something aboyt run

[amarantostaver-cloud]

Hello, the work of PPIFlow is really amazing. If possible, I’d like to ask you a few questions:

1. Regarding the processing of PDB structures: Crystal structures inevitably have missing residues at certain amino acids, and it seems that the flow model tends to fill in gaps near hotspot residues when generating the backbone. How do you handle the processing of PDB structures?
2. In the Stage 1 part, which is mainly aimed at obtaining side chains that provide stable binding, what criteria do you use to determine if the side chains from Stage 1 are sufficiently stable? (In other words, should I increase my sequence number [recommended as 8], or should I increase the sampling temperature?)
   Thank you for reading this email, and I wish you a pleasant day.

[qyQilinYu]

Hi,

1. PPIFlow (the protein generator) and AF3Score (the protein evaluator) do not modify the input target protein. I think the error may come from keeping the residue gap when running the AF3 prediction. You must give AF3 a complete target sequence, without missing gap residues.

2. I don’t fully understand your question. You may want to increase the sampling temperature in stage 1 to get more diverse rotamers at the interface.

Best,
Qilin

[amarantostaver-cloud]

I am currently using PPIFLOW for antibody generation and have encountered some unexpected results during Stage 1. I would appreciate it if you could provide some guidance.

When testing the system with the built-in IL-13 control (sampling_temp: 0.5, samples_per_target: 100, num_seq_per_target: 8), around 20 sequences passed the initial filtering with ipTM > 0.2. However, when I ran the pipeline on my target antigen (sampling_temp: 0.5, samples_per_target: 25000, num_seq_per_target: 8), out of 200,000 generated sequences, only 3 managed to achieve an ipTM > 0.2.

Regarding this issue, I have performed several troubleshooting steps and have a few specific questions:

Pipeline Verification: Since the IL-13 control runs smoothly and yields expected results, I believe the local installation and program environment are functioning correctly.

Structure & Setup: I formatted my target antigen file to strictly match the layout of the IL-13 PDB file, but the low pass-rate persisted, suggesting it might not be a file formatting issue. I also checked the generated CDR loops and found no obvious abnormalities. Visualizing the antigen-antibody complexes showed proper spatial alignment, and my designated hotspot residues look correct. I am stuck as to whether I missed a processing step, or if such a low success rate is actually typical for certain targets.

Impact of Missing Structural Loops: My question regarding the structure stems from the fact that my experimental crystal structure has a small missing loop in the middle. For PPIFLOW, I used a structure modeled by AlphaFold 3 (AF3) instead. Does a missing loop in the original crystal structure (or the AF3-reconstructed version) significantly impact the efficiency of backbone generation and antibody screening in PPIFLOW? What is the recommended practice for handling such structural gaps?

Quantitative Metrics for Stage 1: Stage 1 aims to optimize and mature the side chains, but fixing too many side chains might not be ideal. Do you have any quantitative guidelines or empirical metrics for this? For instance, what is the optimal average number of fixed side chains, or what pass-rate at Stage 1 typically yields the highest downstream screening efficiency?

Thank you very much for your time and assistance. I look forward to hearing from you.

Best regards,

[amarantostaver-cloud]

I am currently using PPIFLOW for antibody generation and have encountered some unexpected results during Stage 1. I would appreciate it if you could provide some guidance. When testing the system with the built-in IL-13 control (sampling_temp: 0.5, samples_per_target: 100, num_seq_per_target: 8), around 20 sequences passed the initial filtering with ipTM > 0.2. However, when I ran the pipeline on my target antigen (sampling_temp: 0.5, samples_per_target: 25000, num_seq_per_target: 8), out of 200,000 generated sequences, only 3 managed to achieve an ipTM > 0.2. Regarding this issue, I have performed several troubleshooting steps and have a few specific questions: 1. *Pipeline Verification:* Since the IL-13 control runs smoothly and yields expected results, I believe the local installation and program environment are functioning correctly. 2. *Structure & Setup:* I formatted my target antigen file to strictly match the layout of the IL-13 PDB file, but the low pass-rate persisted, suggesting it might not be a file formatting issue. I also checked the generated CDR loops and found no obvious abnormalities. Visualizing the antigen-antibody complexes showed proper spatial alignment, and my designated hotspot residues look correct. I am stuck as to whether I missed a processing step, or if such a low success rate is actually typical for certain targets. 3. *Impact of Missing Structural Loops:* My question regarding the structure stems from the fact that my experimental crystal structure has a small missing loop in the middle. For PPIFLOW, I used a structure modeled by AlphaFold 3 (AF3) instead. Does a missing loop in the original crystal structure (or the AF3-reconstructed version) significantly impact the efficiency of backbone generation and antibody screening in PPIFLOW? What is the recommended practice for handling such structural gaps? 4. *Quantitative Metrics for Stage 1:* Stage 1 aims to optimize and mature the side chains, but fixing too many side chains might not be ideal. Do you have any quantitative guidelines or empirical metrics for this? For instance, what is the optimal average number of fixed side chains, or what pass-rate at Stage 1 typically yields the highest downstream screening efficiency? Thank you very much for your time and assistance. I look forward to hearing from you. Best regards, qyQilinYu ***@***.***> 于2026年6月26日周五 17:29写道：

…

*qyQilinYu* left a comment (Mingchenchen/PPIFlow#36) <#36 (comment)> Hi, 1. PPIFlow (the protein generator) and AF3Score (the protein evaluator) do not modify the input target protein. I think the error may come from keeping the residue gap when running the AF3 prediction. You must give AF3 a complete target sequence, without missing gap residues. 2. I don’t fully understand your question. You may want to increase the sampling temperature in stage 1 to get more diverse rotamers at the interface. Best, Qilin — Reply to this email directly, view it on GitHub <#36?email_source=notifications&email_token=CFA73HMJLW5L5FIDHQTMS435BY653A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOBQHAZDONZZGU2KM4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KYZTPN52GK4S7MNWGSY3L#issuecomment-4808277954>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/CFA73HOSY34UQL7SIXF5HMD5BY653AVCNFSNUABGKJSXA33TNF2G64TZHMYTCMRZGUZDGNBUHA5US43TOVSTWNBXGI2TQOBTGA4DNILWAI> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/CFA73HMUIJGG2RID4I4RPNL5BY653A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOBQHAZDONZZGU2KM4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KUZTPN52GK4S7NFXXG> and Android <https://github.com/notifications/mobile/android/CFA73HLVVB43QTV6VOUJC2D5BY653A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOBQHAZDONZZGU2KM4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2K4ZTPN52GK4S7MFXGI4TPNFSA>. Download it today! You are receiving this because you authored the thread.Message ID: ***@***.***>

[qyQilinYu]

Hi,

It is difficult to diagnose the issue with your sampling without additional data. If you could share the following files, I can help pinpoint the problem more accurately:

• design_report.html
• ranked_designs.csv
• Your target structure and the corresponding sampled PDB (the top5-ranked candidate is sufficient).

In the meantime, here are a few suggestions that might improve your results:

• Interface Exploration:​ To identify an optimal hotspot, consider running several small-batch trials (e.g., 100 samples each) to better explore the interface area.
• Target Selection:​ Prioritize high-resolution target structures for better prediction accuracy.
• Structure Truncation:​ Consider trimming the termini to retain only the core target domain. Flexible N- and C-terminal loops often bring down target confidence (pLDDT) and lower ipTM scores.

and its true that some target and some epitope yield low success rate.

Best,
Qilin

[amarantostaver-cloud]

Thank you very much for your prompt help and valuable advice. My research group specializes in antibody-related work, so if you ever have any questions or need insights in this field, I would be more than happy to return the favor. Following your instructions, I have provided the files to help you troubleshoot the issue. Here is a brief explanation of the dataset: - *jianmo and jinti folders*: Both contain the same antigen. - *jianmo folder*: Uses the structure modeled by *AlphaFold 3*. - *jinti folder*: Uses the *crystal structure from the PDB database*. - *Parameters*: These two runs differ slightly in their parameters (specifically, temperature $T$ and the total number/count). Please note that in the jianmo folder, the stage 2 filter did not succeed, which is why the design_report.html and ranked_designs.csv files are missing. The stage 1 folder contains the filtering results from stage 1, which I believe look somewhat abnormal. Thank you once again for your time and assistance! On a side note, I have a small request: if it is convenient for you, would it be possible to connect on WeChat for easier communication? Please forgive my presumption if you prefer to decline due to privacy considerations. qyQilinYu ***@***.***> 于2026年6月28日周日 15:10写道：

…

*qyQilinYu* left a comment (Mingchenchen/PPIFlow#36) <#36 (comment)> Hi, It is difficult to diagnose the issue with your sampling without additional data. If you could share the following files, I can help pinpoint the problem more accurately: - design_report.html - ranked_designs.csv - Your target structure and the corresponding sampled PDB (the top5-ranked candidate is sufficient). In the meantime, here are a few suggestions that might improve your results: - Interface Exploration:​ To identify an optimal hotspot, consider running several small-batch trials (e.g., 100 samples each) to better explore the interface area. - Target Selection:​ Prioritize high-resolution target structures for better prediction accuracy. - Structure Truncation:​ Consider trimming the termini to retain only the core target domain. Flexible N- and C-terminal loops often bring down target confidence (pLDDT) and lower ipTM scores. and its true that some target and some epitope yield low success rate. Best, Qilin — Reply to this email directly, view it on GitHub <#36?email_source=notifications&email_token=CFA73HKAN3SSGCRAEA5GTWL5CDAFDA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOBSGUZDANRZG4Z2M4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KYZTPN52GK4S7MNWGSY3L#issuecomment-4825206973>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/CFA73HJR6K2YLJWJBJOHQY35CDAFDAVCNFSNUABGKJSXA33TNF2G64TZHMYTCMRZGUZDGNBUHA5US43TOVSTWNBXGI2TQOBTGA4DNILWAI> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/CFA73HO7FZLPSOFC2YNMMRL5CDAFDA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOBSGUZDANRZG4Z2M4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KUZTPN52GK4S7NFXXG> and Android <https://github.com/notifications/mobile/android/CFA73HKUKVKNXJB7QL6WXAL5CDAFDA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOBSGUZDANRZG4Z2M4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2K4ZTPN52GK4S7MFXGI4TPNFSA>. Download it today! You are receiving this because you authored the thread.Message ID: ***@***.***>

[amarantostaver-cloud]

Hello, thank you for being able to help me. 
I have compressed and packaged the file and sent it to you via Google Mail. I'm not sure if you have received it (amarantostaver@gmail.com). 
Thanks again for your help despite your busy schedule, and I wish you a pleasant workday.