Olga Lyudovyk

[Olga Lyudovyk @Olga_Lyu]

A great summary of our research on structure-based T cell specificity prediction

[Biology+AI Daily @BiologyAIDaily]

Ensembles of in silico structures enable T cell peptide-MHC binding prediction

1. The paper introduces enFoldX, a structure-ensemble method that predicts whether a TCR and a peptide-MHC (pMHC) are cognate binders by featurizing many AlphaFold3 (AF3) co-folded complex samples, rather than trusting a single “best” predicted structure.

2. Core idea: AF3 can generate plausible-looking false positives for non-cognate TCR:pMHC pairs, especially when sequences differ subtly. enFoldX turns this weakness into signal by measuring how confidence and interface geometry vary across an ensemble (multiple seeds, multiple samples per seed).

3. enFoldX focuses feature engineering on the binding interface, especially peptide–CDR3 contacts, using AF3 confidence outputs such as pLDDT, PAE/pAE, pTM, and chain-pair/interface variants (iPAE, ipTM), plus inter-chain contact probabilities. It summarizes each feature by mean and standard deviation across the ensemble, capturing both “typical” behavior and uncertainty/diversity.

4. A motivating example is a CMV epitope differing by one amino acid (wild-type NLVPMVATV vs mutant NLVPKVATV): both cognate and non-cognate ensembles can contain “hallucinated” misdocked structures, but the non-cognate shows a much higher hallucination rate across seeds—an ensemble-level signature that single-structure methods can miss.

5. Benchmarking on VDJdb (positives only) required synthetic negatives. The authors generate non-cognate pairs by swapping in “decoy” peptides predicted to bind the same MHC (NetMHC/NetMHCpan), aiming to make negatives fail at TCR recognition rather than MHC presentation. They also test a simpler “permuted” negative scheme (random TCR/peptide shuffling within an MHC allele).

6. With ridge regression and careful cross-validation splits to reduce leakage, enFoldX achieves (human VDJdb) mean AUC ~0.82 for TCR-wise CV and ~0.73 for epitope-wise CV; (mouse VDJdb) ~0.98 and ~0.95, respectively. A notable result is cross-species transfer: a model trained on human data can classify mouse TCR:pMHCs.

7. Feature attribution (SHAP) highlights interface-focused signals: iPAE and ipTM at TCR:pMHC and pMHC interfaces dominate, and both means and standard deviations matter—supporting the claim that “distribution over structures” carries non-redundant information.

8. Generalization tests include: new VDJdb release data (AUC ~0.86 on new TCRs, ~0.77 on new epitopes), a functionally validated TCRvdb benchmark (AUC ~0.75), promiscuous TCRs such as 1E6 (AUC ~0.92 on a 185-epitope evaluation), and MEL5/MEL8 where predicted probabilities separate validated cognate vs non-cognate sets.

9. On the IMMREP25 “unseen peptides” challenge setting (high epitope divergence), enFoldX attains AUROC ~0.65; an ensemble of ridge regression + XGBoost + random forest improves macro-AUROC to ~0.66 and performs strongly under low-FPR metrics compared to reported IMMREP25 baselines, despite label noise concerns in that dataset.

10. For the neoantigen-relevant regime of single amino-acid peptide substitutions, enFoldX is evaluated on eight mutational-scan datasets (macro-AUC ~0.71; range ~0.59–0.82). Additional leave-one-position and leave-one-TCR analyses on a multi-concentration activation dataset show consistent above-random performance across all positions/TCRs, supporting sensitivity to point mutations.

11. The paper also introduces BERTie, a sequence-only BERT-style baseline trained on CDR3β + peptide, showing strong performance when test sequences resemble training data but substantial degradation as epitope distance grows. In contrast, enFoldX performance is reported to be comparatively independent of epitope sequence similarity, aligning with its improved out-of-distribution behavior.

12. Limitations and practical notes: ensembling increases compute cost; performance drops for underrepresented HLA groups (notably some HLA-C alleles), pointing to data scarcity. enFoldX remains informative even when AF3 predictions degrade (e.g., shallow/no MSA, no templates), implying systematic ensemble differences persist even with imperfect structures.

💻Code: https://t.co/CuMEcR9C2q
📜Paper: https://t.co/wNayH8DwJv
#ComputationalImmunology #TCR #pMHC #AlphaFold3 #ProteinStructure #MachineLearning #Bioinformatics #CancerVaccines #Neoantigens #Immunotherapy

[Biology+AI Daily @BiologyAIDaily]

Ensembles of in silico structures enable T cell peptide-MHC binding prediction

1. The paper introduces enFoldX, a structure-ensemble method that predicts whether a TCR and a peptide-MHC (pMHC) are cognate binders by featurizing many AlphaFold3 (AF3) co-folded complex samples, rather than trusting a single “best” predicted structure.

2. Core idea: AF3 can generate plausible-looking false positives for non-cognate TCR:pMHC pairs, especially when sequences differ subtly. enFoldX turns this weakness into signal by measuring how confidence and interface geometry vary across an ensemble (multiple seeds, multiple samples per seed).

3. enFoldX focuses feature engineering on the binding interface, especially peptide–CDR3 contacts, using AF3 confidence outputs such as pLDDT, PAE/pAE, pTM, and chain-pair/interface variants (iPAE, ipTM), plus inter-chain contact probabilities. It summarizes each feature by mean and standard deviation across the ensemble, capturing both “typical” behavior and uncertainty/diversity.

4. A motivating example is a CMV epitope differing by one amino acid (wild-type NLVPMVATV vs mutant NLVPKVATV): both cognate and non-cognate ensembles can contain “hallucinated” misdocked structures, but the non-cognate shows a much higher hallucination rate across seeds—an ensemble-level signature that single-structure methods can miss.

5. Benchmarking on VDJdb (positives only) required synthetic negatives. The authors generate non-cognate pairs by swapping in “decoy” peptides predicted to bind the same MHC (NetMHC/NetMHCpan), aiming to make negatives fail at TCR recognition rather than MHC presentation. They also test a simpler “permuted” negative scheme (random TCR/peptide shuffling within an MHC allele).

6. With ridge regression and careful cross-validation splits to reduce leakage, enFoldX achieves (human VDJdb) mean AUC ~0.82 for TCR-wise CV and ~0.73 for epitope-wise CV; (mouse VDJdb) ~0.98 and ~0.95, respectively. A notable result is cross-species transfer: a model trained on human data can classify mouse TCR:pMHCs.

7. Feature attribution (SHAP) highlights interface-focused signals: iPAE and ipTM at TCR:pMHC and pMHC interfaces dominate, and both means and standard deviations matter—supporting the claim that “distribution over structures” carries non-redundant information.

8. Generalization tests include: new VDJdb release data (AUC ~0.86 on new TCRs, ~0.77 on new epitopes), a functionally validated TCRvdb benchmark (AUC ~0.75), promiscuous TCRs such as 1E6 (AUC ~0.92 on a 185-epitope evaluation), and MEL5/MEL8 where predicted probabilities separate validated cognate vs non-cognate sets.

9. On the IMMREP25 “unseen peptides” challenge setting (high epitope divergence), enFoldX attains AUROC ~0.65; an ensemble of ridge regression + XGBoost + random forest improves macro-AUROC to ~0.66 and performs strongly under low-FPR metrics compared to reported IMMREP25 baselines, despite label noise concerns in that dataset.

10. For the neoantigen-relevant regime of single amino-acid peptide substitutions, enFoldX is evaluated on eight mutational-scan datasets (macro-AUC ~0.71; range ~0.59–0.82). Additional leave-one-position and leave-one-TCR analyses on a multi-concentration activation dataset show consistent above-random performance across all positions/TCRs, supporting sensitivity to point mutations.

11. The paper also introduces BERTie, a sequence-only BERT-style baseline trained on CDR3β + peptide, showing strong performance when test sequences resemble training data but substantial degradation as epitope distance grows. In contrast, enFoldX performance is reported to be comparatively independent of epitope sequence similarity, aligning with its improved out-of-distribution behavior.

12. Limitations and practical notes: ensembling increases compute cost; performance drops for underrepresented HLA groups (notably some HLA-C alleles), pointing to data scarcity. enFoldX remains informative even when AF3 predictions degrade (e.g., shallow/no MSA, no templates), implying systematic ensemble differences persist even with imperfect structures.

💻Code: https://t.co/CuMEcR9C2q
📜Paper: https://t.co/wNayH8DwJv
#ComputationalImmunology #TCR #pMHC #AlphaFold3 #ProteinStructure #MachineLearning #Bioinformatics #CancerVaccines #Neoantigens #Immunotherapy

[Olga Lyudovyk @Olga_Lyu]

Excited that @JonAvramLevine will be presenting our enFoldX method for T cell specificity prediction at AACR D3 tomorrow (July 22). Go Jonathan! 👏

[Jonathan Levine @JonAvramLevine]

Our preprint is now available, and the code is available on my github at https://t.co/U94SXfHGUU and I'll also be presenting this work at AACR D3 (#AACRdrugdev26) in Boston tomorrow afternoon (Wednesday, July 22). Big thanks to all my co-authors and to my advisor @bengrbm (3/3)

[Olga Lyudovyk @Olga_Lyu]

The signal isn't in one predicted structure. It's in the ensemble.

The Anna Karenina principle at work: cognate TCR:pMHC complexes converge to similar structures, while non-cognate ones "hallucinate" diverse solutions. enFoldX turns that variability into specificity prediction.

[Quaid Morris @quaidmorris]

Can we predict which T cells recognize which peptides? A lot of the signal is hiding in AlphaFold3's hallucinations.
enFoldX: instead of one predicted structure, we use the whole noisy ensemble to predict TCR:pMHC binding.
💻https://t.co/JOUdSMGO1P
https://t.co/DD8iIn3AaC

[Caleb Lareau @CalebLareau]

Excited to share the first preprint of a new direction for our group. Led by the fearless duo of @arthurwchow and @hoyin_chu, our foray into computational protein design— 1/n
https://t.co/o5kDJ5TiWr

[Olga Lyudovyk @Olga_Lyu]

Awesome talk by @bengrbm Ben Greenbaum on neoantigen quality predictability and coevolution with T-cells at the big stage at #AACR2024

[Transform Finance - Global Finance Community @TformFinance]

Introducing our newest speaker at the FinCrime & Cybersecurity Summit NYC, on 4th April at the Ease 605 Third Ave -

Alexander Eristavi from City of London Police!

Register with code 'VIP-NETWORK': https://t.co/eAPoJBiFJR

#TransformFinance #FinCrimeSummit #CybersecuritySummit

[Olga Lyudovyk @Olga_Lyu]

@Target, you think making a 33-wk pregnant lady roam NYC streets in the hopes to score some freebies will make me MORE likely to become a loyal Target customer???

[Olga Lyudovyk @Olga_Lyu]

@Target entice me to use their registry with a freebie box. 3 nearby NYC stores tell me it's not (and is never) available + don't pick up their phones so I can't easily check. 'Guest' services advise to "continue to check with stores". I am 33 weeks pregnant. @Target, really?

[Olga Lyudovyk @Olga_Lyu]

@AppleSupport Please help... My iPhone got locked out. I tried to get help online, submitted proof of purchase - then I found out this means @AppleSupport will ERASE my phone. Does it make any sense?? @AppleSupport please ignore my proof of purchase until I can get to Apple Store

[Benjamin D. Greenbaum @bengrbm]

Great work by our @Olga_Lyu using #BERTs for TCR-antigen interactions in #COVID19 presented at #ATCR23 !!! With @elhanati_yuval @SantoshVardhana @quaidmorris & Artem Streltsov!!!
🔥🔥🔥

[Olga Lyudovyk @Olga_Lyu]

@Airbnb @AirbnbHelp I’ve been your loyal customer for over a decade with a 4.9 rating. You close my account with no explanation and no way to appeal. Your customer service can’t shed any light. Your team was supposed to call me - nothing. Seriously? What did I do???

[LBC @LBC]

'Whatever Ukraine asks for, give it to them - now.'

Writer on Ukrainian affairs Paul Niland tells @eddiemair there can't be any more 'excuses' from the West about what weapons they send to the country.

@PaulNiland

[Tadeusz Giczan @TadeuszGiczan]

Yesterday, RIA Novosti published a lengthy piece titled "What Russia should do with Ukraine", which explains in detail what Russia understands by denazification. It's truly horrific: 1/6

[Olga Lyudovyk @Olga_Lyu]

Fresh signs of support for Ukraine: Roosevelt Island, New York City. #stopputin #stopputinnow #standwithukraine #standwithukraine💙💛 #planesforukraine #slavaukraini #heroiamslava #glorytoheroes https://t.co/dWJzgoIfWY

[Robin Menikoff @robinmenikoff]

East Village, New York City #Ukraine #UkraineRussiaWar #StopPutin

[Olga Lyudovyk @Olga_Lyu]

On Tuesday, March 22, call your Representative and tell him/her/them that the US needs to demonstrate more support for Ukraine.
Click here for more info: https://t.co/KMvJDWRqVN
#Ukraine #StandWithUkraine #PlanesForUkraine

[Olga Lyudovyk @Olga_Lyu]

Policy makers: Ukraine needs air support & air power. The 1000+ women and children stuck under the rubble of Mariupol theater for 3 days need your help. @JoeBiden @JensStoltenberg @EmmanuelMacron @BorisJohnson @Bundeskanzler @USDoDGov (Razom for Ukraine)
https://t.co/vKB7rYhFAQ

[Olga Lyudovyk @Olga_Lyu]

@FedorovMykhailo we (New york volunteers) suggest launching the Citizen App in Ukraine: it can allow people to report incidents, ask for medical or humanitarian aid, and can help NGOs and medical teams prioritize their efforts. Citizen App is onboard. @CitizenAppNYC

[Olga Lyudovyk @Olga_Lyu]

@mockingbirdbway Is the show on for tonight Dec 27?