTA headers are stripped."
+                ),
             )
-            with gr.Accordion("Sampling settings", open=False):
-                num_images = gr.Slider(64, 1024, value=512, step=64, label="Images (concentration levels)",
-                                       info="More levels give a smoother curve and cost proportionally more time.")
-                num_steps = gr.Slider(20, 200, value=100, step=10, label="ODE steps per image")
-                batch_size = gr.Slider(16, 256, value=64, step=16, label="Batch size")
-                seed = gr.Number(value=pipeline.DEFAULT_SEED, precision=0, label="Random seed",
-                                 info="Generation starts from a random latent; fix the seed for a repeatable curve.")
             run_button = gr.Button("Run analysis", variant="primary")
 
         with gr.Column(scale=3):
@@ -255,7 +244,7 @@ def build_condensate_titration_section():
     preset.change(lambda name: PRESETS[name], inputs=preset, outputs=sequence)
     run_button.click(
         analyse,
-        inputs=[sequence, num_images, num_steps, batch_size, seed],
+        inputs=sequence,
         outputs=[curve_plot, sample_plot, summary, table, download],
     )
 

pipeline.py

@@ -48,11 +48,22 @@ WINDOW_DIVISOR = 8  # window = num_images // 8, matching ana_all_mutation_log_sc
 
 VALID_AA = set("ACDEFGHIKLMNPQRSTVWY")
 
+# CondenSeq is a fixed-length library: all 14,578 training sequences are exactly
+# 66 residues. Anything else is off-distribution, so the demo requires this length
+# rather than silently padding or cropping.
+SEQUENCE_LENGTH = 66
+
 # Generation starts from a random latent, so the curve moves a little between runs.
 # The offline scripts fixed this at 6; keeping the same default makes the demo's
 # numbers repeatable and comparable to them.
 DEFAULT_SEED = 6
 
+# Fixed sampling settings. The concentration ladder length and ODE step count match
+# the offline runs, so demo numbers stay comparable to the published ones.
+NUM_IMAGES = 512
+NUM_STEPS = 100
+BATCH_SIZE = 64
+
 
 def resolve_weights(filename: str) -> str:
     """Local directory if CELLFM_LOCAL_WEIGHTS is set, else pull from the model repo."""
@@ -191,6 +202,12 @@ def clean_sequence(sequence: str) -> str:
             f"Sequence contains non-standard residues: {', '.join(bad)}. "
             "Use the 20 standard amino acids."
         )
+    if len(seq) != SEQUENCE_LENGTH:
+        raise ValueError(
+            f"Sequence is {len(seq)} residues; CELL-FM CondenSeq takes exactly "
+            f"{SEQUENCE_LENGTH}. The whole training library is {SEQUENCE_LENGTH}-mers, "
+            "so other lengths are outside what the model has seen."
+        )
     return seq
 
 
@@ -298,8 +315,8 @@ def sample_panel(images, intensities, k=6):
     return images[idx, 1].astype(np.float32) / 65535.0, intensities[idx]
 
 
-def run(sequence, num_images=512, num_steps=100, batch_size=64, seed=DEFAULT_SEED,
-        progress=None) -> Result:
+def run(sequence, num_images=NUM_IMAGES, num_steps=NUM_STEPS, batch_size=BATCH_SIZE,
+        seed=DEFAULT_SEED, progress=None) -> Result:
     """The whole pipeline: sequence in, titration curve and its areas out."""
     sequence = clean_sequence(sequence)