Are there any of the following applications you would have high confidence in AI- and ML-based toxicity prediction models being useful? (Please select all that apply and feel free to explain and include any citations you think relevant)

They can help to fill data gaps but not for extrapolation.
Quantitative risk assessment is possible strong area where AI, ML can contribute.

Not sure I would trust them for quantitative risk assessment unless samples were large enough. and the evidence was good.

As soon as some data (in vitro, QSARs, biomarkers, blood levels etc.) are available, AI and ML can be very helpful. Such cases mainly comprise Dose-response-modeling, benchmark dose identification, uncertainty analysis and kinetic modeling (including body burden considerations)

These may represent future applications

Currently the main impediment is the lack of a reliable gold standard to train our models. We are often required to use results from existing in vivo assays as if they were the ultimate gold standard to achieve, when they are very noisy and inconsistent, even when repeating the exact same assay with the same model organism.

Predicting results of specific studies

EPA's tox cast program was developed using such approaches as a means to screen and prioritize the US commercial chemical inventory to identify candidate agents that merit priority consideration and those that can wait. As part of this effort the approaches used hazard id/endpoint alerts, relative potency considerations and read across-like considerations to identify candidates. They have also been used in regulatory agencies and industry to fill data gaps such the reliable prediction of pchem properties for certain well characterized classes of chemicals where such have not been actually measured for certain members of that class

A human or better a small group of humans still needs to evaluate the results.

No confidence based on my experience.

Again "high confidence" is a high bar. In general parlance, this may be equivalent to proof beyond a reasonable doubt. If so, none at the moment.

If it is preponderance of the evidence, then this is a different level and then maybe initial hazard ID and initial screening.

AI & ML will certainly help in the screening of candidate agents to prioritize, to identify susceptive groups, risk prediction.

Data gaps require more detailed focussed experiments. Quantitative risk assessment requires a dose-response relationship which might be liner or nonlinear, and it is not always predictable.

As mentioned, the available models are of no use in toxicology for the time being. Of course, this is no impediment for some to use them anyway, but the old adage "garbage in - garbage out" still applies. The least damage would probably be done by using AI for prioritizing purposes or screening of candidate agents. However, better and somewhat more relevant models (in vitro screening assays, HTS) are already avaliable and well-established and I do not see any immediate need for half-baked computer-based models.

For QRA based on actual data

Too early to guess, and the options considered for AI or ML might also be best themselves addressed with AI. Training is necessary but constraint based on, for example, the 1950's bright line decision approaches we rely on, might waste value or create harm

What computational toxicology, cheminformatics, nanoinformatics, etc already does great is summarize patterns and point is to prominent effects. I recommend reading the literature about AI/ML has been used in the past 20 years in medicinal chemistry, e.g. the articles by Yvonne Martin (https://scholar.google.com/citations?hl=en&user=YKsl2jEAAAAJ).

Only non-critical data gaps.

Some level of quantitation is currently achievable for some endpoints allowing use for setting exposure limits for low-exposure compounds for example.

Selected seems less problematic. Prioritization will be the most difficult prediction model given the evolution of values.

statistical analyses.

There will be a need for interpretation and implementation.

Based

Useful but not completely reliable.

With the limited data - AI/ML is great in screening and prioritize chemicals for further testings.