18. What if the rules are wrong — shake one arrow and see · Dr. Seok's Cancer-Cell Digital Twin, First Steps

Where did the fifteen arrows in our circuit come from? I decided them. In a realistic model they would come from papers and databases, but those too are in the end read and transcribed by people. They can be wrong.

So there is a check you must always run.

Does the conclusion hold if you flip the sign of one arrow?

This is a robustness check. If the conclusion collapses on a single shake, then that conclusion was made by an assumption of mine, not by the data.

One flipped arrow and the cancer attractor is gone entirely — the circuit is broken.

Python Runs in your browser

import bn
import re

def flip(net, i):
    """Build a network with the sign of regulatory edge i reversed."""
    RULES = {
        "DDX54": "DDX54 and not p53", "p53": "not MDM2",
        "MDM2": "DDX54 or p53", "CD47": "DDX54 and not p53",
        "PDL1": "DDX54 and not p53",
        "T_Kill": "Proliferation and not PDL1 and not CD47",
        "Apoptosis": "T_Kill",
        "Proliferation": "DDX54 and not p53 and not Apoptosis",
    }
    u, v, sign = net.edges[i]
    expr = RULES[v]
    new = (expr.replace(u, f"not {u}", 1) if sign > 0
           else expr.replace(f"not {u}", u, 1))
    code = re.sub(r"[A-Za-z_]\w*",
                  lambda m: m.group(0) if m.group(0) in ("and", "or", "not")
                  else f"s['{m.group(0)}']", new)
    n2 = bn.twin()
    ns = {}
    exec("def f(s): return " + code, {}, ns)
    n2.rules[v] = ns["f"]
    return n2, f"{v}* = {new}"

net = bn.twin()
n2, desc = flip(net, 0)
print("original:", "DDX54* = DDX54 and not p53")
print("flipped :", desc)
print()
print("landscape of the flipped circuit")
n2.landscape().summary()

Counted honestly, the denominator is only the cases where cancer remained, not all fifteen.

Python Runs in your browser

import bn, re

def flip(net, i):
    RULES = {
        "DDX54": "DDX54 and not p53", "p53": "not MDM2",
        "MDM2": "DDX54 or p53", "CD47": "DDX54 and not p53",
        "PDL1": "DDX54 and not p53",
        "T_Kill": "Proliferation and not PDL1 and not CD47",
        "Apoptosis": "T_Kill",
        "Proliferation": "DDX54 and not p53 and not Apoptosis",
    }
    u, v, sign = net.edges[i]
    expr = RULES[v]
    new = (expr.replace(u, f"not {u}", 1) if sign > 0
           else expr.replace(f"not {u}", u, 1))
    code = re.sub(r"[A-Za-z_]\w*",
                  lambda m: m.group(0) if m.group(0) in ("and", "or", "not")
                  else f"s['{m.group(0)}']", new)
    n2 = bn.twin(); ns = {}
    exec("def f(s): return " + code, {}, ns)
    n2.rules[v] = ns["f"]
    return n2

net = bn.twin()
print(bn.pad("flipped relation", 32) + bn.pad("cancer attractor", 20) + "DDX54 knockout result")
print("-" * 80)
tested = ok = 0
for i, (u, v, sg) in enumerate(net.edges):
    n2 = flip(net, i)
    n_cancer = n2.landscape().count("cancer")
    label = f"{u} {'->' if sg > 0 else '-|'} {v}"
    if n_cancer == 0:
        print(bn.pad(label, 32) + bn.pad("gone", 20) + "cannot test (no cancer)")
        continue
    tested += 1
    cured = n2.landscape({"DDX54": 0}).count("cancer") == 0
    ok += cured
    print(bn.pad(label, 32) + bn.pad(f"{n_cancer} left", 20)
          + ("still removes the cancer [ok]" if cured else "fails [X]"))

print()
print(f"of {len(net.edges)} shakes, {tested} left cancer standing; DDX54 knockout worked in {ok} of those")

The result of a robustness check is grounds for confidence, not a proof.

Python Runs in your browser

import bn

print("mistakes people commonly make in a robustness check")
print("-" * 66)
print("  [X] reporting '15 shakes, DDX54 worked in all 15'")
print("      -> in seven of them the circuit was broken and cancer had vanished.")
print("         You cannot count 'removed the cancer' in a circuit with no cancer.")
print()
print("  [X] choosing only the shakes that favour your conclusion")
print("      -> fix the set of things to shake in advance and report all of them.")
print()
print("  [X] saying 'the model is correct' because it passed a robustness check")
print("      -> passing says only 'it does not collapse under this much shaking'.")
print("         About genes missing from the circuit entirely it says nothing at all.")

🎮 Shake bench

Press the arrows one at a time to flip them. It tells you on the spot whether the conclusion survives.

Tell apart the cases where the circuit breaks from the cases where it holds and the conclusion still stands.

Loading…

🤔 Check yourself

Why run a robustness check?

🤔 Check yourself

Why must a flip that made the cancer attractor disappear not be counted as 'treatment success'?

🤔 Check yourself

What is the structural reason DDX54 knockout stood up to the shaking?