[BUG][Documentation gap] Template alignments skipped in preprocessing. Request details on input .sto requirements.

[ECalfeeAdaptive]

Documentation gap
Could I please get more information on the requirements for the input .sto files for OF3 templates and how these differ from OF2? I have read through the documentation on expected headers for template formats here. The detailed template explanation part of the OF3 documentation is still missing.

I have been training on OF2 and would like to transition over to training on OF3. I would like to re-use pre-computed alignments from the OF2 precompute alignments pipeline for OF3 training if at all possible but don't understand the current requirements for OF3 inputs.

Unexpected behavior
I have tried using unmodified OF2 precomputed alignments for OF3 training. The hmm_output.sto template alignment files from OF2 do not have the query sequence as the first entry (only template sequences are included). As a result, these templates are skipped in the OF3 training preprocessing templates pipeline because they don't pass the match_query_chain_and_sequence() check here The result is no templates are used by OF3.

To Reproduce

• Run scripts/data_preprocessing/preprocess_template_alignments_train_val_of3.py using alignments/chain_id/hmm_output.sto files precomputed by the OF2 hmmsearch pipeline. Alignments are all for protein chains.

        python3 openfold3/scripts/data_preprocessing/preprocess_template_alignments_train_val_of3.py
        --template_alignment_directory alignments
        --template_alignment_filename hmm_output.sto
        --template_structures_directory databases/templates/mmcif_files
        --template_file_format cif
        --query_structures_directory input_fasta_files
        --query_file_format fasta
        --query_seq_load_logic fasta
        --dataset_cache_file training_cache_without_templates.json
        --updated_dataset_cache_file training_cache_with_templates.json
        --template_cache_directory template_cache
        --max_templates_construct 10000
        --max_templates_filter 20
        --is_core_train True
        --num_workers 48

Expected behavior
I expected the aligned templates to be kept for all chains.

Configuration

• GPU A100 node (96 cpus)
• Installation from repo

Additional context
If possible, please also confirm if the OF2 alignments .sto format (precomputed with same script linked above) should be ok for MSA inputs since it's challenging to check the validity of the msa features.

Label: OpenFold Consortium Member

[jnwei]

@vinay-swamy do you have thoughts about how to adapt the OF2 template alignment pipeline for OF3 inference?

[vinay-swamy]

Hmm yeah this is a tricky one, as in our template alignment pipeline we first initially generate the alignment like this

openfold-3/scripts/snakemake_msa/MSA_Snakefile

Lines 448 to 450 in b7940cb

	## build inital hmm output
	$binDir/hmmbuild --hand --amino output.hmm uniref90_hits.sto > /dev/null 2>&1
	$binDir/hmmsearch --cpu 4 --noali --F1 0.1 --F2 0.1 --F3 0.1 -E 100 --incE 100 --domE 100 --incdomE 100 -A hmm_output_no_query.sto output.hmm {input.db} > /dev/null 2>&1

and then manually add the query sequence back in like this :

openfold-3/scripts/snakemake_msa/MSA_Snakefile

Lines 463 to 469 in b7940cb

	cat query.fa > templates_with_query.fasta
	$binDir/esl-reformat -u fasta hmm_output_no_query.sto >> templates_with_query.fasta
	$binDir/hmmalign -o hmm_output_tmp.sto output.hmm templates_with_query.fasta
	echo "# STOCKHOLM 1.0" > hmm_output.sto
	echo "" >> hmm_output.sto
	echo "#=GS $mgyId/1-$seqlen DE [subseq from] mol:protein length:$seqlen UNKNOWN" >> hmm_output.sto
	tail -n+3 hmm_output_tmp.sto >> hmm_output.sto

Updating your .sto files like this is probably the easiest way forward for now. Longer term we may want to support an alternative way for adding the query sequence back into the template alignment directly within the template preprocessing code. @gnikolenyi what do you think? Is this something we could do with kalign ?

[ECalfeeAdaptive]

Thank you! Just to clarify, is this manual step of adding the query back in to the .sto file also expected/required for MSA alignment .sto files or just template alignments?

[ECalfeeAdaptive]

Also, are the input .sto requirements the same for inference MSA and template alignments? The inference pipeline has a kalign step under some conditions for templates.

[vinay-swamy]

Thank you! Just to clarify, is this manual step of adding the query back in to the .sto file also expected/required for MSA alignment .sto files or just template alignments?

For protein MSAs no but for RNA MSAs, yes

Also, are the input .sto requirements the same for inference MSA and template alignments? The inference pipeline has a kalign step under some conditions for templates.

The .sto requirements should be the same for inference compare to training - we use the same pipeline to generate MSAs for both