ADMETlab 3.0 limits & performance

This page helps you size batches and set throughput expectations for ADMETlab 3.0 runs. Batch size, whether you request the core or the full 119-endpoint bundle, and when you schedule heavy runs matter most for turnaround.

Where ADMETlab 3.0 sits

ADMETlab 3.0 is the deeper-profiling counterpart to ADMET-AI: its multi-task DMPNN predicts 119 endpoints (31 more than v2.0) with uncertainty estimates, which is decision-ready for lead optimization but heavier per compound than a fast triage pass. Use ADMET-AI to screen a large library, then bring the survivors here for the wider panel — that keeps the expensive comprehensive run on a shortlist.

Batch sizing

The multi-task DMPNN computes all requested endpoints in a single pass, so batch sizing — not endpoint count — is the main throughput lever.

• Recommended batch size: 50–100 compounds per submission for a good balance of throughput and reliability.
• Maximum batch size: varies by workspace tier; check your workspace limits. Large batches may see longer queue times.
• Start with defaults; downsize batches if retries rise or queue times spike.
• Split heterogeneous libraries into consistent-size batches for steadier throughput.
• Oversize uploads are rejected before queuing; keep under the workspace file-size and molecule-count limits.

Endpoint bundle drives cost

• Core endpoints complete faster than the full 119-endpoint extended panel.
• Request only the categories your decision depends on (for example, hepatotoxicity and hERG for a safety screen) rather than the full panel by default.
• The uncertainty estimates ADMETlab 3.0 returns let you focus review on low-confidence endpoints instead of re-running the whole panel.

Throughput expectations

• Processing time depends on batch size and endpoint-bundle selection.
• Multiple medium jobs (50–100 compounds) often finish sooner than a single very large one.
• Queue depth fluctuates; schedule heavy runs during lower-traffic windows.
• Track retry counts in the logs — persistent retries point to input or parameter issues, not capacity.

Quotas and rate limits

• Honor polling backoff headers to avoid throttling.
• Associate jobs with projects for clearer quota attribution and auditing.
• Prefer steady-state scheduling over bursty submissions for long-running campaigns.

Repro and stability

• Log model_version, the endpoint bundle, batch size, and preprocessing flags for every run.
• Maintain a fixed regression subset to detect shifts when tuning parameters.

Read next