Can AlphaFold3 predict bacterial transcription factor binding sites on DNA? I tried to devise a fair test. The results are quite impressive. My test: I chose LexA (favorite TF at the @ErillLab) from E. coli K-12, and I generated "randomized" binding sites from the PWM. First 1/

of all, from what I read from Supplementary 2.5.2, AF3 was trained on Jaspar, which doesn't contain motifs for bacterial transcription factors! Please let me know if I'm missing some way in which bacterial TF motifs may have been available to AF3. Secondly, by "randomized" 2/

sites I mean that I generate the DNA sequences by picking the base at each position according to the probabilities in the position weight matrix. You can get a seq that was not even among the examples. Then I embed each site into random DNA, for a total of 40 bp. The position 3/

at which the site starts is also random (Uniform) so that the sequence may not be in the center. I input the sequence of LexA, and set "Copies" to 2 because LexA acts as a dimer. I then try AF3 on the sequences I generated. I consider the test passed if: (1) the dimer is 4/

assembled correctly; (2) the LexA dimer contacts DNA with the known DNA binding domains in the DNA grooves; (3) it binds exactly on the "randomized" LexA sites (a pattern of 16 bp within the 40 bp). Here are the results on 8 sequences (we can't run >10 jobs/day at the moment) 5/

In 8/8 cases, the TF complex is perfect, and targets DNA using correctly the DNA binding domains. In 4/8 cases, the binding occurs precisely on my "randomized" LexA sites. Uppercase: from PWM; lowercase: random; underlined red: binding expected; highlighted yellow: bound. 6/

Case 3 is particularly convincing because the site came out quite different from the consensus. It's not the classical CTGT-n8-ACAG, but AF3 predicts that the dimer binds there. AF3 can propose more than one model, but I was only looking at the first proposed. If I consider 7/

also the others we have at least one perfect solution in 6/8 cases. I also tried with totally random DNA. As I expected it binds DNA the normal way (helices in groove) even if there's no LexA binding site. Not necessarily a "wrong" prediction given how bacterial TFs "search" 8/