19. From prediction to experiment — what the twin can and cannot say · Dr. Seok's Cancer-Cell Digital Twin, First Steps

We reached the conclusion "switch DDX54 off and the cancer attractor disappears and the cell returns to normal". Now let us see what it takes to carry that sentence into a lab.

Separating what the model can say from what it cannot is the whole of this lesson.

What the model says	What the model does not say
Take DDX54 to 0 and proliferation stops	How far you have to lower it (70%? 90%?)
The attractor disappears	How many days it takes
It moves to the normal attractor	Whether that state equals a real normal cell
There are no side effects inside this circuit	What happens to genes outside the circuit
PDL1 alone is ineffective	Whether it is ineffective in real patients

This is the price of throwing away everything between 0 and 1 — the other side of the computing power we gained in lesson 4.

Python Runs in your browser

import bn

net = bn.twin()
base = net.landscape()

# The model speaks only in all-or-nothing. What about a partial knockdown?
print("the DDX54 levels a Boolean model can express")
print("-" * 56)
for v, name in [(1, "DDX54 = 1  (normal expression)"), (0, "DDX54 = 0  (complete null)")]:
    L = net.landscape({"DDX54": v})
    print(f"  {name:<32} cancer {L.count('cancer'):>3}/{L.total}")

print()
print("A knockdown at the bench is usually a 70-90% reduction. There is no value in between here.")
print("So when the model says 100%, a partial effect in the experiment is the normal outcome.")
print("Closing that gap needs a multi-valued model or a differential-equation model.")

Check that the model gets right what it already knows, and only then ask it what it does not.

Python Runs in your browser

import bn

net = bn.twin()

# First step of model validation: does it reproduce what we already know?
print("the list of 'things already known' the model has to reproduce")
print("-" * 70)
checks = [
    ("cancer arises more readily when p53 is broken",
     net.landscape({"p53": 0}).count("cancer") > net.landscape().count("cancer")),
    ("a normal cell does not proliferate",
     net.label(net.attractor(net.state(p53=1))) == "normal"),
    ("overexpressing DDX54 is harmful",
     net.landscape({"DDX54": 1}).count("cancer") > net.landscape().count("cancer")),
    ("releasing one checkpoint does not break an immune desert",
     net.landscape({"PDL1": 0}).count("cancer") == net.landscape().count("cancer")),
]
for text, ok in checks:
    print(f"  {'[ok]' if ok else '[X] '}  {text}")

print()
print("These four are not 'new predictions' but 'exams to be passed'.")
print("Fail even one and you have no business offering a new prediction.")

Write down predictions without a falsifier and you will claim you were right whatever comes out.

Python Runs in your browser

import bn

print("turning a prediction into an experimental plan")
print("=" * 70)
plan = [
    ("prediction",  "lower DDX54 and proliferation stops, opening the way to immune attack"),
    ("cell assay",  "introduce shDDX54 into a lung cancer line -> growth curve, cell-cycle analysis"),
    ("markers",     "do CD47 and PD-L1 proteins really come down together (flow cytometry)"),
    ("co-culture",  "culture with T cells and see whether they actually kill"),
    ("in vivo",     "Ddx54 knockdown in mouse tumours +/- anti-PD1 -> tumour size, infiltrating lymphocytes"),
    ("falsifier",   "if DDX54 goes down and CD47/PD-L1 do not, the circuit is wrong"),
]
for k, v in plan:
    print(f"  {bn.pad(k, 14)}{v}")

print()
print("The last line is the most important one.")
print("Writing down in advance what observation would prove your model wrong -")
print("that is what makes it science.")

🎮 Prediction checklist

Match each prediction of the model to the experiment it would take to test it.

One experiment per prediction. A prediction with no partner is a prediction you cannot test.

Loading…

🤔 Check yourself

Why can a Boolean model not express a 'partial knockdown'?

🤔 Check yourself

Why check 'does it reproduce what we already know' first?

🤔 Check yourself

Why write a 'falsifier' into the experimental plan in advance?