Home page
AI is hyped, useless
(Fig.1) Overhyped AI's prediction rate of proteins is very bad = only 2%
The 1st, 6th, 7th, 2nd-last paragraphs of this hyped news (7/24/2025) say
"Precision cancer treatment on a larger scale is moving closer (= still unrealized ) now that researchers have developed an AI platform to tailor protein components and arm the patient's immune cells to fight cancer." ← hype
"Normally, T cells naturally identify cancer cells by recognizing specific protein fragments, known as peptides, presented on the cell surface by molecules called pMHCs. It is a slow and challenging process to utilize this knowledge for therapy"
"a well-known cancer target, NY-ESO-1, which is found in a wide range of cancers. The team succeeded in designing a minibinder that bound tightly to the NY-ESO-1 pMHC molecules. When the designed protein was inserted into T cells,.. which effectively guided the T cells to kill cancer cells in laboratory experiments (= Not inside bodies, so still useless )."
"expects that it will (= still unrealized ) take up to five years before the new method is ready for initial clinical trials in humans."
↑ So this research tried to design minibinder proteins (= by AI, Alphafold with very low success rate of less than 2%, this-p.8-1st-paragraph ) binding to cancer antigen to stimulate T cell to kill cancers in vitro (= Not inside animals or bodies ).
↑ This CAR-T cell therapy for solid tumors did Not succeed despite many attempts, because T-cells are often blocked from reaching cancer cells inside bodies ( this-3. car T-cell theory in solid tumors ).
This-abstract says
"Despite early signs of clinical efficacy, CAR-T cells have repeatedly failed to achieve curative responses in solid cancers."
↑ This research paper ↓
p.2-last-paragraph says "Among various de novo protein design approaches, RFdiffusion (= trained on experimental protein structure in PDB, this-p.68-4.1, Not quantum mechanics ) is considered one of the most promising tools due to an impressive average of ~10% success rate of designed binders (= too bad and low AI success rate )"
p.3 says "In our work, we focused on
designing binders that specifically target the pMHC formed by MHC-I, presenting the cancer-associated
NY-ESO-1 peptide"
"Following backbone generation
through RFdiffusion denoising trajectories, we designed sequences using ProteinMPNN.
The resulting designs were filtered based on AF2 (= Alphafold 2 ) initial guess"
p.5-1st-paragraph says "there is still a high rate of false positives, where predicted binders fail during experimental testing" ← AI prediction of protein structures is bad.
p.6 says "synthetic genes encoding all 44 miBd designs targeting SLLMWITQC/HLA-A*02:01 were cloned into a lentiviral CAR vector for expression on the cell surface "
"From this library, we identified one candidate (NY1-B04)," ← So only one binder-protein out of 44 candidates predicted by AI bound to the target peptide. ← AI success rate was very bad = only 1/44
p.8-1st-paragraph says "We selected 95 miBds for in vitro testing, constituting a success rate of ~2%" ← Also in design of binders targeting other peptides, the AI success rate was very bad = only 2%.
↑ As shown here, today's AI's success rates of actual proteins' structures were very bad ~2%, because the experimental (static) protein structures in PDB on which AI, Alphafold are trained are different from actual protein structures inside bodies or cells.
↑ To cure cancers, we should stop relying on T cell's immunity that is useless, often blocked by cancers inside bodies.
And we should directly measure atomic structures over cancer cells (= which direct observation of atoms and cells is impossible in today's X-ray and cryo-EM ) by practical multi-probe atomic force microscopes that can also design and directly create molecules or effective drugs binding only to cancer cells and preventing them proliferating.
↑ Development of this useful multi-probe atomic force microscopes has been hampered by the current impractical quantum mechanical atomic model.
Feel free to link to this site.