Pv
P.T. van Berkel
info
Please Note
<p>This page displays the records of the person named above and is not linked to a unique person identifier. This record may need to be merged to a profile.</p>
1 records found
1
Valuing Pre-clinical Decisions
Gate structure, valuation engine and the ordering of experiments in an immuno-oncology programme
Most oncology programmes fail pre-clinically, so experiment order determines how much is spent before termination. Yet pre-clinical development is conventionally modelled as one gate with one probability of success, which cannot order experiments. Question-based clinical development (QBCD) instead splits stages into the questions a programme must answer. This thesis applies QBCD to the pre-clinical phase and separates two sources of value: the gate premium from finer staging and the engine premium from a binomial real-options lattice. To our knowledge, no published application separates these effects.
One pre-clinical immuno-oncology asset was valued in a 2×2 factorial design crossing two engines, a risk-adjusted decision tree and a binomial lattice, with two paths, conventional phases and six company-elicited question gates. Both paths shared total cost, duration and probability of success, and both engines the same discount rate. Question order and concurrency were searched exhaustively under precedence constraints, with sensitivity, Monte Carlo and four-indication robustness checks.
Decomposing the pre-clinical block added €4.15 M (17.1 % of the phase-based €24.25 M), whereas the lattice added €0.13 M (0.5 %), a thirty-two-fold difference. Continuing was worth about four times the gate cost even at the tightest gate, so abandonment was rarely optimal and a lattice earns its place mainly on marginal assets. Discounting the lattice at a lower rate than the tree, as published comparisons often do, inflated the engine premium 500-fold.
The optimal ordering was reproduced by sorting questions on ascending cost-to-failure ratio, a 1960 rule this thesis extends to a discounted form, so finding the best sequence is a sort, not a search. Re-ordering the company plan was worth €0.79 M and running questions concurrently added €2.05 M, together 10.0 % of the original plan’s value. Of the 28.8 % increase an rNPV appraisal would report, only that 10.0 % is under the team’s own control. Across 10 000 Monte Carlo scenarios, the question-based tree was worth starting in 97.5 % of cases and the phase-based tree in 76.3 %, and the discounted rule matched the exhaustive optimum in every scenario.
Three decisions follow. Order experiments by ascending cost-to-failure ratio and run them concurrently where precedence allows. Admit an extra experiment only if it lifts a later gate’s probability of success enough to repay its cost and the delay it imposes. Act on differences between execution plans, not on absolute valuations or the choice of engine. What an asset is worth at this stage cannot be established with certainty, but what should be done next can. ...
One pre-clinical immuno-oncology asset was valued in a 2×2 factorial design crossing two engines, a risk-adjusted decision tree and a binomial lattice, with two paths, conventional phases and six company-elicited question gates. Both paths shared total cost, duration and probability of success, and both engines the same discount rate. Question order and concurrency were searched exhaustively under precedence constraints, with sensitivity, Monte Carlo and four-indication robustness checks.
Decomposing the pre-clinical block added €4.15 M (17.1 % of the phase-based €24.25 M), whereas the lattice added €0.13 M (0.5 %), a thirty-two-fold difference. Continuing was worth about four times the gate cost even at the tightest gate, so abandonment was rarely optimal and a lattice earns its place mainly on marginal assets. Discounting the lattice at a lower rate than the tree, as published comparisons often do, inflated the engine premium 500-fold.
The optimal ordering was reproduced by sorting questions on ascending cost-to-failure ratio, a 1960 rule this thesis extends to a discounted form, so finding the best sequence is a sort, not a search. Re-ordering the company plan was worth €0.79 M and running questions concurrently added €2.05 M, together 10.0 % of the original plan’s value. Of the 28.8 % increase an rNPV appraisal would report, only that 10.0 % is under the team’s own control. Across 10 000 Monte Carlo scenarios, the question-based tree was worth starting in 97.5 % of cases and the phase-based tree in 76.3 %, and the discounted rule matched the exhaustive optimum in every scenario.
Three decisions follow. Order experiments by ascending cost-to-failure ratio and run them concurrently where precedence allows. Admit an extra experiment only if it lifts a later gate’s probability of success enough to repay its cost and the delay it imposes. Act on differences between execution plans, not on absolute valuations or the choice of engine. What an asset is worth at this stage cannot be established with certainty, but what should be done next can. ...
Most oncology programmes fail pre-clinically, so experiment order determines how much is spent before termination. Yet pre-clinical development is conventionally modelled as one gate with one probability of success, which cannot order experiments. Question-based clinical development (QBCD) instead splits stages into the questions a programme must answer. This thesis applies QBCD to the pre-clinical phase and separates two sources of value: the gate premium from finer staging and the engine premium from a binomial real-options lattice. To our knowledge, no published application separates these effects.
One pre-clinical immuno-oncology asset was valued in a 2×2 factorial design crossing two engines, a risk-adjusted decision tree and a binomial lattice, with two paths, conventional phases and six company-elicited question gates. Both paths shared total cost, duration and probability of success, and both engines the same discount rate. Question order and concurrency were searched exhaustively under precedence constraints, with sensitivity, Monte Carlo and four-indication robustness checks.
Decomposing the pre-clinical block added €4.15 M (17.1 % of the phase-based €24.25 M), whereas the lattice added €0.13 M (0.5 %), a thirty-two-fold difference. Continuing was worth about four times the gate cost even at the tightest gate, so abandonment was rarely optimal and a lattice earns its place mainly on marginal assets. Discounting the lattice at a lower rate than the tree, as published comparisons often do, inflated the engine premium 500-fold.
The optimal ordering was reproduced by sorting questions on ascending cost-to-failure ratio, a 1960 rule this thesis extends to a discounted form, so finding the best sequence is a sort, not a search. Re-ordering the company plan was worth €0.79 M and running questions concurrently added €2.05 M, together 10.0 % of the original plan’s value. Of the 28.8 % increase an rNPV appraisal would report, only that 10.0 % is under the team’s own control. Across 10 000 Monte Carlo scenarios, the question-based tree was worth starting in 97.5 % of cases and the phase-based tree in 76.3 %, and the discounted rule matched the exhaustive optimum in every scenario.
Three decisions follow. Order experiments by ascending cost-to-failure ratio and run them concurrently where precedence allows. Admit an extra experiment only if it lifts a later gate’s probability of success enough to repay its cost and the delay it imposes. Act on differences between execution plans, not on absolute valuations or the choice of engine. What an asset is worth at this stage cannot be established with certainty, but what should be done next can.
One pre-clinical immuno-oncology asset was valued in a 2×2 factorial design crossing two engines, a risk-adjusted decision tree and a binomial lattice, with two paths, conventional phases and six company-elicited question gates. Both paths shared total cost, duration and probability of success, and both engines the same discount rate. Question order and concurrency were searched exhaustively under precedence constraints, with sensitivity, Monte Carlo and four-indication robustness checks.
Decomposing the pre-clinical block added €4.15 M (17.1 % of the phase-based €24.25 M), whereas the lattice added €0.13 M (0.5 %), a thirty-two-fold difference. Continuing was worth about four times the gate cost even at the tightest gate, so abandonment was rarely optimal and a lattice earns its place mainly on marginal assets. Discounting the lattice at a lower rate than the tree, as published comparisons often do, inflated the engine premium 500-fold.
The optimal ordering was reproduced by sorting questions on ascending cost-to-failure ratio, a 1960 rule this thesis extends to a discounted form, so finding the best sequence is a sort, not a search. Re-ordering the company plan was worth €0.79 M and running questions concurrently added €2.05 M, together 10.0 % of the original plan’s value. Of the 28.8 % increase an rNPV appraisal would report, only that 10.0 % is under the team’s own control. Across 10 000 Monte Carlo scenarios, the question-based tree was worth starting in 97.5 % of cases and the phase-based tree in 76.3 %, and the discounted rule matched the exhaustive optimum in every scenario.
Three decisions follow. Order experiments by ascending cost-to-failure ratio and run them concurrently where precedence allows. Admit an extra experiment only if it lifts a later gate’s probability of success enough to repay its cost and the delay it imposes. Act on differences between execution plans, not on absolute valuations or the choice of engine. What an asset is worth at this stage cannot be established with certainty, but what should be done next can.