On a scale of 1 to 5 (1=No confidence; 3=equivocal, 5=high confidence), what is your confidence that AI- and ML-based prediction models can accurately predict the potential toxicity of new compounds or relate chemical structure or activity to risk? Grouped By What is your level of experience with conducting or interpreting results from AI- and ML-based toxicity prediction models?

It depends on the quality of the database used for training.

QSARs are already being employed in toxicology so I think of AI and ML as potentially better mouse traps, not an entirely new undertaking.

I believe that AI and ML-based prediction models will in the future predict the potential toxicity of new compounds and or relate chemical structure or activity to risk. It will likely help to adjust toxicity and risk assessments of existing compounds WHEN the collected data is properly defined and used. I am expecting that different testing principles need to be defined that will be similar successful as DeepMind AlphaGo self-playing success. I do not think that the current approaches will be successful yet.

While I think the accuracy of such predictive approaches is high, it will never be perfect, and is unlikely to be used by itself for regulatory decision-making. However, it can be a useful component of risk assessment, e.g. as part of weight of evidence approaches, and can be very useful for prioritizing chemicals for further testing.

Accurately predict?? Not really, we do not know enough as yet about toxins and their physiological effects on different species. In the end, animal testing is sadly still the golden rule. BUT AI could reduce the number of animals used through accurate prediction of the likelihood of outcomes of toxicity tests.

Model output is only as good as the input variables

Acutually, it depends on what data and models were applied.

Predictions can be as good as the data they are based on: In several instances the database is very limited...

I feel that AI amd ML could be game changers in the future but their output totally depends upon the quality and depth of input data.

Of course it depends on the clarity of the MoA, however in general we should understand how the molecule works therefore pathway related information and data on labels of similar products should provide sufficient evidence to support appropriate predictions.

Very confident

ML/AI offer unique clustering methodology that can actually risk validate the risk percentage from each technique/algorithm used.

Only with careful expert supervising from early beginning to results valuation

Room can be made for occasional misses with AI and ML

It depends upon what data are available for AI and ML. For example for mutagenicity, (Q)SAR approaches show very good concordance with results. However, predicting other toxicities could be challenging.

I think they can approximate the validity of lab tests (which can of course vary too).

It depends on the chemical class and whether it is a mixture etc. For some things these approaches can be applied and predict reliably but for others not. They must be applied on a case-by-case basis in the context of a well defined problem

Some can be used some not. Still needs expert judgement though this technology should continue to evolve and be evaluated like qsars

Lots of work on NAMs are showing that it is possible to predict toxicity, so as long as a computer can learn the relationships, I dont see a problem. i am less certain about risk because you would also need to factor in exposure.

there will be examples where it works effectively and examples where it does not.

We use different software for the assessment of toxokinetic parametets of newly designed molecules

This is a two-pronged question with different streams that should not be written as such.
My confidence that AI- and ML-based prediction models will be able to accurately predict the potential toxicity of new compounds at the moment is low: see answer above.
My confidence that AI- and ML-based prediction models can accurately predict the potential toxicity of new compounds based on chemical structure or activity is higher as this is being done now to develop new drugs.
So my overall answer will only relate to the first part.

Simulations way more optiibs

For many compounds, prediction is possible. Exceptions do occur and care is needed.

I have full/ high confidence that AI- and ML-based prediction models can accurately predict the potential toxicity of new compounds or relate chemical structure or activity to risk because they can accommodate "Real World scenario".

see my comment for #1

I do not have adequate experience to rate this in a valid manner.

The relationship between chemical structure and biological activity is well established. Still, predictions need constant experimental confirmation.

Depends if it is a read accross with adequate data it may be possible but in such a case it can be done manually

QSAR models have been in use for decades now and their predictivity even for "limited" endpoints like genotoxicity or irritation is rather limited. Complex endpoints like systemic toxicity after repeated dosing still cannot be modelled. Thus, I have medium confidence an improved modelling of simple endpoints and low to no confidence of AI or ML being of any suseful assistance in complex endpoints like systemic toxicity, immunotox, reprotox etc.

See 1 above

For the above reasons I am concerned that AI or ML machine could increase safety or accuracy of risk assessment of chemical components.

not clear where AI and ML can add value. It may not be in the final determination of "always toxic at dose x" as we tend to simplify with our piecemeal decision approaches now. Could be more about predicting population and individual risk in decisions of use scenarios and risk management alternatives, considering exposure and confounding and co-morbidity and aggregate interacting exposures, for example, that we are so bad at doing now.

SAR and sometimes QSAR have become highly refined for new chemistry tox assessment especially when common metabolic pathways comparing surrogates and target molecules are available.

This is the union of the interactions of chemistry with biology. The accuracy of the AI/ML predictions will only be as good at the current knowledge of both fields and their interations.

Le correlazioni fra struttura e tossicità in varie forme fino ad avere veri e propri algoritmi interpretativi è uno dei successi della R.S.più recente

I think that is to be found out, but as said, the potential for accurate predictions is definitively there.

Undoubtedly. They are just math. If there is causation, math can model it to any precision we find important. The bottleneck is not the AI/ML but the precision of our experiments.

If some QSAR has verified that AI/ML can mimic animal results, then they may have a role in predicting toxicity.

There will be some early glitches but it should improve over time.

AI results should at a higher TIer should be vetted by a human expert.

Humans need to work collaboratively across disciplines and sectors to achieve success.

Models made with the earlier proven biological system can provide some guidance

My results from Leadscope and In Vitro MultiFlow were incorrect when further analyses were performed.

Initial hypotheses may be generated but require further testing to validate

Prediction models are the best tools to predict any disease's evolution or potential physiopathological dysregulation. However, in real-world, the things are more complex; so models do not necessarily well fit with data

No one can accurately predict the potential toxicity of new compounds or relate chemical structure or activity to risk, neither experts nor AI. We can only try to predict it with a rather uncertain degree of confidence.

nearly impossible I would say as the entire literature suffers from too high and often way unrealistic exposures

it also depends on the toxicity endpoints, some toxicity endpoints, such as skin sensitization, irritation are more mature than others, because they have large amount of both animal/human and in vitro data. Others, such as developmental and reproductive endpoints, are less mature and because of multiple pathways that are involved.

For the moment these prediction are far away from routine application.

ML/AI in toxicology is still in its infancy, it's ability to consistently and accurately predict novel hazards has yet to be seen.