Most projects assume the design-make-test-analyse loop is limited by chemistry. Measure it and the picture usually changes: the chemistry runs in days, and the cycle takes six weeks because of queues between the steps, assay data that arrives in a form nobody can model, and an analyse step that happens whenever someone finds an afternoon. The time to a decision, not the time to a compound, is what a cycle costs, and the three largest losses are almost never in the fume hood.
Measure the cycle before you optimise it
Timestamp four moments for the last ten iterations: design frozen, synthesis complete, result released, decision taken. The four intervals tell you where the weeks are. Two patterns show up repeatedly:
- Waiting dominates working. The sum of the active steps is a fraction of the elapsed time. Everything else is queueing for an instrument, a plate, a reviewer or a meeting.
- The last interval is the longest. Results exist for a week or two before anyone converts them into the next design, because no one owns that conversion and no slot is reserved for it.
Neither is fixed by working faster.
Loss one: the analyse step has no owner and no slot
The analyse step is the only one in the loop with no equipment, which is exactly why it drifts. A result that lands on a Tuesday and is discussed at the next project meeting in eleven days has added eleven days to the cycle while nothing happened.
The fix is scheduling, not tooling: a fixed, short slot in the week where the available results are read and the next design is frozen, with a named person responsible. Projects that do this routinely take a week or more out of every cycle without changing anything else.
Loss two: the data arrives unusable
A measurement that has to be re-keyed, re-mapped to structures, or cleaned of layout artefacts before it can be read costs more than the delay it creates: it also costs the trust in any analysis built on it. Three recurring forms:
- Identity ambiguity. The result sheet carries an internal code, the design carries a structure, and the link between them is a mapping file someone maintains by hand.
- Values without context. A number with no unit, no limit of quantitation and no replicate information, so a flat "no activity" cannot be told apart from "below the detection limit".
- Censored data silently converted. "Above 100 µM" written as 100 turns a non-result into a measurement and biases every model fitted to it afterwards.
Fixing the export once, so that structure identifier, value, unit, limit and replicate flag arrive in machine-readable form, usually removes several days per cycle and makes the previous year of data usable as well.
Loss three: batches sized for convenience, not information
Two failure modes, opposite in direction. Twenty compounds per round because the plate holds twenty, when five of them were designed to answer the question and fifteen are filler. Or one compound per round, when the question needs a contrast and a single result cannot provide it.
The informative batch is the smallest set whose results can distinguish the hypotheses on the table. That is usually three to eight compounds in lead optimisation, chosen to span the variable in question, including deliberate extremes. A worked example: if the hypothesis is that a substituent at one position controls metabolic stability, a batch with four substituents spanning the electronic range answers it, and twenty analogues around the best current compound do not.
Loss four: parallel steps run in series
The measurable part of the loop usually has slack nobody uses. Analytical method development can start while synthesis runs. The next design round can be prepared on predicted values, ready to be confirmed or discarded when the measurements land. A decision rule agreed before the data arrive ("if the stability of all four is below the threshold, we drop the position") converts the analyse step from a discussion into a lookup.
The counterpoint is honest: preparing designs on predictions means some preparation is wasted. That is a trade of cheap work against expensive waiting, and it usually pays.
Where prediction actually helps
Models shorten the loop in two specific places. They remove compounds that would have failed on a property the project already knows it needs, before anybody makes them. And they turn a one-shot campaign into a series of informed rounds, because a model updated on each round's measurements proposes the next round better than a fixed plan does. What they do not do is replace the measurement that decides, and a cycle that stops measuring stops learning.
Limits of the statement
- Cycle time is not the only objective. A faster loop around the wrong question produces wrong answers sooner, and the design step is where that is prevented.
- The intervals above are averages over projects. Yours may genuinely be limited by synthesis, for example in a multi-step route, and then the analysis says so and the effort goes there.
- Decision rules agreed in advance need an escape clause. A surprising result is a reason to think, not a reason to follow the rule.
- None of this reduces the number of experiments needed to establish a fact. It reduces the time spent not doing them.
Frequently asked questions
What usually limits a DMTA cycle?
Queueing and the analyse step, not the chemistry. Timestamping design frozen, synthesis complete, result released and decision taken over the last ten iterations shows where the elapsed time goes, and the waiting intervals are normally larger than the active ones.
How do you shorten the analyse step?
Give it an owner and a fixed weekly slot in which the available results are read and the next design is frozen. It is the only step with no equipment, so without a reserved slot it drifts to the next project meeting.
How large should a batch be?
Large enough that its results can distinguish the hypotheses under discussion, and no larger. In lead optimisation that is often three to eight compounds chosen to span the variable in question, including the extremes, rather than a plate-sized set of near-analogues.
What makes assay data unusable for modelling?
Ambiguous compound identity, missing units or quantitation limits, and censored values recorded as numbers. The third is the most damaging, because "above 100 µM" stored as 100 biases every model fitted afterwards.
Can predictions replace measurements in the loop?
No. They remove candidates that would fail on a known requirement and they inform which experiments to run next. The measurement is still what settles the decision, and a loop without measurement stops learning.