Computational Safety Report: Guide to Reading and Interpreting

This guide explains how to read and interpret the metrics presented in the Phase 5 Computational Safety report. The reports are generated by deterministic, template‑driven modules; they provide computational measurements, not clinical predictions or investment recommendations. Use this guide to understand each section, the meaning of each metric and how the different pieces fit together.

Computational Safety Report

This report provides detailed outputs behind the fast‑fail summary. It combines ADMET, toxicity, metabolism, exposure, off‑target predictions, synthesis feasibility and biological risk. The sections appear in the order below; metrics labeled (prob.) are model‑predicted probabilities and do not represent observed clinical incidence.

1. Failure Mode Ontology – First‑Failure Ranking

Under Failure Mode Ontology System Output, bars rank potential failure modes by absolute probability. This ranking estimates which failure mode is most likely to break the program first under the current evidence. Each row lists:

Mode – failure mode name (e.g., Exposure Collapse, Reactive Metabolism).
P(first failure) – absolute probability (0–100 %) that this mode occurs first.
Weight – contribution weight for the mode; weights sum to 1 across all modes.

High P(first failure) values and higher weights indicate modes that currently drive near‑term failure pressure. Compare these with the failure mode cards in the risk summary.

2. ADMET Predictions

The ADMET Predictions section lists predicted ADME‑Tox metrics. Key items:

BBB Perm. (prob.), CYP2D6 (prob.), CYP3A4 Inhibition (prob.), Substrate Prob. – model‑predicted probabilities for blood‑brain barrier permeability and metabolic liabilities.
Solubility – points to the solubility assessment section (see below); ADMET predictions do not duplicate solubility.
Clearance (pred.) – predicted clearance in model units.
Physchem (MW / cLogP / TPSA) – molecular weight, calculated logP and topological polar surface area.

These values help assess drug‑likeness and potential ADME risks. Higher probabilities may suggest a need for further assays.

3. Toxicity Profile

Probabilistic predictions for various toxicities are summarized. Each value is a model score rendered as a percentage (0–100 %). Major items include:

hERG liability, hepatotoxicity, mutagenicity, cytotoxicity, nephrotoxicity, neurotoxicity, carcinogenicity, skin sensitization.
Non‑hERG cardiotox and resp. toxicity may be blank if not predicted.

These probabilities do not indicate incidence rates; they provide relative risk signals for prioritizing in‑vitro validation.

4. Mechanistic Confidence Matrix

This matrix summarises mechanistic attributes and whether assay gates should be opened:

Component – the property being assessed (BBB permeability, solubility, CYP drug–drug interaction risk, cardiac assay gate readiness, DILI risk).
Result/Prior – classification under current evidence (e.g., Medium, Low, None, Unknown).
Verdict – recommended level of concern (e.g., Medium or Low).
Assay Gate – indicates whether to run a physical assay (e.g., hERG patch clamp).

Use this matrix to decide if in‑vitro confirmation assays are warranted.

5. Off‑Target Screening

Off‑target predictions identify potential off‑target interactions using similarity‑enrichment analysis (SEA‑lite). The report lists risks for proarrhythmia, 5‑HT2B and MAO‑B; “NONE” indicates no significant similarity. Additional details include the number of reference compounds and maximum similarity scores. If any off‑target risk appears, plan targeted receptor binding assays.

6. PBPK Sensitivity

Physiologically based pharmacokinetic (PBPK) outputs estimate exposure metrics:

Cmax and AUC – model‑reported exposure at maximum concentration and area under the curve.
Tmax and Clearance – may be absent if not emitted.
Status/Method – indicates whether uncertainty quantification (UQ) was available; e.g., SUCCESS using a split_conformal method with a nominal 90 % confidence level.
Prediction interval width – width of the conformal prediction interval; wider intervals indicate greater uncertainty.

This section informs how certain the PBPK estimates are and whether additional pharmacokinetic studies are needed.

7. Metabolite Prediction

The Metabolite Prediction section lists the number of predicted metabolites and reactive metabolites and flags elevated risk if many reactive species are expected. “Transformations” lists typical biotransformations predicted. Use this to anticipate metabolism complexity and design metabolite identification studies.

8. Solubility Assessment

This section is authoritative for solubility:

Class (e.g., INSOLUBLE) and numeric solubility in mg/mL.
LogS and the method used (ADMET_V3_SOLUBILITY_AQSOLDB_PRIMARY).
A comparison between the primary estimate and diagnostic estimates (AQSOLDB / ESOL / Δ) for context.
If conformal quantification was not run, CP status is NOT RUN.

Solubility classification helps interpret the exposure collapse risk card in the fast‑fail summary.

9. Manufacturability & Synthesis Feasibility

Key metrics:

SA Score – synthetic accessibility (lower scores indicate easier synthesis). A score of 1.82 with 6 strategic fragments suggests low complexity.
Synthesis Status – whether the synthetic route found by RDKit+TemplateSearch was successful.
Route found / Blocker signal – indicates if a route exists and whether any synthesis blockers were detected.

These outputs provide a first pass on how feasible synthesis might be; however, they rely on limited templates and are not definitive.

10. Graph & Drug‑Likeness

This section reports graph‑based drug‑likeness metrics (e.g., kinase drug‑likeness score). A high value (0.99) suggests the compound’s features align with known kinase inhibitors. Use this as supporting evidence for target relevance.

11. Virtual Panels

Virtual screening panels evaluate panel‑specific risks (e.g., DILI risk, CYP time‑dependent inhibition). The section lists whether hits were returned and summarises panel coverage. “Unknown” or “no rows returned” indicates no decisive data.

12. Exploratory Optimization & Causal Signals

Brief outputs from exploratory optimization (Monte‑Carlo Tree Search) and causal discovery are provided (e.g., QED score, number of nodes/edges in the causal graph). Lack of improvement in MCTS indicates no better analog found by the computational heuristic; causal discovery highlights relationships among variables that might guide optimization.

13. Target Biology Risk Details

This section delves into biological risk drivers. It reports the overall band (Moderate, score 0.410) and confidence (0.738), then breaks down submodules:

Submodule	Band (score)	Risk	Confidence	Key points
Genetic Intolerance	Favorable (0.125)	12 %	85 %	LOEUF, pLI and mis_z suggest tolerable genetic loss‑of‑function.
Knockout Phenotype	Elevated (0.680)	68 %	80 %	38 knockout phenotypes across five organ systems; lethal at postnatal stage.
Substrate Pleiotropy	Elevated (0.650)	65 %	70 %	Many GO/Reactome processes; potential on‑target toxicity from pleiotropic roles.
Expression Window	High (0.700)	70 %	70 %	Tissue‑specific expression; limited therapeutic window between disease tissue and safety tissue.
Direction Calibration	Favorable (0.000)	0 %	70 %	Pathogenic variants consistent with inhibitory mechanism.

The ranked failure modes summarise the biological failure pressures: Target knockout is lethal (Critical), No expression‑based therapeutic window (High) and Target functional pleiotropy (Elevated). Each card provides severity, confidence and recommended gate actions (e.g., heterozygous knockout review, tissue‑restricted delivery strategies). These inform whether to proceed or design safer modalities.

14. Clinical Precedent

Lists historical clinical programs with similar targets/indications. Each entry shows the program, phase, target/indication and a similarity value. Similarity is not a probability; it contextualises the project against previous trials. Use this to gauge how well explored the target is clinically.

15. Evidence References

A summary of unique evidence IDs used in risk signals and priority findings is provided. Each evidence ID links back to data sources. Reviewing these references can help verify computational outputs.

16. Other Actions & Interpretation Notes

The Other Actions table suggests experimental or computational steps (e.g., run PAMPA/MDCK efflux studies, perform CYP inhibition panel, reactive metabolite trapping) that could strengthen evidence. The Interpretation Notes remind users that model probabilities are not clinical incidence rates and that some metrics (e.g., clearance, half‑life) may be in model units. This section emphasises appropriate use of the data; absence of elevated signals is not proof of safety.

Using the Guide

Start with the CB‑TRI overview to understand the overall risk profile and identify which channels contribute most to risk. Use the semantics legend to interpret scores and bands correctly.
Dive into the projected failure mode cards for actionable insights. Prioritise modes with higher P(first failure) and severity and review recommended experiments.
Consult the computational safety report for detailed metrics behind each risk channel. Examine ADMET predictions, toxicity profiles, metabolism, solubility and manufacturability to validate or challenge the risk summary.
Pay attention to uncertainty and evidence strength. Sections indicating Unknown or Skipped or low evidence strength point to data gaps; design studies to fill them.
Cross‑reference evidence IDs when you need to trace the origin of a metric or verify the underlying data.
Treat predictions as exploratory. These metrics provide hypotheses for early‑stage drug discovery; they are not clinical recommendations.

By following this structured approach, you can systematically evaluate computational risk outputs, identify key liabilities and design evidence‑strengthening experiments.