16. Three traps — smashing the gauge, cutting the brake, the lure of zero · Dr. Seok's Cancer-Cell Digital Twin, First Steps

Before you trust a screen there are three things you must check. In all three the computation is perfectly correct and the conclusion is wrong.

Smashing the gauge — targeting the very node you read the phenotype from
Cutting the brake — switching off a tumour suppressor because it is 'cancer-related'
The lure of zero — judging on the single number 'cancer left = 0'

That is why net.druggable() takes the gauge nodes out in advance.

Python Runs in your browser

import bn

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

# Trap 1 - what if the gauge is allowed into the target list?
print("putting the phenotype readout nodes into the target list")
print("-" * 62)
for g in net.readouts:
    L = net.landscape({g: 0})
    r = L.report(base)
    print("  " + bn.pad(g + " knockout", 28)
          + f"reversion {r['reversion']:>5.0f}%   cancer left {r['cancer_left']}")

print()
print("Proliferation knockout gives 'cancer left = 0' - it looks like full marks.")
print("Of course it does. We judge cancer by Proliferation,")
print("and we just pinned Proliferation to 0.")
print("We smashed the thermometer and announced there is no fever.")

96 harmed states. Taken into the lab, that would have been an expensive failure.

Python Runs in your browser

import bn

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

# Trap 2 - what if you switch off every 'cancer-related gene' on principle?
print("which way should p53 be manipulated?")
print("-" * 62)
for v, name in [(0, "p53 knockout (switch off)"), (1, "p53 overexpression (switch on)")]:
    L = net.landscape({"p53": v})
    r = L.report(base)
    print("  " + bn.pad(name, 30)
          + f"reversion {r['reversion']:>5.0f}%   harm {r['harm']:>3}   "
          + f"cancer left {r['cancer_left']}")

print()
print("It is the exact opposite. p53 is the brake (a tumour suppressor), so it must go ON.")
print("The single phrase 'cancer gene' mixes accelerators and brakes together -")
print("having a computer filter that out in advance is the practical value of screening.")

Look at the last row reading 100%. Metrics do not filter out fakes for you.

Python Runs in your browser

import bn

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

# Trap 3 - gather the treatments with 'cancer left = 0' and look closely
print(bn.pad("treatment", 28) + bn.pad("cancer left", 14) + bn.pad("reversion", 12)
      + bn.pad("dormant", 10) + "what actually happened")
print("-" * 92)
for name, p in [("DDX54 knockout",         {"DDX54": 0}),
                ("MDM2 knockout",          {"MDM2": 0}),
                ("PDL1+CD47 combined",     {"PDL1": 0, "CD47": 0}),
                ("Proliferation knockout", {"Proliferation": 0})]:
    L = net.landscape(p)
    r = L.report(base)
    # Check whether a gauge node was touched - that comes before any number
    if any(g in net.readouts for g in p):
        what = "only switched off the gauge (fake)"
    elif r["reversion"] == 100:
        what = "returned to normal (real reversion)"
    elif r["dormant"] == 100:
        what = "growth suppressed only (dormancy)"
    else:
        what = "partial effect"
    print(bn.pad(name, 28) + bn.pad(r["cancer_left"], 14)
          + bn.pad(f"{r['reversion']:.0f}%", 12) + bn.pad(f"{r['dormant']:.0f}%", 10) + what)

print()
print("All four give 'cancer left = 0', and three of the four even agree on 'reversion 100%'.")
print("The numbers cannot tell them apart - you have to look at what was touched.")

🎮 Trap bench

Press each of the three traps in turn. Each one tells you what went wrong.

The arithmetic can be right and the conclusion still wrong.

Loading…

🤔 Check yourself

Why is 'cancer left = 0' from a Proliferation knockout meaningless?

🤔 Check yourself

What is the lesson of the p53 knockout giving a harmful result?

🤔 Check yourself

What best summarises why 'cancer left = 0' alone must not pick your target?