Rephois MolTrace's reaction-optimization module — it uses Bayesian optimization to recommend the next experiment to run.
Bayesian acquisition over a live Gaussian-process surrogate proposes the conditions most worth trying next — under hard constraints from your spectroscopy evidence and regulatory framework. The loop is half-closed by design: Repho proposes, a chemist approves, and nothing executes on its own.
Why this exists
A typical process-chemistry campaign often runs dozens of reactions to find a robust optimum. Many of those reactions are wasted — they sample regions the team already knew were bad, or repeat conditions a previous campaign explored two years ago and forgot.
At the same time, regulatory burden is climbing. Every condition tried is a potential data point that needs to be traceable. Every impurity that crosses an ICH limit is a campaign-restart event. Spreadsheet-tracked reaction histories can't withstand inspection — and don't survive the analyst who wrote them.
Repho runs the acquisition function in your stead. It remembers every prior campaign, respects every active regulatory constraint, and proposes the experiment most likely to advance the Pareto frontier — with an auditable trail of why.
The optimization loop
Each round emits a typed Pydantic record. The surrogate state, the acquisition proposal, the measured outcome, and the human decision are all recipe-hash-linked so any prior campaign replays bit-identically.
Set the design space: input variables (solvent, temperature, equivalents, time, catalyst loading), discrete or continuous bounds, encoded categoricals. The same Pydantic model the backend's ConditionDomain consumes.
Emits
design_space · variable_bounds · categoricals
Hard limits from Regentry auto-load: ICH Q3C residual-solvent ceilings, Q3D PDE limits, M7 nitrosamine alerts. Plus safety + cost guardrails from your operational profile.
Emits
objective_profile · cost · safety · regulatory_priors
Bayesian acquisition over a Gaussian-process surrogate. Expected Improvement (single objective), upper confidence bound for exploration, and constraint-aware ranking sliced into batches. Multi-objective campaigns surface a deterministic Pareto front. Reproducible — same seed, same batch.
Emits
recommendation_batch · acquisition_score · expected_improvement
A chemist approves the proposed batch and runs it in the lab. Recipe + parameters pinned by recipe-hash; cross-modal evidence pulled in automatically when the batch finishes. Repho executes nothing itself.
Emits
run_id · recipe_hash · operator · timestamp
Spectroscopy evidence lands automatically. Yield, purity, selectivity, impurity profile, residual solvent — all per-condition, all hyperlinked to source spectra.
Emits
yield · purity · impurity_set · ee · spectra_uris
Surrogate refit with the new observation. Posterior mean + variance over the design space recompute. Constraint-aware: violated points get masked out before the next proposal.
Emits
posterior · constraint_mask · pareto_set
Continue, stop, or pivot. Stopping rules built in: hypervolume convergence, budget exhaustion, hyper-parameter saturation. Human reviewer signs every campaign decision into the audit ledger.
Emits
decision · rationale · reviewer · audit_event
At a glance
┌─► Define design space ──► Constrain (regulatory + safety + cost)
│ │
│ ▼
Decide (stop / pivot / continue) Propose (acquisition over GP surrogate)
▲ │
│ ▼
Update surrogate posterior ◄── Measure ◄── Run (manual or closed-loop)
with constraint mask (spectroscopy auto-attaches)Methods we ship
Switch regimes by argument — the design-space + objective + constraint contract stays identical across all four. The backend uses run_bayesian_optimization() with the same acquisition_score + expected_improvement fields.
Single-objective · model-based
Pareto-frontier · batch-aware
Uncertainty-driven · explore + exploit
Human-gated · propose, approve, run
Use cases shipped
Each is a typed pipeline you parameterise — not a bespoke build. The same acquisition engine drives every template; the differences are in design-space encoding and objective weighting.
Hit-finding across solvent / base / catalyst combinatorial space. Discrete-categorical Bayesian search with diversity-aware proposals.
Inputs
Categorical design space · screening budget
Outputs
Top-N hits · uncertainty bounds · suggested follow-up
Continuous-variable optimization over temperature, equivalents, residence time. Multi-objective when selectivity matters as much as yield.
Inputs
Continuous bounds · objective weights
Outputs
Pareto frontier · best-by-objective conditions
Regulatory-Hub impurity limits become hard constraints. Optimizer proposes only conditions predicted to clear ICH Q3A/Q3B thresholds.
Inputs
Impurity limits from regulatory · prior runs
Outputs
Compliant conditions · constraint-aware ranking
Cost surrogate co-trained alongside yield. Proposes the cheapest condition that still hits the objective — useful for late-stage process work.
Inputs
Reagent unit costs · time budget · operator load
Outputs
Cost-Pareto frontier · sensitivity report
Optimizes for robustness at scale: temperature insensitivity, mixing-time tolerance, work-up reproducibility. Variance-weighted objective.
Inputs
Lab-scale data · scale-up risk profile
Outputs
Robust conditions · sensitivity heatmap
HPLC method dev, work-up optimization, crystallisation polymorph search. Same engine, different objectives — purity + recovery + crystal habit.
Inputs
Design variables · purity targets
Outputs
Method parameters designed to support Q2(R2) validation
Anatomy of a campaign
The campaign below is the worked example threading through every module page — acetic-acid impurity suppression. Round 4 violates the Q3C constraint mid-flight; Round 5's acquisition mask routes around it.
bayesian_optimization_run · run_4f7a · 5 rounds · convergence reached
| Round | Acquisition | Proposed | Measured | Status |
|---|---|---|---|---|
| R01 | Latin Hypercube · seed | T 25 °C · 1.2 eq · DCM · 4 h | yield 41% · imp 3.2% · cost low | Explore |
| R02 | EI · α 0.84 | T 50 °C · 1.5 eq · DCM · 6 h | yield 67% · imp 1.8% · cost low | Exploit |
| R03 | Pareto-front batch · k=3 | T 65 °C · 1.5 eq · MeCN · 8 h | yield 74% · imp 1.1% · cost mid | Exploit |
| R04 | UCB · β 1.6 | T 80 °C · 2.0 eq · MeCN · 4 h | ABORT · imp 5.8% > Q3C cap | Constraint |
| R05 | EI · constrained | T 60 °C · 1.8 eq · MeCN · 6 h | yield 81% · imp 0.7% · cost mid | Best so far |
Yield trajectory · convergence after constraint mask
R01 ████████░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░ 41 % explore R02 █████████████░░░░░░░░░░░░░░░░░░░░░░░░░░░ 67 % exploit R03 ██████████████░░░░░░░░░░░░░░░░░░░░░░░░░░ 74 % exploit R04 ██████████░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░ ABORT Q3C limit exceeded R05 ████████████████░░░░░░░░░░░░░░░░░░░░░░░░ 81 % best · constraint-aware Pareto best 81 % yield · 0.7 % impurity (Q3C compliant) · mid cost Stop reason hypervolume Δ < 0.01 over last 2 rounds
The honest comparison
Trial-and-error campaigns aren't bad — they're under-instrumented. Repho keeps the chemist as the decision-maker but moves the experiment-selection step from gut to acquisition function.
| Dimension | Trial-and-error | Repho |
|---|---|---|
| Experiment selection | today Intuition + last-week's lab meeting + grad-school heuristics | with Repho Acquisition-function-driven · Expected Improvement over the live surrogate |
| Constraint handling | today Hope the operator remembers the impurity limit · catch in QC later | with Repho Hard constraint at proposal time · violated points masked before the optimizer ever sees them |
| Batch parallelism | today One reaction at a time · serialised round-trip on the bench | with Repho Batch proposal of N diverse runs · plate-aware suggestions |
| Replay 6 months later | today Lab notebook + memory · 'we tried that, didn't work, can't remember why' | with Repho Recipe-hash replay · reproducible from the pinned recipe-hash |
| Stopping decision | today Run until budget runs out · or until someone gets frustrated | with Repho Hypervolume convergence · explicit stopping criteria · auditable decision |
| Cross-module evidence | today Yield in a spreadsheet · impurity in a PDF · cost in an email | with Repho One typed record per round · spectra hyperlinked · cost + impurity + audit entry together |
Three pillars, one loop
Picking up the cross-module worked example from Spectroscopy + Regentry — here's the role Repho plays when the impurity threshold is crossed mid-campaign.
Round 4 shows acetic acid impurity at 2.10 ppm climbed to 5.8% — over the ICH Q3C Class 3 informational threshold for this drug substance.
Constraint flips from 'informational' to 'hard limit' for subsequent rounds. Repho receives the new constraint vector via the typed cross-module API.
Surrogate masks the violating region. Round 5 acquisition function only proposes conditions predicted to keep acetic acid below 2%. Result: 81% yield, 0.7% impurity.
Cross-module handoff captured as a single audit_event with full provenance. Inspector can replay every decision from FID hash to recipe update.
Trust & reproducibility
Every campaign is replayable from any prior date. Every proposal is acquisition-attributed. Every pivot is signed. The surrogate is a tool — the chemist is the decision-maker.
Seed-reproducible campaigns
Same design space + same seed + same observations are designed to yield the same proposals. Recipe-hash-linked surrogate state pinned per round.
Constraint-aware by design
Regulatory limits, safety thresholds, and cost ceilings enter the optimizer as hard constraints — not soft hints. Proposals are masked against known limits before they reach you.
Human signoff per pivot
AI proposes batches. Humans approve campaigns. Every pivot (stop, continue, escalate) is signed and recorded with reviewer + role + rationale.
Safety abort wired live
Closed-loop campaigns include live sensor streams. Out-of-bounds temperature / pressure / off-gas triggers immediate abort + ledger entry.
Spectra linked per round
Every measured outcome carries the SpectraCheck SHA-256 of the source spectrum + the regulatory verdict + the cost line — one row, full provenance.
Campaign replay from any prior date
Re-derive any historic Pareto front, any historic proposal, any historic decision from any prior date. Deterministically reproducible from the pinned recipe-hash and seed, same-environment.
Pick a reaction where you've already run 30+ conditions and can't tell what to try next. We'll show you the Pareto frontier Repho would propose — with your regulatory constraints already enforced.
FAQ
Repho runs Bayesian optimization over a Gaussian-process surrogate, using Expected Improvement and upper-confidence-bound acquisition, plus random-forest and TPE-style alternatives. Multi-objective campaigns produce a deterministic non-dominated Pareto front with a hypervolume indicator, computed in pure NumPy.
Yes. Repho computes Sheldon E-factor (simple and complete), Trost atom economy, process mass intensity (PMI), reaction mass efficiency, and a CHEM21-derived solvent green-score deterministically from RDKit and transparent arithmetic — no model produces the numbers. Each is selectable as an optimization objective alongside yield and selectivity.
Yes. Repho generates a deterministic 24-, 96-, or 384-well plate over a project's design space using Sobol or Latin-hypercube space-filling, full-factorial enumeration, or a Bayesian-optimization seed set, honoring fixed conditions and excluded combinations. Plates are reproducible per seed and export to CSV/JSON for lab robotics.
Impurity action items from a Regentry dossier (for example ICH Q3A/B or Q3C limits) are injected as reaction constraints, and recorded experiment outcomes are evaluated against those limits with provenance back to the source action item. A high or critical limit flags an experiment as exceeding the limit; lower tiers apply an advisory penalty. This is decision support, not batch disposition.
No — Repho is decision support and human-gated; nothing auto-executes. Its half-closed design-make-test-analyze loop only proposes the next batch as a draft cycle, which still requires qualified human sign-off and a cleared structural-safety gate before any execution. An optional advisor agent can plan and narrate but never computes a quantitative value.
Repho campaigns are designed to be reproducible from a pinned recipe-hash and seed in the same environment, and every proposal, measured outcome, and human decision is written to MolTrace's tamper-evident audit ledger. The controls are built to support 21 CFR Part 11 and GAMP 5 workflows; MolTrace does not claim the product is itself compliant or validated.