On Fri, May 5, 2023 at 9:16 AM Elaine Meng <meng@cgl.ucsf.edu> wrote:

Hi Roger,
We do not have any control over what machines Colab uses to run the calculation, nor would we know why this changed. You would have to ask the Colab people.

As for your other question, AlphaFold does not "detect" protein interactions. The user decides what sequences to input. If the user gives multiple sequences, then AlphaFold tries to build a complex, but it does not mean that the complex exists in real life. Instead AlphaFold will just try to find the most reasonable structure of the complex because the user told it to build a complex.

I hope this helps,
Elaine
-----
Elaine C. Meng, Ph.D.
UCSF Chimera(X) team
Department of Pharmaceutical Chemistry
University of California, San Francisco

> On May 5, 2023, at 12:01 AM, Tom Goddard via ChimeraX-users <chimerax-users@cgl.ucsf.edu> wrote:
>
> Hi Roger,
>
> Please send your questions to the ChimeraX mailing list. I do not understand either of your questions (below), and am on vacation.
>
> Tom
>
>> On May 5, 2023, at 5:29 AM, Roger Leng <rleng@ualberta.ca> wrote:
>>
>>
>> Good afternoon Tom,
>>
>> I did not use RoseTTAFold (Ubuntu). I came back to use ChimeraX linked to Colab. But, I just found that Colab uses A100, or V100, or T4 (was "standard" or "premium"). I used Pro+ with an additional 500 units. I selected the A100 with high RAM. Questions: Why did Colab change this? What is the difference?
>>
>> Another question (I watched a youtube made by you): If you detected two protein interactions, how can we find the "exact" sequences for the interaction? Any suggestion? or any method to be suggested? What I did was use "the mouse key" to find the sequences required for the interaction.
>>
>> Thank you very much for your time and help. I look forward to hearing your reply.
>> Have a nice day!
>> Roger
>>
>> On Wed, Apr 12, 2023 at 7:20 PM Tom Goddard <goddard@sonic.net> wrote:
>> I don't know how multiple sequences are specified in ColabFold. The web page must say. Only ChimeraX separates them by commas I think.
>>
>> Tom
>>
>>
>>> On Apr 12, 2023, at 6:14 PM, Roger Leng <rleng@ualberta.ca> wrote:
>>>
>>> HI Tom,
>>> I input three protein sequences. I add common (,) for each sequence ( seq-1, seq-2, seq-3). It is possible wrong doing.
>>> Thank you.
>>> Roger
>>>
>>> On Wed, Apr 12, 2023 at 5:26 PM Tom Goddard <goddard@sonic.net> wrote:
>>> Hi Roger,
>>>
>>> AlphaFold 2.3 is probably the best. ChimeraX is using AlphaFold 2.2, but the ColabFold notebook I mentioned is using 2.3 -- it is supposed to have improved multimer prediction networks, although it is probably a small improvement. If AlphaFold predicts with high confidence then probably everything else will give the same result. If it predicts with low confidence, then nothing other than experimental structure determination will validate it. If it is medium confidence (e.g. pLDDT 70) then it may be worth trying different variants of prediction software.
>>>
>>> Tom
>>>
>>>
>>>> On Apr 12, 2023, at 4:06 PM, Roger Leng <rleng@ualberta.ca> wrote:
>>>>
>>>> Many thanks, Tom.
>>>> One thing I knew was that I failed many times to use AlphaFold.ipynb (online). I am trying to confirm our experiments (bench) using two-AI based programs. We successfully run AlphaFold with ChimeraX. I am searing for other programs to confirm our results. Excepting Alphafold (in ChimeraX), do any suggestions for two or three protein interactions?
>>>> Thank you again.
>>>> Roger
>>>>
>>>> On Wed, Apr 12, 2023 at 4:54 PM Tom Goddard <goddard@sonic.net> wrote:
>>>> Hi Roger,
>>>>
>>>> Unfortunately the Google Colab notebooks for AlphaFold and clones often have problems causing predictions to fail. There are many reasons for the failure, but all of them come down to poor maintenance of these free prediction services. Here are examples of why they have failed: Google Colab updates Python version (to 3.7 to 3.8 to 3.9) without advance warning, libraries being used update (most often jax gpu calculation) and break the code. Often it takes a week for Google or other Colab notebooks to fix the problem. I try to fix the breakages within a day or two for ChimeraX. I think the ColabFold notebook (https://colab.research.google.com/github/sokrypton/ColabFold/blob/main/AlphaFold2.ipynb) is one of the more reliable ones.
>>>>
>>>> It is not too surprising that these free protein structure prediction services fail frequently. It will be interesting to see what happens when one day Google decides that free Colab is too expensive to maintain and pulls the plug on it.
>>>>
>>>> Another problem with all the AlphaFold-like notebooks is that Google Colab has old GPUs, typically 12 or 16 GB of memory, and this limits the size of the predicted structures. Even high-end desktop graphics like an Nvidia RTX 3090 has 24 GB and handles larger structures, while graphics intended for machine learning often has 40 GB, 48 GB, or 80 GB. Here at UCSF people who often use predictions run them on our UCSF machines. Of course lots of work goes into maintain our software and hardware.
>>>>
>>>> Tom
>>>>
>>>>> On Apr 12, 2023, at 3:25 PM, Roger Leng via ChimeraX-users <chimerax-users@cgl.ucsf.edu> wrote:
>>>>>
>>>>> Many thanks, Tom.
>>>>> I watched your lots of "youtube". It is great to learn and use ChimeraX. I tried unifold.ipynb (online), but, un-luck, failed several times with paid program. Interestingly, I failed several times to use alphafold2.ipynb (online); but I used ChimeraX with Alphafold successfully for complex prediction.
>>>>> In addition, I was told that RoseTTAFold (Linux) could use for complex prediction. I am installing RoseTTAFold in Linux. Again, I filed several times to use RoseTTAFold.ipynb (online). Actually, I do not understand why I always filed to use AlphaFold or RoseTTAFold online (.ipynb with the paid program, Pro+).
>>>>> Thank you again.
>>>>> Have a great day!
>>>>> Roger
>>>>> Roger Leng,
>>>>> Faculty of Medicine,
>>>>> University of Alberta,
>>>>> Canada
>>>>>
>>>>> On Wed, Apr 12, 2023 at 4:05 PM Tom Goddard <goddard@sonic.net> wrote:
>>>>> Hi Roger,
>>>>>
>>>>> Unifold looks interesting.
>>>>>
>>>>> https://github.com/dptech-corp/Uni-Fold
>>>>>
>>>>> It is an open source reimplementation of AlphaFold using the PyTorch machine learning framework done by a China-based company called DPTech (https://www.dptech.com/). From reading the github page it sounds like one of its main advantages over AlphaFold is that the training code and protocol is all open source. Code for the training of AlphaFold was never made available as far as I know.
>>>>>
>>>>> That is all nice, but what are the best reasons for our UCSF lab to try to offer this in ChimeraX? The ChimeraX AlphaFold user interface took about 2 months of development with ongoing maintenance. I'd estimate the cost so far at $30,000. ChimeraX is funded by NIH grants. Is it worth expending similar resources to implement and maintain an interface to UniFold?
>>>>>
>>>>> Tom
>>>>>
>>>>>
>>>>>> On Apr 12, 2023, at 2:24 PM, Roger Leng via ChimeraX-users <chimerax-users@cgl.ucsf.edu> wrote:
>>>>>>
>>>>>> Dear Administrators,
>>>>>>
>>>>>> Is it possible to add "unifold" in your prediction, just like Alphafold?
>>>>>>
>>>>>> Thank you.
>>>>>>
>>>>>> Sincerely,
>>>>>>
>>>>>> Roger Leng
>>>>>> _______________________________________________
>>>>>> ChimeraX-users mailing list
>>>>>> ChimeraX-users@cgl.ucsf.edu
>>>>>> Manage subscription:
>>>>>> https://www.rbvi.ucsf.edu/mailman/listinfo/chimerax-users
>>>>>
>>>>> _______________________________________________
>>>>> ChimeraX-users mailing list
>>>>> ChimeraX-users@cgl.ucsf.edu
>>>>> Manage subscription:
>>>>> https://www.rbvi.ucsf.edu/mailman/listinfo/chimerax-users
>>>>
>>>
>>
> _______________________________________________
> ChimeraX-users mailing list
> ChimeraX-users@cgl.ucsf.edu
> Manage subscription:
> https://www.rbvi.ucsf.edu/mailman/listinfo/chimerax-users