David Fay · Transcript · · Updated
David's AlphaFold WorkShop–Tutorial 2026 – Transcript
Video by David Fay · Watch the original on YouTube
Summary
AI-generated from the transcript. Verify details against the source video.
- Running 4 to 8 seeds can help catch protein interactions that might otherwise be missed by using only a single seed.
- Metrics such as ipSAE, actifpTM, and iLIS perform better than ipTM with a greater area under the curve.
- Metrics can be calculated through a web browser using the LIVIA site, though it is a low throughput process.
Use this video with Pro
Reading, copying and transcript downloads are free. Create saved AI results for this video with a signed-in Pro account ($9/month).
Transcript
Search ignores accents. All transcript text stays visible.
If your subtitles are on, GREAT! leave them on as I have inserted some comments and corrections. I also have a video on proximity labeling: https://www.youtube.com/watch?v=0T9PdvilDnM
Increasing the seed number to 4-8 can help you catch interactions that you might otherwise miss by running only one seed.
Note that PAE co-coloring is based on a ChimeraX algorithm, which imposes a chosen cutoff. But the cutoff isn't a 'God-given' physical constant, like the speed of light. It's useful and suggestive but taken too literally can give false positives and negatives—just like everything else.
If you truly understand how PAE works, then the presence of the thin diagonal line from the top left to the bottom right should make sense.
I should emphasize that using a different"starting" seed can dramatically change the models you get!
If your first seed gives you a negative result, but you have reason to believe there may be an interaction, try running 4-to-8 seeds.
More seeds will reduce missing out on a true positive... But will potentially increase the number of false positives...
Lastly, seeds are not fully deterministic – Running the same seed twice can give you slightly different results
Lower PAE is when i,j goes from a more flexible protein (domain) to a more rigid protein (domain).
The terminal domain in B (iii) is in fact green. There is nothing wrong with your vision or device!
Meant to say upper right and lower left... but you probably knew that.
I mean to say a lot of the true positives. Which you want to keep!
Note: Ah-Ram Kim has made it much easier to determine LIS, iLIS, and other metrics via a web page. (Thanks!) See video comments.
Please see Roland Dunbrack's clarification about actifpTM in comments Note: AF can place residues very close to each other just by chance – even when they have bad (high) PAEs
ROC (Receiver Operating Characteristic) curves are used for clinical diagnostics.
ipSAE, actifpTM, and iLIS all appear to perform better than ipTM (greater area under curve). Note that the newer metrics make a bigger difference when the degree of protein disorder (lower pLDDT - right plot) is higher.
Shame on me for blaming my wife! I retract that statement entirely! I had an actual cowbell somewhere in my office but failed to access it.
These metrics can be calculated via a web browser using the LIVIA site: https://flyark.github.io/LIVIA/universal.html Only downside is that it's a low throughput process.